Digital mental health care: from smart sensing and machine learning classifiers to digital therapy

Abstract

Background: Mental disorders cause substantial personal and societal burdens worldwide. To lower the burden, timely and reliable diagnosis and effective interventions are essential. The ongoing digitalization in daily living and the development of novel technological and methodological approaches may provide an opportunity to augment current clinical practices from screening and assessment to treatment. For instance, supervised machine learning prediction models could improve the accuracy of screening instruments for mental disorders, or the omnipresence of sensor data may enable novel paradigms (e.g., smart sensing) to unobtrusively collect objective data and overcome current biases in the assessment of psychopathology. Furthermore, the Internet and mobile applications could provide a location- and time-independent delivery format for interventions. Research is needed to explore the feasibility and effectiveness of these approaches. Objective: Contributing to this, the present thesis focused particularly on depression as the mental disorder causing the highest disease burden worldwide and investigated the research questions: 1) Can supervised machine learning models improve screening for major depressive episodes? 2) Can smart sensing be used to infer depression severity? 3) What are the heterogeneity, magnitude, and moderators of the effects of Internet- and mobile-based interventions (IMI) against depression? 4) What is the quality of commonly available mobile health apps, and how can quality be measured? 5) What are the acceptance levels and their predictors of IMI and 6) smart sensing with a focus on the effects of an acceptance facilitating intervention (AFI)? Structure: To answer these research questions, study I compared supervised machine learning models for predicting major depressive episodes based on commonly used screening instruments to their best-practice scoring procedures (i.e., cut-offs for sum scores). Studies II to IV explored the potential of sensor data for depression inference in 1) a systematic review and meta-analysis of correlations between location features and depression (study II), 2) an observation study with regression models determining the explained variance of depression by smart sensing and by ecological momentary assessment (EMA) (study III), and 3) a chapter conceptualizing and reviewing smart sensing enhanced expert systems for clinical practice (study IV). Concerning question 3, the variance ratios and standardized mean differences between IMI and control groups in randomized controlled trials were investigated in a systematic review with a three-level Bayesian meta-regression analysis (study V). In an international individual participant data (IPD) analysis, study VI explored the evidence and quality of mobile health applications in the app markets using the Mobile Application Rating Scale (MARS) alongside the metric properties of the MARS. To answer the fifth question, an IPD analysis (study VII) was used to determine the acceptance of IMI and its predictors based on the Unified Theory of Acceptance and Use of Technology (UTAUT). Similarly, the acceptance of smart sensing and its predictors were investigated in study VIII, but with a primary focus on the effectiveness of a video-based AFI following a randomized controlled trial design. Results: Study I found that supervised machine learning models predicting major depressive episodes can significantly outperform scoring procedures of some screening instruments (i.e., QIDS C 16: ΔAUC = 0.04, 95%-CI: 0.02 to 0.05, p < .001; PHQ-9: ΔAUC = 0.01, 95%-CI: 0.00 to 0.02, p = .009). Also, smart sensing seems to be suited to inferring depression: In study II, meta-analytically robust correlations between depression and location features were found (e.g., distance: r = -0.25, 95 %-CI: -0.29 to 0.21) alongside risk for publication bias and poor adherence to international reporting standards in the field. The results of study III underlined that smart sensing can substantially explain variance in depression as a stand-alone assessment (adj. R2 = 20.45 %, 95 %-CI: 7.81 % to 35.59 %) and as an augmentation to existing approaches, specifically EMA (adj. R2 = 45.15 %, 95 %-CI: 30.39 % to 58.53 %; Δadj.R2 to smart sensing only: 24.70 %; Δadj.R2 to ema only: 9.87 %). While finding a lack of evidence for smart sensing-enhanced expert systems in clinical settings, study IV outlined promising application areas from diagnoses to just-in-time-adaptive interventions. Investigating digital treatments in mental health, study V found strong evidence for the effectiveness of IMI (g = -0.56, 95%-CrI: -0.46 to 0.66) without substantial treatment-by-patient interactions overall (lnVR = -0.02, 95%-CrI: 0.07 to 0.03). However, sensitivity analysis revealed important insights into the role of guidance and baseline severity. In contrast to the profound evidence found for IMI in the academic literature, study VI showed poor evidence for mobile health apps available in the app stores (< 5 % evaluated) and an overall moderate quality (M = 3.74, SD = 0.59). Good metric properties were found for the MARS. The acceptance of IMI (study VII) was moderate (M = 2.82, SD = 1.12) and low (35.41 %) to moderate (47.92 %) for smart sensing (study VIII). Modified versions of the UTAUT were replicated with performance expectancy as the main predictor of acceptance in both studies. The AFI in Study VIII did not significantly influence the acceptance (p = .357). Discussion: The present research project highlights supervised machine learning and smart sensing enhanced screening and assessment, as well as IMI, as promising ways to improve mental health care from diagnosis to treatment. However, for supervised machine learning and smart sensing applications, future confirmatory studies in clinical settings are strongly needed before the implementation in clinical practice can be recommended. Conversely, the meta-analytically shown efficacy and effectiveness of IMI provide a convincing rationale to include IMI in mental health care. However, baseline severity and the involvement of guidance need to be considered when recommendations are made. Furthermore, the lack of evidence and quality of interventions available in common app markets strongly calls for systematic quality assurance platforms to guide users and caregivers to adequate interventions. Lastly, to maximize the potential impact of digital innovations in the screening and treatment of mental health, acceptance levels pose a major barrier and make acceptance facilitating and implementation strategies a prime target for future research.