Detecting major depressive disorder presence using passively-collected wearable movement data in a nationally-representative sample.

Major Depressive Disorder (MDD) is a heterogeneous disorder, resulting in challenges with early detection. However, changes in sleep and movement patterns may help improve detection. Thus, this study aimed to explore the utility of wrist-worn actigraphy data in combination with machine learning (ML) and deep learning techniques to detect MDD using a commonly used screening method: Patient Health Questionnaire-9 (PHQ-9). Participants (N = 8,378; MDD Screening = 766 participants) completed the and wore Actigraph GT3X+ for one week as part of the National Health and Nutrition Examination Survey (NHANES). Leveraging minute-level, actigraphy data, we evaluated the efficacy of two commonly used ML approaches and identified actigraphy-derived biomarkers indicative of MDD. We employed two ML modeling strategies: (1) a traditional ML approach with theory-driven feature derivation, and (2) a deep learning Convolutional Neural Network (CNN) approach, coupled with gramian angular field transformation. Findings revealed movement-related features to be the most influential in the traditional ML approach and nighttime movement to be the most influential in the CNN approach for detecting MDD. Using a large, nationally-representative sample, this study highlights the potential of using passively-collected, actigraphy data for understanding MDD to better improve diagnosing and treating MDD.

Depressive symptoms as a heterogeneous and constantly evolving dynamical system: Idiographic depressive symptom networks of rapid symptom changes among persons with major depressive disorder.

Major depressive disorder (MDD) is conceptualized by individual symptoms occurring most of the day for at least two weeks. Despite this operationalization, MDD is highly variable with persons showing greater variation within and across days. Moreover, MDD is highly heterogeneous, varying considerably across people in both function and form. Recent efforts have examined MDD heterogeneity byinvestigating how symptoms influence one another over time across individuals in a system; however, these efforts have assumed that symptom dynamics are static and do not dynamically change over time. Nevertheless, it is possible that individual MDD system dynamics change continuously across time. Participants (N = 105) completed ratings of MDD symptoms three times a day for 90 days, and we conducted time varying vector autoregressive models to investigate the idiographic symptom networks. We then illustrated this finding with a case series of five persons with MDD. Supporting prior research, results indicate there is high heterogeneity across persons as individual network composition is unique from person to person. In addition, for most persons, individual symptom networks change dramatically across the 90 days, as evidenced by 86% of individuals experiencing at least one change in their most influential symptom and the median number of shifts being 3 over the 90 days. Additionally, most individuals had at least one symptom that acted as both the most and least influential symptom at any given point over the 90-day period. Our findings offer further insight into short-term symptom dynamics, suggesting that MDD is heterogeneous both across and within persons over time. (PsycInfo Database Record (c) 2024 APA, all rights reserved).

Using digital phenotyping to capture depression symptom variability: detecting naturalistic variability in depression symptoms across one year using passively collected wearable movement and sleep data.

Major Depressive Disorder (MDD) presents considerable challenges to diagnosis and management due to symptom variability across time. Only recent work has highlighted the clinical implications for interrogating depression symptom variability. Thus, the present work investigates how sociodemographic, comorbidity, movement, and sleep data is associated with long-term depression symptom variability. Participant information included (N = 939) baseline sociodemographic and comorbidity data, longitudinal, passively collected wearable data, and Patient Health Questionnaire-9 (PHQ-9) scores collected over 12 months. An ensemble machine learning approach was used to detect long-term depression symptom variability via: (i) a domain-driven feature selection approach and (ii) an exhaustive feature-inclusion approach. SHapley Additive exPlanations (SHAP) were used to interrogate variable importance and directionality. The composite domain-driven and exhaustive inclusion models were both capable of moderately detecting long-term depression symptom variability (r = 0.33 and r = 0.39, respectively). Our results indicate the incremental predictive validity of sociodemographic, comorbidity, and passively collected wearable movement and sleep data in detecting long-term depression symptom variability.

Depression deconstructed: Wearables and passive digital phenotyping for analyzing individual symptoms.

Wearable technology enables unobtrusive collection of longitudinally dense data, allowing for continuous monitoring of physiology and behavior. These digital phenotypes, or device-based indicators, are frequently leveraged to study depression. However, they are usually considered alongside questionnaire sum-scores which collapse the symptomatic gamut into a general representation of severity. To explore the contributions of passive sensing streams more precisely, associations of nine passive sensing-derived features with self-report responses to Center for Epidemiologic Studies Depression (CES-D) items were modeled. Using data from the NetHealth study on N=469 college students, this work generated mixed ordinal logistic regression models to summarize contributions of pulse, movement, and sleep data to depression symptom detection. Emphasizing the importance of the college context, wearable features displayed unique and complementary properties in their heterogeneously significant associations with CES-D items. This work provides conceptual and exploratory blueprints for a reductionist approach to modeling depression within passive sensing research.

Leveraging deep learning models to understand the daily experience of anxiety in teenagers over the course of a year.

INTRODUCTION: Anxiety disorders are a prevalent and severe problem that are often developed early in life and can disrupt the daily lives of affected individuals for many years into adulthood. Given the persistent negative aspects of anxiety, accurate and early assessment is critical for long term outcomes. Currently, the most common method for anxiety assessment is through point-in-time measures like the GAD-7. Unfortunately, this survey and others like it can be subject to recall bias and do not fully capture the variability in an individual's day-to-day symptom experience. The current work aims to evaluate how point-in-time assessments like the GAD-7 relate to daily measurements of anxiety in a teenage population. METHODS: To evaluate this relationship, we leveraged data collected at four separate three week intervals from 30 teenagers (age 15-17) over the course of a year. The specific items of interest were a single item anxiety severity measure collected three times per day and end-of-month GAD-7 assessments. Within this sample, 40 % of individuals reported clinical levels of generalized anxiety disorder symptoms at some point during the study. The first component of analysis was a visual inspection assessing how daily anxiety severity fluctuated around end-of-month reporting via the GAD-7. The second component was a between-subjects comparison assessing whether individuals with similar GAD-7 scores experienced similar symptom dynamics across the month as represented by latent features derived from a deep learning model. With this approach, similarity was operationalized by hierarchical clustering of the latent features. RESULTS: The aim clearly indicated that an individual's daily experience of anxiety varied widely around what was captured by the GAD-7. Additionally, when hierarchical clustering was applied to the three latent features derived from the (LSTM) encoder (r = 0.624 for feature reconstruction), it was clear that individuals with similar GAD-7 outcomes were experiencing different symptom dynamics. Upon further inspection of the latent features, the LSTM model appeared to rely as much on anxiety variability over the course of the month as it did on anxiety severity (p < 0.05 for both mean and RMSSD) to represent an individual's experience. DISCUSSION: This work serves as further evidence for the heterogeneity within the experience of anxiety and that more than just point-in-time assessments are necessary to fully capture an individual's experience.

An unsupervised machine learning approach using passive movement data to understand depression and schizophrenia.

INTRODUCTION: Schizophrenia and Major Depressive Disorder (MDD) are highly burdensome mental disorders, with significant cost to both individuals and society. Despite these disorders representing distinct clinical categories, they are each heterogenous in their symptom profiles, with considerable transdiagnostic features. Although movement and sleep abnormalities exist in both disorders, little is known of the precise nature of these changes longitudinally. Passively-collected longitudinal data from wearable sensors is well suited to characterize naturalistic features which may cross traditional diagnostic categories (e.g., highlighting behavioral markers not captured by self-report information). METHODS: The present analyses utilized raw minute-level actigraphy data from three diagnostic groups: individuals with schizophrenia (N = 23), individuals with depression (N = 22), and controls (N = 32), respectively, to interrogate naturalistic behavioral differences between groups. Subjects' week-long actigraphy data was processed without diagnostic labels via unsupervised machine learning clustering methods, in order to investigate the natural bounds of psychopathology. Further, actigraphic data was analyzed across time to determine timepoints influential in model outcomes. RESULTS: We find distinct actigraphic phenotypes, which differ between diagnostic groups, suggesting that unsupervised clustering of naturalistic data aligns with existing diagnostic constructs. Further, we found statistically significant inter-group differences, with depressed persons showing the highest behavioral variability. LIMITATIONS: However, diagnostic group differences only consider biobehavioral trends captured by raw actigraphy information. CONCLUSIONS: Passively-collected movement information combined with unsupervised deep learning algorithms shows promise in identifying naturalistic phenotypes in individuals with mental health disorders, specifically in discriminating between MDD and schizophrenia.

Predicting symptom response and engagement in a digital intervention among individuals with schizophrenia and related psychoses.

Introduction: Despite existing work examining the effectiveness of smartphone digital interventions for schizophrenia at the group level, response to digital treatments is highly variable and requires more research to determine which persons are most likely to benefit from a digital intervention. Materials and methods: The current work utilized data from an open trial of patients with psychosis (N = 38), primarily schizophrenia spectrum disorders, who were treated with a psychosocial intervention using a smartphone app over a one-month period. Using an ensemble of machine learning models, pre-intervention data, app use data, and semi-structured interview data were utilized to predict response to change in symptom scores, engagement patterns, and qualitative impressions of the app. Results: Machine learning models were capable of moderately (r = 0.32-0.39, R2 = 0.10-0.16, MAE norm = 0.13-0.29) predicting interaction and experience with the app, as well as changes in psychosis-related psychopathology. Conclusion: The results suggest that individual smartphone digital intervention engagement is heterogeneous, and symptom-specific baseline data may be predictive of increased engagement and positive qualitative impressions of digital intervention in patients with psychosis. Taken together, interrogating individual response to and engagement with digital-based intervention with machine learning provides increased insight to otherwise ignored nuances of treatment response.

Association of Selective Serotonin Reuptake Inhibitor Use With Abnormal Physical Movement Patterns as Detected Using a Piezoelectric Accelerometer and Deep Learning in a Nationally Representative Sample of Noninstitutionalized Persons in the US.

Importance: Selective serotonin reuptake inhibitors (SSRIs) are a common first-line treatment for some psychiatric disorders, including depression and anxiety; although they are generally well tolerated, SSRIs have known adverse effects, including movement problems, sleep disruption, and gastrointestinal problems (eg, nausea and upset stomach). No large-scale studies using naturalistic, longitudinal, objective data have validated physical activity findings, and actigraphy data are well suited to address this task. Objectives: To evaluate whether differences in physical movement exist among individuals treated with SSRIs compared with control participants and to identify the unique features of the movement of patients treated with SSRIs. Design, Setting, and Participants: This cross-sectional study examines longitudinally collected wearable movement data within a cross-sectional sample of 7162 participants from the 2005-2006 National Health and Nutrition Examination Survey (NHANES), a nationally representative population-based sample of noninstitutionalized persons in the US having both medication information and passive movement data. Statistical analysis was performed from April 1, 2021, to February 1, 2022. Exposures: The use of SSRIs (sertraline hydrochloride, escitalopram oxalate, fluoxetine hydrochloride, paroxetine hydrochloride, and citalopram hydrobromide), as reported by participants interviewed by NNHANES personnel, was the primary exposure, measured as a binary variable (taking an SSRI vs not taking an SSRI). Main Outcomes and Measures: The primary outcome was the intensity of body movement as recorded by a piezoelectric accelerometer worn on the right hip for more than 1 week. Results: Of the 7162 participants included in the study, the mean (SD) age was 33.7 (22.6) years, 266 (3.7%) were taking an SSRI, 3706 (51.7%) were female, 1934 (27.0%) were Black, 1823 (25.5%) were Mexican American, 210 (2.9%) were other Hispanic, 336 (4.7%) were other or multiracial, and 2859 (39.9%) were White (per the NHANES data collection protocol). A cross-validated, deep learning classifier was constructed that achieved fair performance predicting SSRI use (area under the curve, 0.67 [95% CI, 0.64-0.71] for the validation set and 0.66 [95% CI, 0.64-0.68] for the test set). To account for possible confounding by indication, we constructed a parallel model incorporating depression severity, finding only marginal performance improvement. When averaged across individuals and across 7 days, the results show less overall movement in the SSRI group (mean, 120.1 vertical acceleration counts/min [95% CI, 115.7-124.6 vertical acceleration counts/min]) compared with the non-SSRI control group (mean, 168.8 vertical acceleration counts/min [95% CI, 162.8-174.9 vertical acceleration counts/min]). Conclusions and Relevance: This cross-sectional study found a moderate association between passive movement and SSRI use, as well as SSRI detection capacity of passive movement using time series deep learning models. The results support the use of passive sensors for exploration and characterization of psychotropic medication adverse effects.

Using Smartphone App Use and Lagged-Ensemble Machine Learning for the Prediction of Work Fatigue and Boredom.

INTRO: As smartphone usage becomes increasingly prevalent in the workplace, the physical and psychological implications of this behavior warrant consideration. Recent research has investigated associations between workplace smartphone use and fatigue and boredom, yet findings are not conclusive. METHODS: To build off recent efforts, we applied an ensemble machine learning model on a previously published dataset of N = 83 graduate students in the Netherlands to predict work boredom and fatigue from passively collected smartphone app use information. Using time-based feature engineering and lagged variations of the data to train, validate, and test idiographic models, we evaluated the efficacy of a lagged-ensemble predictive paradigm on sparse temporal data. Moreover, we probed the relative importance of both derived app use variables and lags within this predictive framework. RESULTS: The ability to predict fatigue and boredom trajectories from app use information was heterogeneous and highly person-specific. Idiographic modeling reflected moderate to high correlative capacity (r > 0.4) in 47% of participants for fatigue and 24% for boredom, with better overall performance in the fatigue prediction task. App use relating to duration, communication, and patterns of use frequency were among the most important features driving predictions across lags, with longer lags contributing more heavily to final ensemble predictions compared with shorter ones. CONCLUSION: A lag- specific ensemble predictive paradigm is a promising approach to leveraging high-dimensional app use behavioral data for the prediction of work fatigue and boredom. Future research will benefit from evaluating associations on densely collected data across longer time scales.

Using the COVID-19 Pandemic to Assess the Influence of News Affect on Online Mental Health-Related Search Behavior Across the United States: Integrated Sentiment Analysis and the Circumplex Model of Affect.

BACKGROUND: The digital era has ushered in an unprecedented volume of readily accessible information, including news coverage of current events. Research has shown that the sentiment of news articles can evoke emotional responses from readers on a daily basis with specific evidence for increased anxiety and depression in response to coverage of the recent COVID-19 pandemic. Given the primacy and relevance of such information exposure, its daily impact on the mental health of the general population within this modality warrants further nuanced investigation. OBJECTIVE: Using the COVID-19 pandemic as a subject-specific example, this work aimed to profile and examine associations between the dynamics of semantic affect in online local news headlines and same-day online mental health term search behavior over time across the United States. METHODS: Using COVID-19-related news headlines from a database of online news stories in conjunction with mental health-related online search data from Google Trends, this paper first explored the statistical and qualitative affective properties of state-specific COVID-19 news coverage across the United States from January 23, 2020, to October 22, 2020. The resultant operationalizations and findings from the joint application of dictionary-based sentiment analysis and the circumplex theory of affect informed the construction of subsequent hypothesis-driven mixed effects models. Daily state-specific counts of mental health search queries were regressed on circumplex-derived features of semantic affect, time, and state (as a random effect) to model the associations between the dynamics of news affect and search behavior throughout the pandemic. Search terms were also grouped into depression symptoms, anxiety symptoms, and nonspecific depression and anxiety symptoms to model the broad impact of news coverage on mental health. RESULTS: Exploratory efforts revealed patterns in day-to-day news headline affect variation across the first 9 months of the pandemic. In addition, circumplex mapping of the most frequently used words in state-specific headlines uncovered time-agnostic similarities and differences across the United States, including the ubiquitous use of negatively valenced and strongly arousing language. Subsequent mixed effects modeling implicated increased consistency in affective tone (SpinVA ß=-.207; P<.001) as predictive of increased depression-related search term activity, with emotional language patterns indicative of affective uncontrollability (FluxA ß=.221; P<.001) contributing generally to an increase in online mental health search term frequency. CONCLUSIONS: This study demonstrated promise in applying the circumplex model of affect to written content and provided a practical example for how circumplex theory can be integrated with sentiment analysis techniques to interrogate mental health-related associations. The findings from pandemic-specific news headlines highlighted arousal, flux, and spin as potentially significant affect-based foci for further study. Future efforts may also benefit from more expansive sentiment analysis approaches to more broadly test the practical application and theoretical capabilities of the circumplex model of affect on text-based data.