Can machine learning with data from wearable devices distinguish disease severity levels and generalise across patients? A pilot study in Mania and Depression

Abstract

Mood disorders are severe and chronic mental conditions exacting high costs from society. The lack of reliable biomarkers to aid clinicians in tailoring pharmacotherapy based on distinguishable patient-specific traits means that the current prescribing paradigm is largely one of trial and error. Previous studies showed that different biological signatures, such as patterns of heart rate variability or electro-dermal reactivity, are associated with clinically meaningful outcomes. Against this backdrop, the advances in machine learning and the spread of wearable devices capable of providing continuous and ecological monitoring of patients may unlock great opportunities in mental healthcare. We herewith present a pilot study on mania and depression where we moved beyond the simple disease state binary classification but pursued the more informative and clinically meaningful task of differentiating between levels of disease severity. While most previous similar endeavours used recording segments extracted from the same subjects for both training and testing, we explicitly carried out model development and evaluation on segments from different groups of patients, in order to have a fair assessment of the model out-of-sample generalisation. This illustrated how individuals heterogeneity and non-disease-related dimensions of variations (e.g. sex, age, physical fitness) may dominate the signal so that in low sample size regimes a model might learn and overfit subject-specific patterns rather than capturing disease-relevant traits generalisable across disorders. Lastly, we developed a viable baseline for pre-processing raw data from wristband recordings and compared three classical and two deep-learning models to identify levels of disease severity.

Automated mood disorder symptoms monitoring from multivariate time-series sensory data: Getting the full picture beyond a single number

“…Channel encoders Since the sampling rate varies across the recorded signals within a segment, these are typically time-aligned, e.g. to the level of a second in wall-time [1,45]. However, the down-sampling process employed to time-align data (usually via max-pooling or averaging) can risk removing useful information in the raw recordings.…”