Development and Feasibility Study of HOPE Model for Prediction of Depression Among Older Adults Using Wi-Fi-based Motion Sensor Data: Machine Learning Study (Preprint)

Published: Oct. 18, 2024

Depression, characterized by persistent sadness and loss of interest in daily activities, greatly reduces quality life. Early detection is vital for effective treatment intervention. While many studies use wearable devices to classify depression based on physical activity, these often rely intrusive methods. Additionally, most classification involve large participant groups single-stage classifiers without explainability. This study aims assess the feasibility classifying using nonintrusive Wi-Fi–based motion sensor data a novel machine learning model limited number participants. We also conduct an explainability analysis interpret model’s predictions identify key features associated with classification. In this study, we recruited adults aged 65 years older through web-based in-person methods, supported McGill University health care facility directory. Participants provided consent, collected 6 months activity sleep via sensors, along Edmonton Frailty Scale Geriatric Depression data. For classification, proposed HOPE (Home-Based Older Adults’ Prediction) feature selection, dimensionality reduction, stages, evaluating various combinations accuracy, sensitivity, precision, <i>F</i>₁-score. Shapely addictive explanations local interpretable model-agnostic were used explain predictions. A total participants enrolled study; however, 2 withdrew later due internet connectivity issues. Among 4 remaining participants, 3 classified as not having depression, while 1 was identified depression. The accurate model, which combined sequential forward selection principal component decision tree achieved accuracy 87.5%, sensitivity 90%, precision 88.3%, effectively distinguishing individuals those revealed that influential order importance, “average duration,” “total interruptions,” “percentage nights duration “Edmonton Scale.” findings from preliminary demonstrate sensors highlight effectiveness our even small sample size. These results suggest potential further research larger cohort more comprehensive validation. collection method architecture offer promising applications remote monitoring, particularly who may face challenges devices. Furthermore, importance patterns aligns previous research, emphasizing need in-depth role mental health, suggested explainable study.

Language: Английский

Predicting Adolescent Female Stress with Wearable Device Data Using Machine and Deep Learning

Published: Oct. 9, 2023

The prevalence of mental health issues in adolescent females has become a significant concern recent years. To investigate the potential wearable biosensors predicting stress responses this understudied demographic, we collected wearables data from eight teenage girls over 1-4 months and explored prediction using several machine learning (ML) deep (DL) models. Various person-dependent person-independent schemes, feature extraction methods, classifier types were systematically investigated to provide recommendations for effective prediction. Feature importance physiological signals was also analyzed insights into responses. study provides actionable classifiers, extraction, personalization schemes enhance accuracy, enhancing understanding early detection females.

Language: Английский

Towards Personalised Depression Modelling and Explanation from Wearable Data

Research Square (Research Square), Journal Year: 2023, Volume and Issue: unknown

Published: June 9, 2023

Abstract Depression and anxiety are the leading causes of health loss globally, Covid-19 pandemic has significantly exacerbated effect these disorders. There is a widening gap between available resources mental needs globally. Digital applications using artificial Intelligence (AI) promising opportunity to address this gap. Increasingly, passively acquired data from wearables augmented with carefully selected active participants develop machine learning (ML) models depression. However, ML black-box in nature, hence outputs not explainable. also multi-modal, reasons for depression may vary individuals. Explainable personalised will thus be beneficial clinicians determining main features that lead decline mood state patient, enabling suitable therapy. This currently lacking. Therefore, study presents first methodology developing accurate deep (DL)-based depression, along novel methods identifying key facets exacerbation depressive symptoms. We illustrate our approach an existing multi-modal dataset containing longitudinal ecological momentary assessments lifestyle wearables, neurocognitive 14 mild moderately depressed over one month. train classification- regression-based DL predict participants’ scores - discrete score given participant based on severity their The trained inside eight different evolutionaryalgorithm-based optimisation schemes optimise model parameters maximum predictive performance. A 5-fold cross-validation scheme used verify performance, error as low 6% some participants. use best process extract indicators, SHAP, ALE Anchors AI literature explain why certain predictions made how they affect mood. These feature insights can assist professionals incorporating interventions into patient’s treatment regimen.

Language: Английский