Objective: This study aims to test the feasibility of using digital tools to evaluate motor performance in a Parkinson’s disease clinical trial.

Background: Modern smartphones and wearable devices can implement multiple digital sensors, e.g. accelerometer, gyroscope and photoplethysmogram, that are able to collected behavioral and physiological signals relevant to disease symptoms, including motor impairments of Parkinson’s disease (PD). It provides an unprecedented opportunity to develop digital biomarkers (DBM) using these wearable devices for frequent and objective diagnosing, monitoring disease progression and detecting treatment-associated effect in both clinical trials and real-world situations.

Method: A deep learning algorithm was developed for the Parkinson’s disease digital biomarker DREAM challenge (PDDB challenge) with the aim to classify PD patients and non-PD subjects using digital sensor data collected during gait and balance test in the public mPower app. Lilly Trial App [1] collected similar digital sensor data from 24 PD patients in the D1PAM phase 1B clinical trial, which enabled us to apply the algorithm to Lilly data with minor modifications. Quantitative DBM scores were further derived from the algorithm, in order to test whether the DBM scores correlate with motor performance measured by UPDRS and whether the scores can differentiate treatment group from placebo group.

Results: The algorithm had highest classification accuracy among all the algorithms submitted in the PDDB challenge, and showed a very high true positive rate in Lilly clinical trial data. DBM scores derived from the algorithm were able to detect significant differences between treatment and placebo groups during intervention period, whereas the UPDRS failed to do so. The correlation between UPDRS scores and DBM scores were generally positive, but not statistically significant.

Conclusion: Due to small sample size and lack of significant correlation with UPDRS, further evaluation is warranted to establish clinical relevancy of the treatment-associated DBM score changes. Nevertheless, this work demonstrated the potential of applying deep learning algorithm to digital sensor data as surrogate biomarker and paved a new way of developing objective and near real-time DBM for PD.

References: [1] Jian et al. Treatment monitoring using objective and frequent digital testing in the D1PAM phase 1B Parkinson’s disease clinical trial, MDS Congress (2019)

« Back to 2019 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/use-digital-sensor-and-deep-learning-to-evaluate-motor-performance-in-the-d1pam-ly3154207-phase-1b-parkinsons-disease-clinical-trial/