Mobile phone sensors can discern medication-related gait quality changes in Parkinson's patients in the home environment

•Accelerometer data allows abnormal gait to be identified in Parkinson's patients.•Up to 92.5% gait classification accuracy was achieved using raw smartphone data.•Individual random forest models outperformed neural networks and logistic regression. Patients with Parkinson's Disease (PD) experience daytime symptom fluctuations, which result in small amplitude, slow and unstable walking during times when medication attenuates. The ability to identify dysfunctional gait patterns throughout the day from raw mobile phone acceleration and gyroscope signals would allow the development of applications to provide real-time interventions to facilitate walking performance by, for example, providing external rhythmic cues. Patients (n = 20, mean Hoehn and Yahr: 2.25) had their ambulatory data recorded and were directly observed twice during one day: once after medication abstention, (OFF) and once approximately 30 min after intake of their medication (ON). Regularized generalized linear models (RGLM), neural networks (NN), and random forest (RF) classification models were individually trained for each participant. Across all subjects, our best performing classifier on average achieved an accuracy of 92.5%. This study demonstrated that smartphone accelerometers and gyroscopes can be used to distinguish between ON versus OFF times, potentially making smartphones useful intervention tools.