A Machine Learning Approach for Continuous Mining of Nonidentifiable Smartphone Data to Create a Novel Digital Biomarker Detecting Generalized Anxiety Disorder: Prospective Cohort Study

Abstract

Background Anxiety is one of the leading causes of mental health disability around the world. Currently, a majority of the population who experience anxiety go undiagnosed or untreated. New and innovative ways of diagnosing and monitoring anxiety have emerged using smartphone sensor–based monitoring as a metric for the management of anxiety. This is a novel study as it adds to the field of research through the use of nonidentifiable smartphone usage to help detect and monitor anxiety remotely and in a continuous and passive manner. Objective This study aims to evaluate the accuracy of a novel mental behavioral profiling metric derived from smartphone usage for the identification and tracking of generalized anxiety disorder (GAD). Methods Smartphone data and self-reported 7-item GAD anxiety assessments were collected from 229 participants using an Android operating system smartphone in an observational study over an average of 14 days (SD 29.8). A total of 34 features were mined to be constructed as a potential digital phenotyping marker from continuous smartphone usage data. We further analyzed the correlation of these digital behavioral markers against each item of the 7-item Generalized Anxiety Disorder Scale (GAD-7) and its influence on the predictions of machine learning algorithms. Results A total of 229 participants were recruited in this study who had completed the GAD-7 assessment and had at least one set of passive digital data collected within a 24-hour period. The mean GAD-7 score was 11.8 (SD 5.7). Regression modeling was tested against classification modeling and the highest prediction accuracy was achieved from a binary XGBoost classification model (precision of 73%-81%; recall of 68%-87%; F1-score of 71%-79%; accuracy of 76%; area under the curve of 80%). Nonparametric permutation testing with Pearson correlation results indicated that the proposed metric (Mental Health Similarity Score [MHSS]) had a colinear relationship between GAD-7 Items 1, 3 and 7. Conclusions The proposed MHSS metric demonstrates the feasibility of using passively collected nonintrusive smartphone data and machine learning–based data mining techniques to track an individuals’ daily anxiety levels with a 76% accuracy that directly relates to the GAD-7 scale.

Machine Learning for Multimodal Mental Health Detection: A Systematic Review of Passive Sensing Approaches

“…Generally, they also showed more active application engagement in the early hours or midnight compared to healthy controls, who showed diluted engagement patterns throughout the day. Meanwhile, Choudhary et al [ 212 ] revealed that individuals with anxiety exhibited more frequent usage of applications from “passive information consumption apps”, “games”, and “health and fitness” categories.…”

“…XGBoost and AdaBoost were gradually favored by researchers due to their better predictive performance. Specifically, few studies [ 134 , 158 , 184 , 212 ] revealed XGBoost as the most effective among SVM, RF, K-nearest neighbor, logistic regression, and DNN models. In contrast, researchers also proposed novel hierarchical ensemble architectures by stacking algorithms (e.g., XGBoost [ 194 ], Extreme Learning Machine (ELM) [ 192 ]) into layers where models in subsequent layers receive outputs from previous layers as inputs for ensemble predictions.…”

“…Generally, they also showed more active application engagement in the early hours or midnight compared to healthy controls, who showed diluted engagement patterns throughout the day. Meanwhile, Choudhary et al [212] revealed that individuals with anxiety exhibited more frequent usage of applications from "passive information consumption apps", "games", and "health and fitness" categories. (3) Sociability Features: Sociability features, such as the number of incoming/outgoing phone calls and text messages and the duration of phone calls, were also potential indicators of MH disorders [164,175].…”

“…ElasticNet regression [437], stochastic gradient descent (SGD) regression, Gaussian staircase model, partial least square (PLS) [438] regression (useful for collinear features), generalized linear models Learn relationship between features to predict continuous values (scores of assessment scales) or probabilities (correspond to output classes) [20,43,49,53,55,68,70,99,104,105,126,130,134,140,150,154,163,164,167,175,179,182,188,200,[211][212][213]219,222,223] SVM Find a hyperplane that best fits features (regression) or divides features into classes (classification), secondary model in score-level fusion [47,50,79,99,104,105,115,121,130,134,140,148,162,163,…”

Multidimensional digital biomarker of cognitive health: unobtrusive and continuous monitoring of cognitive changes using smartphones

“…Accelerometer data, app use, number of text messages and phone calls, and GPS-derived measures (number of distinct locations visited, number of hours outside, distance traveled) were used to detect depressive and manic episodes in bipolar patients with high precision and recall rates 62,63 . In a recent study 64 , a new metric, the Mental Health Similarity Score, derived from passively collected smartphone app use data, for tracking individual-level daily anxiety levels, was shown to correlate with the 7-item Generalized Anxiety Disorder Scale 65 . App use data were also used in the context of sleep, mood, and stress 57,[66][67][68] .…”