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

Abstract

Mood disorders (MDs) are among the leading causes of disease burden worldwide. Limited specialized care availability remains a major bottleneck thus hindering pre-emptive interventions. MDs manifest with changes in mood, sleep, and motor activity, observable in ecological physiological recordings thanks to recent advances in wearable technology. Therefore, near-continuous and passive collection of physiological data from wearables in daily life, analyzable with machine learning (ML), could mitigate this problem, bringing MDs monitoring outside the clinician's office. Previous works predict a single label, either the disease state or a psychometric scale total score. However, clinical practice suggests that the same label may underlie different symptom profiles, requiring specific treatments. Here we bridge this gap by proposing a new task: inferring all items in HDRS and YMRS, the two most widely used standardized scales for assessing MDs symptoms, using physiological data from wearables. To that end, we develop a deep learning pipeline to score the symptoms of a large cohort of MD patients and show that agreement between predictions and assessments by an expert clinician is clinically significant (quadratic Cohen's κ and macro-average F1 score both of 0.609). While doing so, we investigate several solutions to the ML challenges associated with this task, including multi-task learning, class imbalance, ordinal target variables, and subject-invariant representations. Lastly, we illustrate the importance of testing on out-of-distribution samples.

Fig. 1. The same severity level can be realized from different symptom combinations, underlying different treatment needs.

Top row: a pair of patients with Major Depressive Disorder on a Major Depressive episode; while both share the same severity levels, total Hamilton Depression Rating Scale (HDRS) ≥ 23 [33]. Patient (a), with total HDRS = 24, exhibits high levels of anxiety (H9, H10, H11), whereas patient (b), with total HDRS = 26, displays a marked insomnia component (H4, H5, H6). Bottom row: a pair of patients with Bipolar Disorder on a Manic Episode with a total Young Mania Rating Scale (YMRS) ≥ 25. Patient (c), with total YMRS = 30, has an irritable/aggressive profile (Y2, Y5, Y9) whereas patient (d), with total YMRS = 30, has a prominently elated/expansive presentation (Y1, Y3, Y7, Y11). Knowing what specific symptoms underlie a given state may allow clinicians to tailor treatment accordingly: e.g., a molecule with a stronger anxiolytic profile such as paroxetine or a short course of a benzodiazepine as an antidepressant is introduced may be appropriate in patient (a) whereas patient (b) might benefit from a compound with marked hypnotic properties such as mirtazapine.

Fig. 2. Analysis workflow.

Patients had up to four assessments. At the start of each assessment, a clinician scored the patient on the Hamilton Depression Rating Scale (H in the figure) and Young Mania Rating Scale (Y) and provided an Empatica E4 device asking the patient to wear it for ~48 h (i.e., average E4 battery life). An Artificial Neural Network (ANN) model is fed with recording segments and is tasked with recovering clinician scores. The quadratic Cohen’s κ measures the degree to which the machine scores are in agreement with those of the clinician. The ANN model is made of Classifier (CF) and Critic (CR). The former comprises three main modules: (1) Encoder (EN), projecting input sensory channels onto a new space where all channels share the same dimensionality, regardless of the native E4 sampling frequency; (2) Representation Module (RM), extracting a representation h that is shared across all items; and (3) one Item Predictor IP_j for each item. CR is tasked with telling subjects (S in the figure) apart using h and is pitted in an adversarial game against RM(EN(⋅)), designed to encourage the latter to extract subject-invariant representations.

Fig. 3. All physiological modalities contributed to the test performance across items, however, this was particularly pronounced for Acceleration (ACC) and relatively modest for Blood Volume Pressure (BVP).

Effect of dropping individual channels on item performance. The dotted line is at the level of baseline model performance while each bar indicates the performance upon re-training the best model including all channels but the one corresponding to the bar color code, as shown in the legend.