Comparing machine learning and deep learning models to predict cognition progression in Parkinson's disease

Abstract

Cognitive decline in Parkinson's disease (PD) varies widely. While models to predict cognitive progression exist, comparing traditional probabilistic models to deep learning methods remains understudied. This study compares sequential modeling techniques to identify cognitive progression in individuals with and without PD. Using data from the Parkinson's Progression Marker Initiative, shallow Markov, deep recurrent (long short-term memory [LSTM]), and nonrecurrent (temporal fusion transformer [TFT]) models were compared to predict cognitive status over time. Cognitive status was categorized into normal cognition (NC), mild cognitive impairment (MCI), and dementia. Predictions were made annually for up to 3 years using clinical data, including demographics, cognitive assessments, PD severity, and medical history. Each approach was evaluated using inverse probability weighted (IPW-) F1 scores. An ensemble method combined outputs from the Markov, LSTM, and TFT models. The dataset included 917 individuals (53% PD; 30% at risk for PD; 17% Healthy Controls). The TFT model outperformed others across all annual periods (IPW-F1 = 0.468) compared to the Markov (IPW-F1 = 0.349) and LSTM (IPW-F1 = 0.414) models, with improved performance using an ensemble approach (IPW-F1 = 0.502). For MCI and dementia predictions, which were rarer occurrences compared to NC status (ratios: 50:8:1), the TFT model consistently outperformed competing models, achieving IPW-F1 scores of 0.496 and 0.533 for MCI and dementia, respectively. In conclusion, sequential deep learning models like TFT, which mitigate long-term memory loss and can interpret complex, high-dimensional data, perform best overall in predicting clinically important cognitive transitions. These methods should be further explored for predicting degenerative conditions.