Parkinson's progression prediction using machine learning and serum cytokines

Abstract

The heterogeneous nature of Parkinson's disease (PD) symptoms and variability in their progression complicates patient treatment and interpretation of clinical trials. Consequently, there is much interest in developing models that can predict PD progression. In this study we have used serum samples from a clinically well characterized longitudinally followed Michael J Fox Foundation cohort of PD patients with and without the common leucine-rich repeat kinase 2 (LRRK2) G2019S mutation. We have measured 27 inflammatory cytokines and chemokines in serum at baseline and after 1 year to investigate cytokine stability. We then used the baseline measurements in conjunction with machine learning models to predict longitudinal clinical outcomes after 2 years follow up. Using the normalized root mean square error (NRMSE) as a measure of performance, the best prediction models were for the motor symptom severity scales, with NRMSE of 0.1123 for the Hoehn and Yahr scale and 0.1193 for the unified Parkinson's disease rating scale part three (UPDRS III). For each model, the top variables contributing to prediction were identified, with the chemokines macrophage inflammatory protein one alpha (MIP1α), and monocyte chemoattractant protein one (MCP1) making the biggest peripheral contribution to prediction of Hoehn and Yahr and UPDRS III, respectively. These results provide information on the longitudinal assessment of peripheral inflammatory cytokines in PD and give evidence that peripheral cytokines may have utility for aiding prediction of PD progression using machine learning models.

Keywords: Parkinson's disease; Predictive markers.

Fig. 1

Increased levels of platelet-derived growth factor in LRRK2-PD serum. a Overview of patient sample numbers and data collection. Baseline platelet-derived growth factor b and monocyte chemoattractant protein 1 c levels in Parkinson’s disease patients with (red squares) and without (blue circles) the LRRK2 G2019S mutation. Data were compared using Student’s t test. Graphs show mean ± SEM, n = 80 per group. d Log2 changes over 1 year for all 27 cytokine measurements in all patients. N = 160

Fig. 2

Random forest prediction of longitudinal clinical outcomes. Random forest machine learning algorithms using baseline cytokine measures were used to predict the 2-year outcomes for the geriatric depression scale a, Hoehn and Yahr b, Schwab and England activities of daily living c, and UPDRS III d. The root mean square error (RMSE) and normalized RMSE (NRMSE) are shown as an indication of performance for each predictive algorithm. For each algorithm the variables that contributed the most to the predictive performance are listed. Data points indicate the machine learning prediction for each individual in the test dataset. The lower the NRMSE value and the more the test data points lie along the prediction line, the better the prediction accuracy