Wearable devices are frequently used to continuously collect physiological and behavioral data using integrated sensors. The strong correlation between activity levels and psychiatric disorders implies that these data offer potential in the diagnosis of depression. The objective of this master thesis is to answer the question of how activity data from wearables can be used to diagnose depression. To this end, the following research question is posed: How can wearables be applied to automatically detect depression states? To answer the research question, a secondary data analysis of the Depresjon dataset was conducted. The dataset includes motor activity data from 23 unipolar and bipolar depressed subjects and 32 healthy controls. Statistical features were extracted from the motor activity data to subsequently feed a random forest classifier. Using the motor activity signal from the wearable, the results show a sensitivity value of 0.941, indicating that depressed subjects are correctly classified 94.1% of the time, and a specificity value of 0.936, indicating that healthy control subjects are correctly classified 93.6% of the time.
Table of Contents
List of Abbreviations
List of Figures
List of Tables
1 Introduction
1.1 Problem and objective
1.2 Methodical procedure
2 Theoretical Framework
2.1 Depression
2.1.1 Biological aspects
2.1.2 Psychosocial aspects
2.1.3 Biopsychosocial pathogenesis model
2.2 Bipolar disorder
2.2.1 Mania and hypomania
2.2.2 Bipolar depression
2.3 Wearables
2.3.1 Categories of wearables
2.3.2 Actigraphy
2.3.3 Wearables in depression research
2.4 Interim conclusion
3 Methodology
3.1 Data collection
3.1.1 Rating scale data
3.1.2 Motor activity data
3.2 Data preprocessing
3.3 Feature extraction
3.4 Classification analysis
3.4.1 Machine learning approaches
3.4.2 Tree-based methods
3.4.3 Oversampling
3.5 Validation
4 Results
4.1 Training and testing
4.2 Validation and comparison
5 Discussion
5.1 Interpretation and implications
5.2 Recommendations for future research
6 Threats for Validity
7 Conclusion and Outlook
References
A Program Code
A.1 data_exploration.ipynb
A.2 machine_learning.ipynb
Abstract
Wearable devices are frequently used to continuously collect physiological and behavioral data using integrated sensors. The strong correlation between activity levels and psychiatric disorders implies that these data offer potential in the diagnosis of depression. The objective of this master thesis is to answer the question of how activity data from wearables can be used to diagnose depression. To this end, the following research question is posed: How can wearables be applied to automatically detect depression states? To answer the research question, a secondary data analysis of the Depresjon dataset was conducted. The dataset includes motor activity data from 23 unipolar and bipolar depressed subjects and 32 healthy controls. Statistical features were extracted from the motor activity data to subsequently feed a random forest classifier. Using the motor activity signal from the wearable, the results show a sensitivity value of 0.941, indicating that depressed subjects are correctly classified 94.1% of the time, and a specificity value of 0.936, indicating that healthy control subjects are correctly classified 93.6% of the time. Based on these results, it is concluded that motor activity can be used to distinguish between the two classes and that the use of wearables is suitable for automatic depression detection. This study suggests an approach both for physicians to diagnose depression and for manufacturers of commercial wearables to integrate automatic depression screening into their IT ecosystem.
Keywords: Wearable, depression, bipolar disorder, motor activity, machine learning, random forest
List of Abbreviations
illustration not visible in this excerpt
List of Figures
2.1 EU-citizens reporting depression in 2019
2.2 Biopsychosocial pathogenesis model for major depressive disorder
2.3 Progression of bipolar disorder
2.4 Categories of wearables
3.1 Flowchart of the applied methodology
3.2 MADRS Development
3.3 MADRS between marital status and biological gender
3.4 MADRS between affection types
3.5 Full activity time series of a control subject
3.6 Full activity time series of a condition subject
3.7 Activity count for a depressed patient
3.8 Activity count for non-zero activity
3.9 Prolonged time without activity
3.10 Nonzero proportion of conditions of condition and control subjects
3.11 Comparison of motor activity at different times of the day
3.12 Locality, spread, and skewness of activity between conditions and controls
3.13 Classification tree
4.1 OOB error rate in RF
4.2 RFC confusion matrix at baseline
4.3 RFC confusion matrix after hyperparameter tuning
4.4 RFC confusion matrices after random oversampling
4.5 RFC confusion matrix after ADASYN
4.6 ROC curve and precision-recall curve for RFC
4.7 Feature importance for RFC
4.8 Partial Dependence Plots for RFC
List of Tables
3.1 Demographic characteristics of condition and control subjects
3.2 Demographic and clinical characteristics of condition subjects
3.3 Statistical features extracted from temporal motor activity data
3.4 Overview of the random forest hyperparameters and typical default values
3.5 The standard confusion matrix M
3.6 Outcome metrics used for validation
4.1 RF classification results
4.2 Random forest comparison
1 Introduction
The use of wearable devices to monitor personal health became increasingly common in recent years. Wearable devices, or wearables, are connectable, body-worn computers with integrated sensors capable of measuring physiological and behavioral data to support users in their daily activities. These include steps, heart rate, energy expenditure, sleep patterns, respiratory rate, blood oxygen saturation, skin temperature, and skin conductance139. Every day, people collect large amounts of data to improve their quality of life, track their fitness levels, or even change harmful behaviors. A consumer survey in 2022 found that 30% of the German population use wearables in their personal life155. Since heart rate and activity levels are often recorded continuously, this data offers significant potential in addition to tracking daily step count or calories burned. There is growing recognition in psychiatry that these activity data are associated with a range of mental disorders, including changes in mood, personality, difficulty coping with daily difficulties or stress, and withdrawal from friends and hobbies [56, 123]. The close association between daily activity, physiology, and psychological well-being makes wearable digital diagnostic devices particularly attractive for diagnosing depression. Such devices can potentially be used to diagnose depression risk and improve mental health screening in the general population. Moreover, due to the high level of granularity of the available data, digital diagnostics have the potential to expand knowledge on the development of depression. To prevent and cure depression, complementary smartphone applications (apps) can be used for digital health treatments and tailored cognitive behavioral therapy.
1.1 Problem and objective
Despite the fact that the market for wearables is constantly growing, the application of wearable technologies for the diagnosis and treatment of depression remains limited. Depression is the third leading cause of lost work years due to disability, affecting around 5% of the adult population worldwide (approximately 280 million people)49. Regardless of its high prevalence, depression is misdiagnosed and untreated in 50% of all cases98. At the same time, the developing COVID-19 pandemic and related economic crises are worsening the mental health of the population176. The economic costs of the disease are considerable: depression is responsible for about one in fifteen days of absence from work in Germany. According to the German Employees' Health Insurance Fund, absenteeism due to depression was 41% higher in 2021 than ten years earlier. During this time, the number of prescriptions for psychotropic medications to treat depression roughly tripled55.
Changes in motor activity are known to be a symptom of depression. Motor activity generally describes the range of active coordinated muscle activity of the human body controlled from the brain124. In psychology, motor activity is viewed as a behavioral phenomenon, while behavior is primarily understood as the observable and measurable activity of an individual. Findings suggest that depressed patients are less likely to be physically active than non-depressed individuals. Numerous studies found that depressed patients tend to have more sedentary lifestyles47. Changes in motor activity are associated with changes in the severity of depression. The onset of depression is associated with a transition from physical activity to a sedentary lifestyle167. Depression may be related to a decrease in motor activity or an inability to maintain prescribed physical activity after a cardiac event138. Physical activity also explains the increased risk of death associated with depression after a cardiac event172. Several hypotheses are proposed to explain the association between depression and the development of a sedentary lifestyle. The lower motor activity in depressed patients may be explained by the association between depression and lower motivation and energy. The finding that negative health habits accumulate in depressed individuals may also help explain the association between depression and motor inactivity. Physical activity is associated with successful treatment of unipolar depressive disorder109 and improves fitness, cognitive function, and overall well-being while reducing or preventing depressive symptoms3. Case-control studies found that patients with depression were less active during the day, but longitudinal studies showed an increase in daytime activity and a decrease in nighttime activity over the course of therapy20. Wearables are able to record and process motor activity data from their users, giving reason for this master thesis to pose the following research question: How can wearables be applied to automatically detect depression states?
The objective of this master thesis was to develop a system for detecting depressive states in patients using motor activity data from a wristband with an integrated accelerometer and a method for automatic classification, allowing diagnosis and immediate treatment. Such a system could potentially be integrated into the smartphone apps of commercial wearables. In this context, it is critical to improve understanding of the variables involved in the development and maintenance of initial depressive symptoms. Information on motor activity can be used for this purpose, as depressed individuals tend to have lower motor activity than healthy controls. Improving these characteristics may contribute to both the prevention of the disorder and the development of effective treatments.
1.2 Methodical procedure
To achieve this objective, a secondary data analysis of the Depresjon 1 dataset was conducted. This dataset contains information about patients with depression and control subjects without depression. Motor activity levels were monitored using a wristband worn on the right wrist with an integrated accelerometer. Initially, data preparation was performed for the analysis, selecting samples from the Depresjon dataset and standardizing the data. In the next step, statistical features were extracted for classification. Subsequently, these features were classified using machine learning (ML) algorithms. ML developed as a subfield of artificial intelligence (AI) using self-learning algorithms that derive knowledge from data to make predictions132.1
Due to the fact that the depressive class is underrepresented in the dataset, different data oversampling methods were used and their effects on classification performance were investigated. In a final step, classification results were validated based on the associated confusion matrices and compared with similar work in the research area.
This master thesis is structured as follows: In the second chapter, a theoretical framework is first established by explaining relevant terminology and fundamentals. Subsequently, the methodological procedure for the secondary data analysis is explained in the third chapter. Results from the analysis are disclosed and visualized in chapter four. In the fifth chapter, the results are discussed and recommendations for research and practice are issued that result from the study. Potential threats to the validity of the study are discussed in the sixth chapter. Finally, a summary of the results and an outlook on possible further research is given in the seventh chapter.
2 Theoretical Framework
The objective of this chapter is to explain the clinical conditions of depression and bipolar disorder in terms of essential, underlying aspects and interrelationships based on existing scientific literature. In addition, the field of study surrounding the development of wearables is defined theoretically and illustrated with application examples. First, the state of depression is defined and explained using a cause-and-effect model in Section 2.1. Subsequently, similarities and dissimilarities to bipolar disorder are presented (Section 2.2). This is followed by a discussion of wearables (Section 2.3) and an interim conclusion (Section 2.4).
2.1 Depression
Depression, or major depressive disorder (MDD), is defined by ICD-11 as a depressed mood and a decline in activity, where the ability for enjoyment, interest, and concentration is reduced114. Self-esteem and confidence are typically low and emotions of shame and self-doubt arise136. The depressed mood fluctuates little from day to day, is indifferent to life circumstances, and may be accompanied by symptoms such as early awakening and morning low37, reduction of psychomotor activity (e.g., slow speaking and walking), agitation, loss of appetite, weight loss, and loss of libido114. A depressive episode is classified as mild, moderate, or severe, depending on the number of symptoms present and the extent to which the patient's daily life is affected. MDD patients are typically treated with cognitive behavioral therapy and/or interpersonal therapy, as well as antidepressant medications such as selective serotonin reuptake inhibitors (SSRIs) and tricyclic antidepressants (TCAs)112. Drug treatment appears to be effective, but the effect may be limited to the most severely depressed patients53. In cases involving self-neglect or significant risk of danger to self or others, hospitalization (possibly involuntary) may be required112.
MDD is a both epidemiologically and socioeconomically significant disorder. It creates considerable suffering for people affected and their families, evokes emotions of helplessness and hopelessness, and results in high costs for the healthcare system. The condition often has a harmful influence on the individuals' social network and ability to work, which is expressed by isolation and loneliness, as well as job loss and early disability14. In 2011, the average 12-month prevalence2 estimates for MDD were 5.5% in high-income countries and 5.9% in low-income countries. Women were almost twice as likely to be affected as men16. According to the European Health Interview Survey (EHIS), Germany has one of the highest proportions of MDD patients in the European Union (EU). 11.6% of people in Germany reported having MDD in 2019, implying that 9.7 million people were affected (Figure 2.1)48. An increased risk of depression was identified in the context of the evolving COVID-19 pandemic, particularly among people with various chronic diseases, people in quarantine, and medical personnel. In these populations, the point prevalence3 was found to range from 31.0% to 44.8%176. To provide the large number of people affected with health services that meet their needs, substantial resources are required. In 2016, the global net present value of spending needed until 2030 to significantly increase effective treatment of depressive disorders was estimated at C146 billion. However, aside the inherent benefits of better health, it was predicted that increased treatment of MDD would lead to large economic productivity gains, yielding a return on investment of C83 billion29.
illustration not visible in this excerpt
Figure 2.1: EU-citizens reporting depression in 2019 (own illustration). Illustrated are the ten EU countries with the highest proportion of reported depression in 2019 in the total population and the average of all 27 current EU member states (country codes according to ISO-3166). The averageintheEUis7.2%. At 11.6%, Germany (orange bar) ranks fifth in the comparison.
Although the course of depression is often episodic, a growing number of patients are affected by a chronic illness19. Chronic MDDs are among the most burdensome diseases worldwide, not only medically but also economically125. Because chronic MDD is considered extremely difficult to treat and has a very low chance of spontaneous remission (below 10%), the reasons for chronicity are increasingly becoming the focus of scientific investigation. Research has shown that childhood trauma is associated with a significantly higher risk of chronic MDD in adulthood71. It is unlikely that dissociation between early trauma and predisposition to depression in stressful life situations occurs until old age31. Some research also suggests that neurotic personality traits may be a risk factor for the development of chronic MDD70. Acute and chronic MDD are understood to be inherent in a heterogeneous process that is influenced by psychosocial factors in addition to biological vulnerability160.
2.1.1 Biological aspects
The best evidence to date suggests that genetic factors play a significant role in the occurence of MDD. However, only the vulnerability to the disease is inherited. The genetic disposition4 can only cause depression when combined with psychosocial triggers. Which neurological and psychological elements are involved, and whether or not they are inherited, is the subject of intensive research that has yet to be completely resolved14. In studies on the relative risk of developing MDD, the likelihood of disease for relatives of depressives was compared to that of relatives of healthy individuals. Numerous studies suggest that first-degree relatives of depressed individuals show a significantly increased relative risk. The lifetime prevalence of MDD among first-degree relatives is 10-30%, based on a 1.5 to 3-fold increased risk of developing depression compared to the general population [14, 142]. Based on twin studies, the heritabil- ity of MDD is estimated to be 31-42%80. According to adoption studies, 38% of adopted children whose birth parents had suffered from MDD also develop the condition (compared to 9% in the control group)23. These findings support the importance of genetic predisposition in depression development14.
One hypothesis is that MDD is caused by a disruption in neurotransmitter5 networks. This concept is communicated to most patients as part of their disease education, and it is also a common starting point for research into the neurological causes of depression. Biological depression research is primarily concerned with serotonin6 receptors. In recent years, a large number of distinct serotonin receptors were discovered, leading biological depression research to focus on these receptor types. Several animal studies suggest that dysregulation of biologic processes at specific serotonin receptors is disease-causing (pathogenic) in MDD, making them therapeutic targets14. Animal studies also show that not only the duration of stress exposure, but also the ability to cope with stress has a significant impact on behavior and changes in the neurotransmitter system. Repeated exposure to uncontrollable stress factors leads to behavioral changes in animals that are described as depression-like (e.g., reduction in locomotion, food intake, mating behavior). This is commonly referred to as learned helplessness 173. Changes in neurotransmitter and receptor levels, e.g., when sadness is triggered, are likely to result in significant differences in local brain activity between acutely ill or remitted MDD and healthy individuals. Psychotherapy is able to alter these circuits and partially reverse the described changes59.
MDD patients tend to have elevated levels of cortisol in their blood. Cortisol is a stress hormone that is secreted by the body in stressful situations and is involved in the fight or flight response of mammals. In addition to cortisol, corticotropin-releasing hormone (CRH) is also investigated. CRH reduces feelings of hunger, increases feelings of anxiety, and causes short-term actitivies such as improving concentration. CRH is significantly increased in newborn animals separated from their mothers. Alterations in CRH were also observed in humans with early trauma. Based on these findings, it was suggested that lowering CRH concentrations in the central nervous system may have antidepressant effects. Yet, it remains uncertain whether a CRH inhibitor has a direct antidepressant effect or only affects symptoms such as anxiety, sleep disturbances, and loss of appetite14.
One line of research is based on the assumption that in MDD not only the balance of neurotransmitters, but also the so-called neuroplasticity is disturbed. These disturbances can be structural and affect the quantity or shape of neurons, or they can affect their function. For example, studies show a loss of volume in the hippocampus in MDD patients. The hippocampus is responsible for controlling emotions such as fear, anxiety and pleasure. It is also involved in sexual behavior and the control of autonomic functions. Volume reduction correlates specifically with cumulative illness duration (but not with episode severity) and appears to persist for years or decades after the depressive episode has faded15. People with MDD also have lower gray matter in some parts of the brain and frontal lobe61. Many elements of MDD could be explained by a disturbance in neuroplasticity. In particular, this would establish the link between stress and depression and help answer the question of why stressful life experiences trigger a depressive episode or why learning processes can alter molecular neurobiological processes (e.g., through psychotherapy)14.
More than 90% of all MDD patients suffer from sleep disorders. MDD is also one of the most common causes of insomnia. Slow-wave sleep is often shortened in depressed individuals, as is the latency between sleep onset and the occurrence of the first rapid eye movement (REM) episode, and the intensity of REM sleep. These problems with sleep architecture usually disappear when MDD improves. However, the relationship between sleep and MDD remains unclear14.
2.1.2 Psychosocial aspects
According to psychodynamic theories, MDD is caused by a fragile self-esteem system and inadequate coping strategies that develop in early infancy. Life event research has also shown that stressful life events can trigger MDD. These are often loss-related experiences (e.g., divorce or end of deployment)129. The weaker the social support network, the more likely MDD is to be triggered148. However, there are also findings that show that an experience of loss is only temporally related to the onset of the illness in about a quarter of depressed people. Moreover, only about 20% of those who experience a loss develop MDD14. This underscores the importance of vulnerabilities in MDD development.
Modern behavioral-cognitive multifactorial models of MDD integrate behavioral and cognitive treatment concepts21. One noteworthy model is the theory of reinforcer loss25. According to this theory, the lack of behaviorally contingent positive reinforcement (reward) is a critical factor in the development and maintenance of MDD. The brain's reward center generates pleasant emotions when triggered by something enjoyable. When active, it is a circuit in the brain that reinforces behavior. When such rewards become rare, a melancholic mood and resignation are triggered, and the individual gradually reduces actions that could enable them to obtain other, alternative reinforcers (depressive vicious cycle). Numerous empirical studies have found a relationship between MDD and a low incidence of positive reinforcers following depressiontypical behavior21. The cognitive psychological hypothesis64 and the cognitive theory of learned helplessness150 are two other important behavioral cognitive models. Modern
cognitive behavioral therapy (CBT) integrates these approaches and theories14.
Interpersonal theories always assume a psychosocial and interpersonal context to explain the occurrence and duration of MDD148. Thus, MDD is primarily viewed as a relational disorder. Early social learning experiences are thought to shape people's cognitive, emotional, and interpersonal schemas that serve as prototypes for their relationships with others. MDD patients have become resigned to not getting what they want from others, to the extent that these experiences do not allow them to articulate their needs in interpersonal relationships14. Numerous studies from life events research, social support research, and epidemiological findings have supported the role of interpersonal elements in the manifestation of MDD [38, 171]. Low social support, loneliness, drastic life events such as marriage, divorce, moving, retirement, unemployment, a chronic illness of the patient or a relative, or chronic stress are therefore considered empirically established risk factors for MDD14.
2.1.3 Biopsychosocial pathogenesis model
In summary, dynamic and complex models that take into account genetic and neurobiological factors, early trauma, certain personality traits, and psychosocial stress factors can provide a causal description of acute and chronic MDD. To this end, Brakemeier et al.2008 developed a cause-and-effect model (Figure 2.2) of the disease development (pathogenesis) of MDD. The above suggest that vulnerability factors for the development of MDD can be divided into three categories: biological, psychological, and environmental factors.
- Biological vulnerability factors: genetic predisposition, stress regulation disorder, dysregulation of neurotransmitter systems, dysregulation of central nervous system functions.
- Psychological vulnerability factors: early acquired low self-esteem, learning deficits and dysfunctional schemes, attachment disorders and neuroticism.
- Environmental vulnerability factors: traumatic events (especially in chronic depression), such as sexual or physical abuse, neglect, early childhood isolation, adverse social conditions and interactions.
Individuals with the described vulnerabilities and the resulting specific cognitions, skills, and actions are more likely to develop episodic or chronic MDD when exposed to situational stressors. These can take various forms, but especially social stressors (e.g., loss of loved ones), appear to have a significant impact. Since chronic MDD is considered extremely difficult to treat and has a low chance of remission, validated support tools for better illness management are of particular interest in this group.
2.2 Bipolar disorder
Affective disorders can be classified by the extent and severity of mood elevation and range from unipolar to bipolar II to bipolar I. Individuals with unipolar MDD only have depressive episodes, whereas individuals with bipolar II disorder (BD-II) or bipolar I disorder (BD-I) have more intense episodes of mood elevation60. Unlike MDD, medical treatment for bipolar disorder (BD) includes lithium carbonate, anticonvulsant medications, and/or antipsychotic medications.
illustration not visible in this excerpt
Figure 2.2: Biopsychosocial pathogenesis model for major depressive disorder (own illustration based on14 ). Biological, psychological, and environmental vulnerabilities predispose to the development of acute and chronic MDD. The occurrence of acute psychosocial stressors may favor development of acute MDD. Certain personality traits increase the likeliness of developing chronic MDD. In cases of high vulnerability, chronic MDD may occur even in the absence of acute stressors.
BD affects more than 1% of the world population5. In a German mental health survey75, the 12-month prevalence was 0.6% for BD-II and 1.0% for BD-I. BD is more common in younger people, with the highest prevalence in adults under 35 years of age. Since BD is most commonly diagnosed in young adults, it affects the economically active population and therefore imposes high societal costs33. BD affects biological men and women equally75.
2.2.1 Mania and hypomania
According to ICD-11, BD is characterized by at least two episodes in which the mood and activity level of the affected person are significantly disturbed. In this context, the disturbances alternate between elevated mood and increased activity (hypomania or mania) and lowered mood and decreased activity (depression; Figure 2.3). The difference in diagnostic criteria between BD-I and BD-II is primarily in the intensity of manic episodes. BD-I involves episodes of severe mania, whereas BD-II involves episodes of less intense hypomania111. A hypomanic episode is defined in the ICD-11 as lasting at least several days, whereas a manic episode lasts at least one week114. A manic episode impairs social or occupational functioning and may involve psychotic symptoms or even require hospitalization, whereas a hypomanic episode may be noticed by others but usually does not cause severe impairment or necessitate hospitalization. Psychotic symptoms occur in about 75% of individuals suffering from an acute manic episode. However, in some cases of hypomania, job performance may temporarily improve due to increased productivity and good mood60.
illustration not visible in this excerpt
Figure 2.3: Progression of bipolar disorder (own illustration based on60 ). In bipolar disorder, depressive episodes alternate with episodes of mania and hypomania. Often, depressive episodes are significantly longer than (hypo)manic episodes.
2.2.2 Bipolar depression
The ICD-11 criteria for bipolar and unipolar depression are the same, and the severity of the episode is measured with the same scales: the Hamilton Depression Rating Scale (HDRS)65 or the Montgomery-Âsberg Depression Rating Scale (MADRS)100. The main difference between these two scales is that the MADRS is more responsive to change and does not focus as much on somatic symptoms as the HDRS60. With the MADRS, clinicians score 10 depression- related criteria based on observation and contacts with the patient, and the total score (060) indicates the severity of depression. Scores below 10 indicate an absence of depressive symptoms, whereas scores above 30 indicate a severe depressive state56.
Bipolar depression, in contrast to unipolar depression, often begins at a younger age161, has more frequent episodes of shorter duration52, has an abrupt onset and offset4, and is triggered by stressors at earlier stages104. A family history of mania is also a risk factor for bipolar depression36. One of the diagnostic challenges is the long delayed distinction between BD and unipolar MDD. 40% of BD patients are initially classified as unipolar MDD. This ambiguity is compounded by the fact that depression is the most common polarity in BD. The total duration of depressive episodes in BD as well as the duration of depressive episodes are much longer than in mania or hypomania (Figure 2.3)90. Often, BD is not recognized until a mood shift to hypomania or mania occurs. Other indirect indicators of BD include: (i) psychosis or mental breakdown; (ii) early disease development, usually with depression; (iii) recurrent episodes (e.g., more than 4 depressive episodes within 10 years); and (iv) depression with severe agitation, anger, other hypomanic features, or psychotic symptoms6.
Given the long duration of depression in BD patients, depression can be expected to be associated with dysfunction and impairments such as limited academic progress and poorer occupational performance. Approximately 80% of BD patients lose their employment to some extent, and 30-40% are unemployed for long periods during their working lives, with much of this impairment attributable to depression179. BD patients are more vulnerable to developing vascular conditions and shortened life expectancy. Obesity, diabetes, migraines, and viral disorders are more common in people with BD96. In BD patients, the risk of heart muscle infection is 37% higher (88% higher in women), stroke is 60% higher, and heart insufficiency is 230% higher175. Various general medical disorders in BD patients lead to worse clinical outcomes, resulting in a reduced life expectancy of 12-15 years. Substance abuse, smoking, being overweight, being unmarried, and having poor access to medical care are all co-occuring variables, while reduced lifespan may be linked to depression in particular6.
The reported annual suicide rate in Germany averages 12.3 per 100,000 citizens or 0.012% per year, with 18.6/100,000 (0.019%/year) in the male population and 6.2/100,000 (0.006%/year) in the female population113. The suicide risk among BD patients is about 20-times above general population risk rates and among the highest of all psychiatric disorders7. One-third to one-half of BD patients attempt suicide at least once in their lifetime, with approximately 1520% of attempts completed146. In patients with affective disorder, depressive episodes are more associated with suicide than other illness states, especially if accompanied by (hypo-)manic features, co-occuring substance abuse, and following previous suicidal acts. The prolonged delay of recognition and intervention in BD contrasts strikingly with observations that 50% of suicidal acts among BD patients occured within the first 2-3 years of illness6. Suicidal ideation in BD patients requires immediate and comprehensive evaluation and care, whereby the availability and potential lethality of methods, and the presence of protective factors should be considered60.
To summarize, there are challenges in differentiation between bipolar depression and unipolar MDD, and misdiagnosis or underdiagnosis can have dramatic consequences. Depression rating scales help determine the severity of depression but provide little information about the distinction between BD and MDD. Because bipolar depression is a recurrent and chronic disorder, verified and objective assessment methods to detect future episodes are critically important for this population.
2.3 Wearables
The rapid development of the Internet of Things (IoT) has enabled the development of small electronic and computer devices that can be placed on an individuals' body. As the smartphone market becomes increasingly saturated, users are becoming more interested in wearable computing devices to improve their quality of life. Such devices are commonly referred to as wearables and include smart watches, wristbands, glasses, jewelry, clothing, patches and similar technological artifacts151. According to recent market reports35, global shipments for wearables reached 533.6 million units in 2021, an increase of 20% over 2020. Shipments are expected to grow to 637.1 million units per year through 202417, resulting in a compound annual growth rate (CAGR) of 6.1% from 2020 to 2024. The global market value was C58.1 billion in 2021, and is projected to reach C181.4 billion by 2031, growing at a CAGR of 10.9% from 2021 to 2031143. Wearables provide a variety of value-added services, such as indoor navigation and location, financial payments, sports analytics, and health insurance analytics151. Different definitions for wearables exist in the literature107. Steve Mann is considered the inventor ofmodern wearables and his definition1996 is provided here:
“Wearable computing is the study or practice of inventing, designing, building, or using miniature body-borne computational and sensory devices. Wearable computers may be worn under, over, or in clothing, or may also be themselves clothes.” 94
In relation to the research question, the definition indicates that wearables are capable of obtaining data from sensors located close to the body, making them suitable devices for collecting motor activity data. A sensor is a device that measures a physical property and generates a corresponding output. Sensors are constantly monitoring people today. Many sensors are incorporated into wearables to track location, movement, communication or social interaction, light, audio, surrounding digital devices and other variables. Wearables with built-in sensors record movement and physiological processes. This records contain a wealth of information about people's behavior, and ultimately their mental health99.
Wearables have the potential to bring multiple benefits to health research because they are typically unobtrusive, cost less thanstandard research devices, are easyto handle, and are affordable to consumers73. The quality and accuracy of wearables increased in recent years, leading to more clinical certifications54. Wearables can collect long-term data in participants' natural environments, enabling momentary assessments. Therefore, wearables are valuable innovations, especially for health research data collection in large study populations, such as global health or epidemiological studies, or in low-income settings [46, 73]. As a result, wearables are used in a variety of health-related areas, e.g. as acoustic gastrointestinal sensors for predicting intestinal obstruction78, measuring UV radiation exposure8, measuring heat-related illness39, electrolyte monitoring for exercise management133, panic attack prediction163, continuous noninvasive blood glucose monitoring24, and smart inhalers and activity trackers for asthma monitoring165. Smart wristbands combined with ML methods are proven to pre-symptomatically detect COVID-19 infections135. This opens up the possibility of using wearable technologies for early detection of other infectious diseases. In the near future, wearables have the potential to be used for the development of innovative diagnostic procedures or therapies. For example, wearables could detect human movement using sensors such as inertial sensors, bending sensors, and muscle activity signal sensors45.
2.3.1 Categories of wearables
In recent years, a wide range of wearables with different functions and wearing options were introduced in the market. Seneviratne et al.2017 propose to categorize wearables into three groups: 1) accessories, 2) e-textiles, and 3) e-patches (Figure 2.4). Accessories are wearables that are worn by the user outside of clothing and are not part of the main clothing items. Smartwatches, smart wristbands, smart glasses, smart rings, and various clip-on widgets are examples of such devices. Accessories are by far the most popular category of wearables.
The most popular subcategory of accessories are wrist-worns. As the name implies, these are wearables that are worn on the wrist. The majority of wrist-worn products are either smart-
illustration not visible in this excerpt
Figure 2.4: Categories of wearables (own illustration based on151 ). Wearables are grouped into accessories, e-textiles and e-patches. Accessories include the popular wrist-worns such as smartwatches and smart wristbands, as well as head-worn devices and smart jewelry. Etextiles include smart garments and hand/footwear, while e-patches include sensor patches and e-tattoos. Capturing human physiological and biomechanical signals is one of the most important applications of wearables.
watches or smart wristbands, while smartwatches are the most popular form of wearables. The value of the smartwatch market generated worldwide was around C37.4 billion in 2021, and according to recent predictions, global market value will reach C57.9 billion by 2026 (CAGR 2021-2026: 7.6%)116. In terms of market share, Apple Inc. (Cupertino, CA, USA) dominates the smartwatch sector with a 36% share of global shipments in the first quarter of 2022, followed by Samsung Electronics Co., Ltd. (Suwon, South Korea) with a 10% market share134. Inaddition to smartphone manufacturers such as Apple and Samsung, several traditional wristwatch manufacturers such as Casio Computer Co., Ltd. (Tokyo, Japan), TAG Heuer S.A. (La Chaux- de-Fonds, Switzerland), and Rolex S.A. (London, UK) have also introduced smartwatches to the market, indicating the popularity of this group of devices. Current smartwatches usually have two functions. First, they serve as communication and notification tools, complementing smartphones with features such as receiving notifications (e.g., phone calls, text messages, email, voice commands, and weather updates) and performing microinteractions (e.g., launching an app on the phone, limited web browsing, adding reminders, and taking voice commands). Second, most smartwatches can also detect some human physiological and biomechanical signals, allowing them to act as fitness monitoring devices. This, in turn, allows users to record their daily activities, such as automatically recording exercise duration, heart rate, step count, and calories burned. The data collected is transmitted to the users' smartphone or a cloud server for analysis and enhanced user display, e.g. in the form of dashboards151.
Another common type of wrist-worn wearables are smart wristbands. The smart wristband market is currently growing swiftly, and the growing interest in health and wellness is likely to drive sales growth in the coming years. In 2022, the smart wristband market revenue is expected to reach C16.0 billion. Revenue is estimated to grow at an annual rate of 12.3% (CAGR 2022-2027), reaching a market value of C32.1 billion by 2027156. While smart wristbands and smartwatches share certain similarities, smart wristbands are primarily designed to track a specific set of health and fitness activities. As a result, some smart wristbands do not have a display for notifications, and most have a limited form factor compared to smartwatches that are designed to replace traditional watches. The embedded system7 design ensures that the typically long battery life of smart wristbands meets the need for extensive outdoor use in sports such as cycling, jogging and hiking. On the other hand, such designs minimize interaction with the user, and the main role of the device is to passively observe user behavior. Overall, wrist- worn wearables are a major contributor to the growing popularity of wearables. Currently, the two main subcategories of smartwatches and smart wristbands cover two distinct user needs: they replace the traditional wristwatch and serve as an extension device for the smartphone, while they provide accurate and specialized tracking of a variety of fitness activities, with some overlap of basic fitness functions. For everyday fitness monitoring, it is possible that these two types of products will merge in the future. Nonetheless, more sophisticated fitness wristbands are likely to continue to exist for users who require advanced analytics151.
Further accessories include head-mounted devices worn over the head or around the neck. Smart eyewear as well as headsets and earbuds are two major subcategories of head-mounted wearables. Smart eyewear includes eyeglasses or contact lenses equipped with sensors and wireless connectivity. Recent advances in augmented reality (AR) and virtual reality (VR) are likely to improve popularity in this category, as there are numerous new applications. For example, it is predicted that smart glasses may become popular as productivity enhancers in the workplace. However, the main limitations with existing models are the limited battery life and the weight of the devices. Bluetooth headsets were the first consumer wearables and are the most popular form of wearables today. While voice communication and music playback remain the most common uses for Bluetooth headsets, there are now a variety of devices that can record fitness activities and act as smart assistants. Smart jewelry, such as smart rings and bracelets, is another accessory subgroup. Smart jewelry is typically designed to alert the user of smartphone notifications, make payments, recognize physiological and biomechanical signals from the user, or sense the environment. The ability to comfortably wear sensor functions in smart jewelry is a significant advantage151.
E-textiles are main clothing items that also function as wearables, such as smart clothing and smart shoes. This items are primarily used to monitor human physiological signals and biomechanics for a variety of applications, including health and sports117, environmental sensing for military or hazardous environments145, and sensory/haptic applications such as therapeutic massage92. Smart garments and hand/footwear are the two main subcategories of etextiles. Smart garments include main clothing items such as shirts, pants and undergarments. The subject of smart garments is to develop smart clothing items with conductive yarns and to integrate electronics into this items. Subsequently, smart textiles collect touch and gesture data and communicate wirelessly with smartphones. Google LLC (Mountain View, CA, USA), in collaboration with textile company Levi Strauss & Co. (San Francisco, CA, USA), has developed a jacket that connects to the smartphone, thus enabling the user to answer calls or take photos and videos through gesture control on the sleeve of the jacket22. Another form of smart garments monitors human vital signs for purposes such as health, rehabilitation, and safety. Early examples of wearable health systems use knitted textiles to monitor biomechanical signals such as heart rate, respiration rate, and dehydration [40, 118, 119]. Smart hand/footwear, which include smart shoes, smart socks, and smart gloves, are another subcategory of e-textiles. Like smart garments, smart hand/footwear offers applications for monitoring human physiological signals and biomechanical processes [79, 151].
E-patches are skin patches that can be adhered or tattooed. Therefore, e-patches are divided into two types: Sensor patches, which are tiny devices with sensors, and e-tattoos, which are skin-like e-patches with miniaturized sensors and electronics. Sensor patches are adhered to the skin and contain a foam component and integrated electronics that capture physiological data such as the user's pulse. Sensor patches are being used in medical studies, for example to monitor body temperature81. E-tattoos use flexible and elastic eletronic circuits to enable sensor integration and wireless data transmissions. Monitoring of human physiological signals and biomechanics, but also communication and notification to the user and contactless payment are available features in various existing devices151.
2.3.2 Actigraphy
Actigraphy is a quantitative tool for monitoring and analyzing overall motor activity, as well as daily variations in motor activity10. Actigraph units, or actigraphs, are compact, lightweight electronic devices based on accelerometers, often worn around the waist, wrist, and/or ankle, that digitize motion in one or more dimensions at specified time intervals (e.g., every 60 sec- onds)forasmanydaysasinternalmemoryisavailable(typicallyoneweektoonemonth)162. Since the 1990s, researchers were investigating the use of actigraphy to monitor depression symptoms, which entails employing wristwatch-like sensors to collect activity or sleep data gen- eratedbymovements141. Actigraphs are produced byvarious companies, with the majority of research using Cambridge Neurotechnology Ltd. (Cambridgeshire, UK) devices. However, new devices and algorithms are continually being developed, and there is no consensus on which technology or algorithm is best for screening depressed patients91. In addition, although most actigraphs use accelerometers to measure activity levels, there are no common units of measurement for the devices from different manufacturers140.
Accelerometers in wearables can measure activity parameters, sleep patterns, and circadian rhythms. Piezoelectric accelerometers are frequently deployed. The core of a piezoelectric accelerometer is a piezoelectric element, usually ceramic. When physically stressed, whether by tension, compression, or shear, it generates an electrical charge that corresponds to the applied force. According to Newton's second equation of motion (F = m x~a), the piezoelectric element of an accelerometer is constructed so that when the accelerometer vibrates, the mass exerts a force on the piezoelectric element that is proportional to the acceleration of vibration12. Human motion can be characterized in terms of anatomical planes that correlate with three-axis accelerometry using accelerometers. When motion occurs along the axial plane or horizontal axis, it is either up or down. If the motion is along the lateral plane or lateral axis, it is from right to left and vice versa. If the movement is in the frontal plane or vertical axis, it can be forward or backward144.
Circadian rhythms are biological activities that fluctuate over a 24-hour period and regulate body functions and processes of mammals. While it is possible to study circadian rhythms during a single 24-hour period, evaluation over a week or month provides far more reliable results. The average activity level during a 24-hour period or multiple periods is called the measure. The acrophase is the period during which the maximum activity level occurs. Environmental signals, called zeitgebers, control circadian rhythms. The solar light-dark cycle is a well-known zeitgeber, and the sleep-wake cycle is the circadian rhythm associated with it. A well-coordinated sleep-wake cycle allows individuals to sleep and rest sufficiently at night while being productive during the day. Therefore, the circadian rhythm is synchronized with the light-dark alternation of the Earth's day-night cycle41. When this circadian rhythm is disrupted, it can lead to sleep disorders, including insomnia. According to research, the circadian rhythm plays an important role in many areas of physical and mental health158. Moreover, several studies have found an association between factors related to circadian rhythm and MDD symptoms. Bright light exposure (BLE) was found to be helpful in restoring circadian rhythms in individuals with seasonal-linked MDD174. Other findings suggest that BLE and sleep deprivation can also be useful in treating seasonal-independent MDD57. Actigraphy can help track the sleep-wake cycle and thus sleep quality. Consequently, actigraphy is considered a helpful and cost-effective approach for screening certain sleep disorders, such as insomnia and schedule disorders, and tracking their treatment progress141.
In actigraphy, essential methodological aspects to be considered are the site of attachment and statistical procedures. A sensor is attached to a body location such as the wrist, waist or ankle to measure activity. The device responds only to its own movement, which corresponds to the movements of the attachment point. Placing the device on various body sites provides information about those sites, but only those sites. Both walking and running use all parts of the body, albeit in different ways. Sitting in a chair reading a magazine immobilizes the waist and ankles, but not the arms and hands, which are needed for turning pages and possibly taking notes162.
Another difficulty is that activity levels do not remain constant over time, but change in two significant ways. When behavior is assessed in time epochs (typically one minute), autocorrelation is a feature of how activity levels vary over time. Autocorrelation occurs when observations in a time series are not independent of each other, but correlate with each other at a different point in time. This point can be either in the past, or in the future. Since it usually takes more than one minute to reach a place, if a person walks during one minute, it is likely that they will continue to walk during the next minute. If a person sits during one minute, it is very likely that they will also sit during the following minute. Most statistical procedures, such as t-tests, are ineffective because autocorrelation contradicts the general premise of independence. Autocorrelation is reduced when activity levels are aggregated over sufficiently long periods, such as one hour. The second characteristic of repeated activity measurements is that the activity values tend to form a Poisson distribution, where the standard deviation is equal to the mean in the one-minute period of temporal resolution. The majority of activity data is within one standard deviation of the Poisson mean, as opposed to three standard deviations for a normal distribution, whereby a plot of activity versus time using one-minute epochs shows that the activity level varies dramatically from one minute to the next, resulting in enormous fluctuations162.
2.3.3 Wearables in depression research
The development of wearables has also led to their active use in depression research. Wearable sensors such as activity trackers capture fine-grained data that characterize physiological markers that could be used for prompt, unobtrusive, and scalable depression diagnosis. As such, wearable data can be used to predict symptoms of depression and anxiety101, and to assess sleep as a physiological marker for depression178. Data obtained through wearables is expected to enable tailored, interactive, and contactless healthcare in a cost-effective manner91. The lack of data that characterize physiological markers that directly reflect disease severity remains a major challenge in clinical trials in psychiatry9. The advantage of introducing wearables into depression research is that they allow continuous and objective monitoring of patients. With this technology, real-time changes between patients' hospitalizations can be objectively documented and treatment effects can be more reliably assessed. Mood surveys such as the MADRS are commonly used to assess depression (Section 2.2.2). However, this type of mood assessment is often compromised by various biases, such as recall bias8 88. Depending on the degree of variance between patients or physicians, this can lead to erroneous results2. Furthermore, current psychiatric assessment is limited to the day of the physician visit and does not necessarily happen during a crisis (e.g., bad day or relapse); it does not adequately reflect the patient's subjective experience or the real-life effects of therapy152. In addition, errors that may occur during manual data entry by physicians or researchers can be avoided by using wearables91. Patients may also use wearables to track their symptoms. For example, patients can use wearables to measure their current state of stress or depression. Personalized feedback therapies were shown to help patients improve their depressive symptoms and avoid actions that could worsen their mood. Providing depressed patients with regular information about their mood was shown to help them control their depression. Communicating such information via wearables has a similar effect153. In addition, speech patterns and voice analysis can be used to diagnose depression severity and therapy response using sensors103. There is limited evidence linking sensor-based data to affective disorder status or depression severity1. However, effective data management is expected to usher in new delivery methods of healthcare characterized by targeted and timely treatments152. For example, Stanford researchers developed a wearable that continuously records physiological data and enables practicioners to monitor and analyze the cortisol content in the patients' sweat to analyze the response to certain antidepressants72. Another study using activity data from smartphones found that physical activity was related to scale rating outcomes in BD patients115. Activity data from smartphones was also found to be feasible for predicting clinical symptoms of MDD127.
Several studies have found a link between physical activity and depression. In MDD patients, the hippocampus in the brain is reduced in size (Section 2.1.1). Physical exercise supports the formation of neurons in the hippocampus and strengthens neuronal connections, which can help with MDD154. People who do not engage in regular physical activity are more prone to depressive symptoms44, and depression maybe a major risk factor for a sedentary lifestyle47. Most research on wearable devices has produced results consistent with these assertions. Case control studies revealed that both MDD and BD patients had reduced daytime activity, but longitudinal studies revealed an increase in daytime activity and a decrease in nighttime activity during the course of therapy20. Compared with healthy controls, the depressed state in MDD and BD patients is also associated with greater variability in activity levels and less complexity of activity patterns. However, contrasting physical activity patterns were found in certain MDD and BD patients, characterized by higher average activity levels, lower variability, and greater complexity of activity patterns more similar to those ofmanic individuals85. In contrast to the general loss of initiative and interest characteristic of depressive states with lower psychomotor activity, such depressed states are often associated with irritability, restlessness, and aroused inner tension76. When MDD inpatients were treated with antidepressants, their activity levels increased dramatically after discharge, as determined byactigraphy131. Compared to healthy controls, BD patients were found to have decreased mean motor activity, prolonged sleep, and lower sleep quality43. Motor activity and actigraphic measurement patterns can also reliably predict hospital discharge dates using a linear regression model122. Since insomnia is a risk factor for MDD (Section 2.1.1), sleep-related measures can be used to predict MDD50. Individuals with MDD report poorer sleep quality than healthy controls95. However, most studies using actigraphy for sleep evaluation in MDD patients do not confirm this conclusion91. Nevertheless, one study showed that insomnia symptoms measured by actigraphy can be used to predict pain symptoms in MDD patients30, implying that further research is needed to gain insight into the relationship between sleep and MDD. However, when BD patients are compared with healthy controls, there are significant differences in total sleep time, sleep latency, and waking after sleep onset159.
Most studies of wearables for depressed individuals use either actigraphs or commercial wearables that are not designed for medical use, such as the Fitbit Charge (Fitbit Inc., San Francisco, CA, USA) and the Apple Watch (Apple Inc.). The wrist is the primary site of use for research participant assessment. Wearables worn on the wrist are actively used for research because they are not distracting and are familiar to most individuals91. Two examples of relevant medical wearables are the Parkinson's KinetiGraph (PKG) and the Empatica E4 wristband (Empatica Inc., Boston, MA, USA). The PKG is a wrist-worn actigraph with a three-axis accelerometer that was originally developed to detect motor symptoms such as tremor in Parkinson's disease patients62. This device was used to measure motor symptoms in MDD patients128. The Em- patica E4 is a wrist-worn device that includes a heart rate sensor, a sensor to measure skin electrical conductivity, an optical thermometer to monitor peripheral skin temperature, and a three- axis accelerometer to assess movement and sleep characteristics. It was used to track changes in the intensity of patients' MDD symptoms121. Two examples of consumer wearables that were not initially designed for medical usage are the Fitbit Charge and Apple Watch. The Fitbit Charge is a commercially available activity tracker that contains a three-axis accelerometer that tracks user movement patterns. It was used to find associations between MDD symptoms and increased nighttime heart rate fluctuations139. Additionally, step and sleep data obtained
by Fitbit devices were used for prediction of scores on depression rating scales93. The Apple Watch is a smartwatch that includes a variety of mobile applications (apps) as well as phone and text messaging capabilities. It has a small touchscreen, which was used to provide study subjects with cognitive and emotional assessments34. This proves that - in addition to collecting physiological data - wearables can also be used for high-frequency assessment of cognition and mood91.
Challenges associated with the use of wearables in depression research include issues of patient adherence and compliance. The degree of comfort with which wearables can be used varies, as does the user comfort with wearing the devices. The effectiveness of mobile health therapies is highly dependent on the design of the intervention, and user acceptance of wearables depends strongly on the age ofthe user157. Otherissueswiththeuseofwearablesincludeunreliability and inaccuracy. The versatility of wearables makes them difficult to use for monitoring clinical symptoms. The unreliability of wearable sensing systems and data processing algorithms also makes their introduction into medical practice problematic. In addition, data can be interrupted by a variety of noises generated by the environment and by the physical condition of the person wearing the device91. Another problem is feature extraction in terms of reliability. Mohr et al.2017 developed a multilayer hierarchical framework for how raw sensor data (e.g., position, movement) are translated into low-level features (e.g., activity, total sleep time). Low- level features are combined to form high-level behavioral indicators (e.g., psychomotor activity, sleep disturbance). A combination of high-level behavioral cues is used to infer clinical state (e.g., depression). Although this is a conceptually sound framework, it is uncertain whether the data, such as accelerometry, can actually indicate psychomotor agitation or delay. For example, sleep is approximated by the patient's movements and pulse rather than by monitoring brain waves. There is also the issue of data protection and ethics. Since data received via wearables is stored on external servers, there is a possibility of data leakage. For this reason, legal controls are critical for use in the medical field91.
2.4 Interim conclusion
In conclusion, there is a significant prevalence of MDD and BD in the world population, and Germany is having a significant proportion of MDD patients in total population in particular. MDD can be episodic or chronic and is caused by a dynamic and complex interplay of biological, psychological, and environmental vulnerabilities and stress factors. Psychotherapy is a proven means of treating MDD, and psychologists assess both MDD and BD patients with rating scales based on qualitative questionnaires. MDD and BD are both affective disorders, with BD displaying additional episodes of (hypo)mania. Diagnostic distinction between BD and MDD patients is challenging, especially since BD patients mostly dwell in a depressive episode.
Reduced physical activity is one manifestation of depression in unipolar MDD and BD. Such patients are less active during the day in their depressed state and show greater variability in activity levels. In this scenario, evaluation of physical activity data collected by wearables may provide an opportunity for objective assessment of MDD and BD patients. Wearables were used in a number ofhealth research settings, and potential uses related to the treatment ofdepression were also explored. Various wearables provide the user with different services, have different functions, and wearing options. The vast majority of wearables used are wrist-worn devices, such as smartwatches and wristbands. Equipped with various sensors, these are capable of capturing digital physiological markers of the user. Actigraphs are one type of wristband commonly used in psychiatric research, which can provide an objective method of observing depression by recording the user's physical activity using accelerometers. Recently, several studies were conducted using commercial wearables such as smartwatches and smart wristbands for screening and assessment of depression.
illustration not visible in this excerpt
3 Methodology
To determine how motor activity data can be used for automatic depression detection, a secondary data analysis was conducted. Figure 3.1 shows the applied methods in this study. The methodological workflow includes five major steps:
1. Data collection: Data was obtained from the Depresjon dataset.
2. Data preprocessing: This step includes the selection of samples and subjects from the original dataset, standardization of data, and removal of missing cases.
3. Feature extraction: A collection of statistical features was extracted for classification.
4. Classification analysis: The statistical features were classified using the random forest method. Oversampling approaches including random oversampling and ADASYN were used to balance the class distribution.
5. Validation: This step includes evaluation of the results based on the corresponding confusion matrix as well as measuring the receiver operating characteristic and the corresponding area under the curve (AUC) value.
illustration not visible in this excerpt
Figure 3.1: Flowchart of the applied methodology (own illustration based on177 ). Blue squares refer to the methodical step in the investigation of this thesis, while gray squares include detailed task descriptions in each step.
Python (version 3.9.12) is a general-purpose programming language capable of statistical calculations and visualizations67 and was used for the analysis in this master thesis. The libraries required for this analysis are NumPy (version 1.21.5)66, Pandas (version 1.4.2)97, SciPy (version 1.7.3)168, Matplotlib (version 3.5.1)74, and Scikit-learn (version 1.1.1)120.
3.1 Data collection
In terms of the research question, a reanalysis of records of motor activity obtained from an observational cohort study was performed. This study assessed variations in motor activity patterns in depressed, schizophrenic, and healthy control participants10. Since the research question of this thesis addresses depression, but not schizophrenia, the presentation of the primary study is limited to depressed patients and the control group. An actigraph was used to monitor motor activity, and data were collected over a period of around two weeks. It was found that motor activity was significantly lower in depressed subjects compared to the control group. Table 3.1 shows characteristics of the depressed patients and healthy controls. Control subjects were ranging in age from 21 to 66 years and consisted of hospital staff (n = 23), students (n = 5), and primary care patients without psychiatric symptoms (n = 4). There was no history of affective or psychotic symptoms among control subjects.
Table 3.1: Demographic characteristics ofcondition and control subjects
illustration not visible in this excerpt
Table 3.2 shows clinical characteristics ofdepressed patients. 15 ofthe 23 MDD and BD patients were taking antidepressants, some were also taking lithium, mood stabilizers, or antipsychotics, while eight were not taking any10. Inpatients were treated at Haukeland University Hospital (Bergen, Norway). Diagnostic assessments of depressive patients was performed using a semi-structured interview and DSM-IV1 criteria, and depressive symptoms were assessed by MADRS scores (Section 2.2.2). Motor activity was measured with an actigraph watch worn on the right wrist. The actigraph in use was the Actiwatch AW4 (Cambridge Neurotechnology Ltd.). It records activity using a piezoelectric accelerometer (Section 2.3.2) that measures the intensity, amount, and duration of movement in all directions. Movements greater than 0.05 g are recorded at a sampling rate of 32 Hz. The sampling rate of accelerometers is the frequency at which they acquire data. An accelerometer with a sampling rate of 32 Hz measures motion 32 times per second. A corresponding voltage is generated as an activity count in the actigraph's memory unit. The number of counts is determined by the intensity of the movement. Total activity was recorded in one-minute intervals. The right wrist was chosen because most participants wear their watches on their left wrist and it is impractical to have two devices on the same arm177. Other studies have found insignificant differences between the left and right wrist166. Both patients and control subjects were instructed to wear their actigraphs at all times except when showering10.
Table 3.2: Demographic and clinical characteristics ofcondition subjects
illustration not visible in this excerpt
All data are given as mean (standard deviation) unless otherwise specified.
Total activity and night-time activity of conditions and controls were compared. Other statistical parameters used to analyze the actigraphy data were inter-day stability (IS) and intra-day variability (IV). IS measures the consistency of motor activity between days, whereas IV measures the frequency and magnitude of transitions between rest and activity. All parameters were calculated for the entire measurement period. There was a significant decrease in total activity (35%less than in the control group) and night-time activity (31% less) in the depressed subjects, whereas there was little difference in IS (4% less). In the condition group, IV was 8% lower. No differences between BD and MDD patients were found in the parameters examined. The results of the comparison of biological males and females in the condition and control groups were similar. In addition, no significant relationships were found between MADRS scores and actigraphy motor parameters10.
Access to the Depresjon dataset is available under the Creative Commons Zero (CC0) license. This dataset is available at http://datasets.simula.no/depresjon/ or can be downloaded directly from http://doi.org/10.5281/zenodo.1219550. It is divided into two folders, one containing the data for the controls and the other containing the data for the conditions. The actigraphy data are provided in a separate file for each subject. In addition, the MADRS scores of the conditions at the beginning and at the end of the study are provided in a separate file. This file also contains data on the number of days of measurement, biological gender (female/male), age, type of affective disorder (MDD/BD-I/BD-II), place of treatment (outpatient or inpatient), education in years, as well as marital status and current work status for each subject in the condition group.
3.1.1 Rating scale data
MADRS scores of condition subjects were assessed at the beginning and at the end of the study (Figure 3.2). For 15 patients2, the score decreased during the observation course, indicating improvements in depressive symptoms. 4 patients had no change in MADRS score, while another 4 patients had higher scores in the end compared to the beginning of observation. Thus, there were improvements in the MADRS score in about two-thirds of the conditions, with a value of - 13 points as the largest observed decrease. The question arises to what extent this improvements might be related to demographic categories or the type of affective disorder. Striking patterns associated with the development of depressive symptoms and specific demographic and/or clinical categories might allow conclusions about the motor activity behavior of specific groups of depressed patients.
illustration not visible in this excerpt
Figure 3.2: MADRS development (own illustration). For each patient, the difference in MADRS score between the start and end of observation is shown. During the observation period, the score decreased in most patients, indicating improvement in their depressive symptoms. Subject 2 showed the most favorable development in this regard. Subjects 1, 5, 9, and 11 showed no alteration, while subjects 3, 7, 15, and 16 showed unfavorable development.
MADRS development was further examined in relation to marital status and biological gender (Figure 3.3). In this context, the average MADRS development is the largest in the group of single males. This can be explained in part by the fact that the patient with the most significant development was among the single males. Apart from this subject, the MADRS development for single males is comparable to that of the other groups. However, it is worth noting that the two subjects with the largest unfavorable MADRS development were one married female patient and one married male patient, respectively. Regarding this, the group of single female patients is the only group with no outliers. Overall, the MADRS development in the marital status and biological gender categories does not reveal any particularly striking deviations. In addition, properties of a normal distribution were found in all four categories.
illustration not visible in this excerpt
Figure 3.3: MADRS between marital status and biological gender (own illustration). Black dots reflect individual subjects, while colored areas represent kernel density estimations for data smoothing. White dots represent the respective median and surrounding bars represent respective inter quartile ranges. Among the married subjects, one male and one female had the least favorable MADRS development, whereas among the single males was the subject with the most favorable MADRS progression. However, in the majority of patients, development is in the interval [-5, 0].
Similarly, the development of the MADRS can be assessed in relation to the patients' affective disorder (Figure 3.4). All patients' scores were between 10 and 30 points each, indicating a moderate depressive episode (Section 2.2.2). There is a visual association between the level of the score in the beginning (MADRS1) and in the end of the observation (MADRS2). Patients with a relatively high MADRS1 score had relatively high MADRS2 scores, and vice versa. The Pearson correlation was used to statistically calculate this relationship147. The Pearson correlation coefficient, which can range from -1 to 1, is a measure of the degree of linear relationship between two variables. A value of -1 represents a completely negative linear correlation, while a value of 1 indicates a completely positive linear correlation. This coefficient is calculated as
illustration not visible in this excerpt
with r as the correlation coefficient, x t as the observed value of x, y t as the observed value of y, μ χ as the arithmetic mean of all observed values of x, y y as the arithmetic mean of all observed values of y, and n as the total number of all observations. Assuming x and y as and MADRS2, r = 0 . 65. Thus, there is a significant positive correlation between MADRS1 and MADRS2 across all three types of affective disorders. This provides empirical suggestion that MADRS development follows a similar trajectory across patients, regardless of MADRS score and type of affective disorder.
Linear regression is another method of statistical analysis170. A linear regression model predicts the value of a dependent variable y based on a given independent variable x. Thus, linear
illustration not visible in this excerpt
Figure 3.4: MADRS between affection types (own illustration). The association between MADRS score in the beginning and end of observation for each patient is shown. The optimal-fit regression line illustrates that regardless of the type of affective disorder, there is a positive association between the level of the score in the beginning and in the end of observation (own illustration).
regression can discover a linear relationship between x and y. Using the hypothesis function y = ß1 + ß2 x x, the linear regression model finds the optimal regression fit line by optimizing ß1 and ß2. By finding the optimal regression fit line, the model attempts to predict y with the goal of minimizing the error difference between the predicted and actual values. MADRS1 is represented by x in Figure 3.4, while MADRS2 is represented by y. The best-fit regression line shows an upward trend: higher MADRS1 values are associated with higher MADRS2 values and vice versa. From the regression analysis, there is a significant linear relationship between MADRS1 and MADRS2. This relationship can be observed equally in MDD patients and BD-II patients. There is only one BD-I patient in the data set, making regression analysis impossible for this group. Still, it can be noted that with regard to the linear relationship in the development of the MADRS score, no distinction can be made between bipolar depression and unipolar MDD. In both groups, the regression analysis for a MADRS1 value that is one point higher results in a MADRS2 value that is approximately 0.64 points higher. Accordingly, a significant linear relationship between MADRS1 and MADRS2 can be observed by linear regression, similarly strong as with the Pearson correlation. Thus, it can be said that the degree of a patient's subjective depressive symptomatology at the end of observation is primarily dependent on the degree of the patient's subjective depressive symptomatology at the beginning of observation, and the effects of treatment or observation during the course of the study are secondary. Consequently, possible effects of treatment success on a patient's motor activity pattern may be disregarded, and no prior MADRS assessment is necessary to classify an individual based on motor activity. Considering these aspects, the MADRS assessments were not included in the classification analysis.
illustration not visible in this excerpt
3.1.2 Motor activity data
To answer the research question, the activity data of the subjects was examined. The motor activity over the observation course (time series) of a control subject in Figure 3.5 gives a feasible example to explain how the measurement of activity was conducted. In this example, the measurement was performed for 14 full days. Potential movement was sampled 32 times per second by the actigraph, and the intensity of motor activity was recorded every 60 seconds. A higher intensity corresponds to a higher value for the activity attribute. Each observation for a 60-second interval corresponds to one point in the diagram. Each mark on the horizontal axis represents 0 a.m. of the given day. It is apparent that the recorded motor activity is associated with recurring patterns. In accordance with circadian rhythms, first of all the sleep-wake cycle, the intensity is comparatively high during the day, while at night it is comparatively low. At certain times, activity was particularly high, as shown by the upward outliers. With these outliers, an activity value of 4 000 or more was observed.
illustration not visible in this excerpt
Figure 3.5: Activity time series of a control subject (own illustration). Using the motor activity time series of control group subject 14 as an example, it is possible to see how the recorded activity fluctuates during the observation time in daily recurring high and low periods.
In comparison with a depressed patient (Figure 3.6), the differences in activity between depressed and non-depressed subjects are immediately recognizable. The recurring patterns that are visually detectable in the control are barely visible in the condition. Disturbed circadian rhythms play a significant role in association with depression, and the time series illustrates how disturbed rhythms and motor activity are related. Thus, daytime activity does not appear to be significantly different from nighttime activity. In addition, the maximum activity per day differs significantly from the control. As such, the upward outliers that were observed for almost every day in the control are not seen in the condition. The observations with the highest values are about 1 500 for the condition which is a significantly lower value than for the control. Figure 3.6 shows the recorded time series of motor activity of the only BD-I patient in the study.
[...]
1 Depresjon is the Norwegian term for depression.
2 Prevalence indicates what percentage of the population is affected by a disease (in this case, over a time period of one year).
3 The prevalence at a given time, such as a specific date, is called the point prevalence.
4 A disposition induces that the chance of having a specific disease is increased.
5 Neurotransmitters are biochemical compounds that transmit impulses from one nerve cell to another.
6 Serotonin is a neurotransmitter with mood-lifting and stress-reducing effects.
7 Embedded systems, as opposed to general-purpose computers, are cyber-physical systems specifically designed to perform a particular function. Cyber-physical systems describe digital functionalities embedded in physical objects.
8 Recall bias is an important issue in (self)reporting and refers to systematic inaccuracies that occur when individuals do not remember past events or omit facts.
1 The Diagnostic and Statistical Manual of Mental Disorders (DSM) is a system for classifying psychiatric conditions.
2 Whether these were the same 15 patients who were taking medication is not specified in the article.
- Quote paper
- Tim Kreutzmann (Author), 2022, The Potential of Wearables to Automatically Detect Depression, Munich, GRIN Verlag, https://www.grin.com/document/1363675
-
-
-
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X. -
-
Upload your own papers! Earn money and win an iPhone X.