Deep Learning-Based Human Activity Recognition for Continuous Activity and Gesture Monitoring for Schizophrenia Patients With Negative Symptoms

Abstract

Background: We aimed to develop a Human Activity Recognition (HAR) model using a wrist-worn device to assess patient activity in relation to negative symptoms of schizophrenia.

Methods: Data were analyzed in a randomized, three-way cross-over, proof-of-mechanism study (ClinicalTrials.gov: NCT02824055) comparing two doses of RG7203 with placebo, given as adjunct to stable antipsychotic treatment in patients with chronic schizophrenia and moderate levels of negative symptoms. Baseline negative symptoms were assessed using the Positive and Negative Syndrome Scale (PANSS) and Brief Negative Symptom Scale (BNSS). Patients were given a GeneActiv^™ wrist-worn actigraphy device to wear over a 15-week period. For this analysis, actigraphy data and behavioral and clinical assessments obtained during placebo treatment were used. Motivated behavior was evaluated with a computerized effort-choice task. A trained HAR model was used to classify activity and an activity-time ratio was derived. Gesture events and features were inferred from the HAR-detected activities and the acceleration signal.

Results: Thirty-three patients were enrolled: mean (±SD) age 36.6 ± 7 years; mean (±SD) baseline PANSS negative symptom factor score 23.0 ± 3.5; and mean (±SD) baseline BNSS total score 36.0 ± 11.5. Activity data were collected for 31 patients with a median monitoring time of 1,859 h per patient, equating to ~11 weeks or 74% monitoring ratio. The trained HAR model demonstrated >95% accuracy in separating ambulatory and stationary activities. A positive correlation was seen between the activity-time ratio and the percent of high-effort choices (Spearman r = 0.58; P = 0.002) in the effort-choice task. Median daily gesture counts correlated negatively with the BNSS total score (Spearman r = -0.44; P = 0.03), specifically with the diminished expression sub-score (Spearman r = -0.42; P = 0.03). Gesture features also correlated negatively with the BNSS total score and diminished expression sub-scores. Activity measures showed similar correlations with PANSS negative symptom factor but did not reach significance.

Conclusion: Our findings support the use of wrist-worn devices to derive activity and gesture-based digital outcome measures for patients with schizophrenia with negative symptoms in a clinical trial setting.

Keywords: body-worn sensor; digital endpoints; digital health technology; digital outcome measures; gesture detection; human activity recognition; negative symptoms; schizophrenia.