Objective: To construct deep learning (DL) models using finger tapping data and apply them in PRESENCE to evaluate treatment effect in motor function.

Background: PRESENCE was a phase 2 clinical trial assessing mevidalen (LY3154207), a D1 receptor positive allosteric modulator, to treat symptoms of Lewy body dementia. Participants received daily doses (10, 30 or 75mg) of mevidalen or placebo. Sensors such as accelerometer on smartphones are increasingly used in clinical studies to assess Parkinson’s disease (PD) symptoms [1-3]. To derive clinically meaningful information from sensor data, DL has performed better [4, 5] than feature engineering and signal processing.

Method: We developed DL regression models using tapping data and patient-reported MDS-UPDRS part II (MDS-part2) scores from Bot el al. [4]. Deep 1D convolutional neural network models from accelerometer (accelerometer model, AM) and coordinate (coordinate model, CM) data were constructed using data from 1932 PD patients. We applied the models to pre, during, and posttreatment data collected via the PRESENCE tapping app. The relationship of MDS-part2 scores at pretreatment from model prediction and clinical assessment was analyzed. Model prediction scores were averaged for each patient and each hand. Treatment effect was determined as change from pretreatment. Pearson’s correlation coefficients and unadjusted p-values are reported.

Results: DL model-predicted scores correlated with patient-reported MDS-part2 scores in mPower (AM: r=.53, p<.001; CM: r=.55, p<.001). When DL models were applied to PRESENCE pretreatment data, model-predicted scores correlated with physician-assessed MDS-part2 scores (affected hands: AM: r=.34, p<.001; CM: r=.33, p<.001; preferred hands: AM: r=.3, p<.001; CM: r=.28, p<.001). During treatment, model-predicted scores from both AM and CM tapping DL models trended in the direction of improvement at 75 mg for the more affected hand from weeks 8 to 12. Statistical significance was not observed. No differences were observed between the other treatment arms (10 mg and 30 mg) and placebo arm.

Conclusion: Results showed DL models trained from tapping data predict scores correlating with clinical motor function scores. More evaluation is needed to determine the treatment effect assessed by DL model-predicted scores.

References: [1] Willey, J. Open-sourcing health. How working together is changing big data into life-changing healthcare. Roche website https://www.roche.com/sustainability/open-sourcing-health.htm . Published 12 September 2018. Accessed 3 March 2021. [2] Mov Disord. 2019; 34 (suppl 2). https://www.mdsabstracts.org/abstract/treatment-monitoring-using-objective-and-frequent-digital-testing-in-the-d1pam-ly3154207-phase-1b-parkinsons-disease-clinical-trial/. Accessed March 2, 2021. [3] Mov Disord. 2019; 34 (suppl 2). 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/. Accessed March 2, 2021. [4] Bot, B. M. et al. The mPower Study, Parkinson disease mobile data collected using ResearchKit. Sci. Data 3:160011 (2016) [5] BEAT-PD (Biomarker and Endpoint Assessment to Track Parkinson’s Disease) DREAM challenge (2020). The Michael J. Fox Foundation for Parkinson’s Research website https://www.michaeljfox.org/grant/beat-pd-biomarker-and-endpoint-assessment-track-parkinsons-disease-dream-challenge Published 13 January 2020. Accessed 3 March 2021.

« Back to MDS Virtual Congress 2021

MDS Abstracts - https://www.mdsabstracts.org/abstract/evaluation-of-motor-function-using-motor-testing-app-and-deep-learning-for-the-treatment-of-lewy-body-dementia/