Abstract
BACKGROUND: Longitudinal tracking of multiple sclerosis (MS) symptoms in an individual's environment may improve self-monitoring and clinical management for people with MS. Conventional symptom tracking methods rely on self-reports and clinical visits, which can be infrequent, subjective, and burdensome. Digital phenotyping using passively collected sensor data from smartphones and fitness trackers offers a promising alternative for continuous, real-time symptom monitoring with minimal patient burden. OBJECTIVE: We aimed to develop and evaluate a machine learning (ML)-based digital phenotyping approach to monitor the severity of clinically-relevant MS symptoms. We used passive sensing data to predict short-term fluctuations in patient-reported symptoms, including depressive symptoms, global MS symptom burden, severe fatigue, and poor sleep quality. Further, we examined the impact of incorporating behavioral context features and ecological momentary assessments on prediction performance. METHODS: We conducted a 12- to 24-week longitudinal study involving 104 people with MS, collecting passive sensor and behavioral health data. Smartphone sensors recorded call activity, location, and screen use, while fitness trackers captured heart rate, sleep patterns, and step count. We extracted patient-level behavioral features and categorized them into 2 feature sets: one from the prediction period (called action) and one from the preceding period (called context). Using an ML pipeline based on support vector machines and AdaBoost, we evaluated the predictive performance of sensor-based models, both with and without ecological momentary assessment inputs. RESULTS: Between November 16, 2019, and January 24, 2021, overall, 104 people with MS (women: n=88, 84.6%; non-Hispanic White: n=97, 93.3%; mean age 44, SD 11.8 years) from a clinic-based cohort completed 12 weeks of data collection, including a subset of 44 participants (women: n=39, 89%; non-Hispanic White: n=42, 95%; mean age 45.7, SD 11.2 years) who completed 24 weeks of data collection. In total, we collected approximately 12,500 days of passive sensor and behavioral health data from the participants. Among the best-performing models with the least sensor data requirement, the ML algorithm predicted depressive symptoms with an accuracy of 80.6% (F(1)-score=0.76), high global MS symptom burden with an accuracy of 77.3% (F(1)-score=0.78), severe fatigue with an accuracy of 73.8% (F(1)-score=0.74), and poor sleep quality with an accuracy of 72.0% (F(1)-score=0.70). Further, sensor data were largely sufficient for predicting symptom severity, while the prediction of depressive symptoms benefited from minimal active patient input in the form of responses to 2 brief questions on the day before the prediction point. CONCLUSIONS: Our digital phenotyping approach using passive sensors on smartphones and fitness trackers may help patients with real-world, continuous self-monitoring of common symptoms in their own environment and assist clinicians with better triage of patient needs for timely interventions in MS and potentially other chronic neurological disorders.