Develop a Personalized Mental Health Assessment based on Keystroke Dynamics and Behavioural Characteristics

Abstract

Depressive Disorder is a leading cause of disability worldwide. Passive screening tools for detecting the symptoms of depression are essential for monitoring and limiting the spread of the disease. Remarkably, individuals’ kinetic expression and activities, including their interaction with touchscreen smartphones, can reveal their mental status. Such widely available data in everyday life form a promising source of information on keystroke dynamics and their characteristics. Therefore, studying these typing patterns can assist in developing effective screening tools for depressive disorder. This research explores how keystroke dynamics derived from touchscreen typing patterns have revealed the diagnosis of mental disorders, particularly depressive disorders. Moreover, different learning-based models were established to detect the subjects’ depressive tendencies and predict the severity of the disease, as indicated by the self-administered Patient Health Questionnaire-9 (PHQ-9) score, based on keystroke digital biomarkers. The keystroke timing sequences were collected unobtrusively during routine interaction with touchscreen smartphones in a non-clinical setting. This study utilized 23,264 typing sessions provided by 24 subjects with various depression severity levels. The proposed framework was investigated under two keystroke feature combinations, namely the hold time and flight time keystroke variables, and validated using a nested cross-validation scheme. Different feature engineering techniques were employed to extract statistical features from the keystroke time sequences and feed the machine learning pipeline. In contrast, the proposed deep learning models were trained and tested using the raw keystroke time sequences. The best-performing depression detection pipeline achieved an Area Under Curve (AUC) of 0.97 with the correlated probabilities of the subjects’ status [95% confidence interval (CI):0.89–1.00]. Whereas the best-performing depression severity score prediction model achieved an f1-score of 0.83. Overall, the proposed learning-based models could effectively and dynamically capture depressive tendencies, considering the users’ behavioural characteristics, which would potentially provide the users and their attending physiologists information regarding the evolution of their mental health. This work contributes to the improvement of digital tools for the objective screening of mental disorders in the wild.

Date of Award	Apr 2023
Original language	American English
Supervisor	Leontios Hadjileontiadis (Supervisor)