More than one piece of the puzzle: considering non-clinical factors for personalisation in digital phenotyping

BackgroundDigital phenotyping is an emerging field that aims to contribute to the clinical care of patients with mental disorders by offering objective, passive behavioural monitoring. This monitoring could be used for applications such as predicting the onset of episodes of mental illness. However, behaviours are often not unique to clinical disorders, and other factors in a persons life may contribute to their digital phenotyping behavioural pattern. ObjectiveWe aimed to investigate non-clinical factors that may be relevant for personalisation in digital phenotyping, such as the area in which participants live and their regular phone habits, and discussed their implications in a depression relapse case study. MethodsIn the MENTALPRECISION study we collected passive smartphone data (phone usage and location behaviours) in a predominantly healthy cohort (n=73) using the Behapp application. We administered a novel questionnaire, the "Smartphone Usage and Lifestyle Questionnaire" (SULQ), to the participants to gather information on their phone usage and lifestyle habits that could impact their digital phenotyping data. We trained a hidden Markov model (HMM) on the smartphone data and developed two types of digital phenotyping measures from the identified hidden behavioural states of the HMM. The "total dwell time" gave the percentage of time each participant spent in each state. The "individual transition probability" was extracted from the HMM itself for each participant, giving their personalised probabilities of transitioning between each of the hidden states. We compared the HMM-generated hidden state sequences and reported events such as holidays and illness. We carried out logistic regression between the digital phenotyping measures and various SULQ measures. We then provide a proof-of-concept for predicting depression relapse in recurrent depression using a HMM and consider the implication of non-clinical factors for this clinical application. ResultsVisible differences in behaviour surrounding holidays and illness were observed in the generated hidden state sequences from the MENTALPRECISION study, as well as surrounding the depression relapse in the proof-of-concept. Participants who use another phone in addition to their personal smartphone spent significantly less time in the "socially inactive home time" state (FDR-corrected P=0.03, odds ratio 0.9196, 95% CI 0.8583-0.9808). iOS users spent significantly less time in the "socially active home time" state (FDR-corrected P=0.04, odds ratio 0.9330, 95% CI 0.8804-0.9857) than Android users, and participants reporting a smartphone addiction spent significantly more time in this state (FDR-corrected P=0.009, odds ratio 1.0787, 95% CI 1.0301-1.1272) when compared to participants reporting no smartphone addiction. Relationships between the mean transition probabilities and SULQ measures did not survive multiple comparison correction. We observed decreases in the monthly likelihood surrounding the depression relapse period, providing a possible metric for relapse prediction. ConclusionsWhen searching for clinically relevant behavioural signals, digital phenotyping researchers should consider additional non-clinical factors that may be contributing to the measured digital signal. Including this additional information could be used to improve personalisation of digital phenotyping models, leading to improved modelling abilities for goals such as depression relapse prediction.