Learning Robust Spatiotemporal Features from RS-fMRI for Parkinson’s Disease Classification Using CanICA, Dictionary Learning, and LSTM

Abstract

Parkinson’s disease (PD) is a progressive neurodegenerative disorder where early and accurate diagnosis remains a critical challenge. Machine learning applied to functional MRI (fMRI) data offers promising avenues for automated classification by extracting meaningful spatiotemporal features. In this preliminary study, we evaluate two data-driven feature extraction methods—Canonical Inde- pendent Component Analysis (CanICA) and Dictionary Learning (DL)—in combination with Long Short-Term Memory (LSTM) networks. Comparative experiments were conducted using different hidden unit configurations and region definition strategies. Results indicate that CanICA features, when paired with compact LSTM architectures, achieved peak accuracy of 85.17% with a loss of 0.54, outperforming DL in terms of discriminability. Conversely, DL features provided more sta- ble performance across configurations, particularly with connected components, yielding consistent accuracy around 71.43% and lower variance in loss values. These findings highlight the complemen- tary strengths of CanICA and DL: CanICA excels in compact models with high accuracy, while DL ensures robustness and generalization.