Abstract
BACKGROUND: Working memory (WM) training (WMT) is a popular intervention approach against cognitive decline in patients with Parkinson's disease (PD). However, heterogeneity in WM responsiveness suggests that WMT may not be equally efficient for all patients. OBJECTIVE: The present study aims to evaluate a multivariate model to predict post-intervention verbal WM in patients with PD using a supervised machine learning approach. We test the predictive potential of novel learning parameters derived from the WMT and compare their predictiveness to other more commonly used domains including demographic, clinical, and cognitive data. METHODS: 37 patients with PD (age: 64.09±8.56, 48.6% female, 94.7% Hoehn & Yahr stage 2) participated in a 5-week WMT. Four random forest regression models including 1) cognitive variables only, 2) learning parameters only, 3) both cognitive and learning variables, and 4) the entire set of variables (with additional demographic and clinical data, 'all' model), were built to predict immediate and 3-month-follow-up WM. RESULT: The 'all' model predicted verbal WM with the lowest root mean square error (RMSE) compared to the other models, at both immediate (RMSE = 0.184; 95% -CI=[0.184;0.185]) and 3-month follow-up (RMSE = 0.216; 95% -CI=[0.215;0.217]). Cognitive baseline parameters were among the most important predictors in the 'all' model. The model combining cognitive and learning parameters significantly outperformed the model solely based on cognitive variables. CONCLUSION: Commonly assessed demographic, clinical, and cognitive variables provide robust prediction of response to WMT. Nonetheless, inclusion of training-inherent learning parameters further boosts precision of prediction models which in turn may augment training benefits following cognitive interventions in patients with PD.