Spontaneous eye blink-based machine learning for tracking clinical fluctuations in Parkinson's disease.

Abstract

In this uncontrolled, open-label exploratory clinical study, the authors explore the potential of blink data as a digital biomarker for estimating clinical indices of Parkinson's disease (PD) using a machine learning approach. Blink data were collected from 20 patients with PD before and after (up to 4 h) L-dopa/decarboxylase inhibitor administration. Concurrent assessments of patient diary-based ON/OFF and dyskinesia, L-dopa plasma concentration, and MDS-UPDRS Part III scores were conducted at 30 min intervals. The models were developed to predict clinical symptoms based on blink data collected at 3 min intervals. The most effective post-processing models accurately predicted the ON/OFF states (mean area under the receiver operating characteristic curve (AUC_ROC) = 0.87) and the presence of dyskinesia (mean AUC_ROC = 0.84). They also moderately predicted MDS-UPDRS Part III scores (mean Spearman's correlation ρ = 0.54) and plasma L-dopa concentrations (ρ = 0.57). Our findings highlight the potential of the spontaneous eye blink as a noninvasive, real-time digital biomarker for PD.