Objective: Develop and evaluate a multimodal, transformer-based pipeline for detecting Parkinson’s disease (PD) from resting-state fMRI time series, aiming to expose short-lived neural connectivity disruptions that may inform early diagnosis and therapeutic strategies.

Background: Parkinson’s disease gradually alters basal ganglia–cortical circuitry, often eluding clinical detection until overt motor deficits manifest. Conventional approaches that aggregate fMRI data risk smoothing out brief but pathologically significant connectivity perturbations, which may coincide with dopaminergic oscillations. A time-resolved method is needed to capture these hidden neural events.

Method: Two independent resting-state fMRI datasets (n=83) were preprocessed to extract region-of-interest (ROI) time series via the Harvard–Oxford atlas. Independent component analysis (ICA) yielded large-scale network features, while a sliding-window correlation pipeline captured dynamic connectivity patterns. All three feature streams—ROI signals, ICA components, and window-based connectivity—were processed within a transformer architecture (IMPACT: Interpretable Multimodal Pipeline for Advanced Connectivity and Time-series). Temporal self-attention identified key time points for classification, and a cross-modal fusion block gated contributions from each modality. Comparisons involved baseline models, including 1D/2D convolutional networks, LSTM, GRU, TCN, MLP, and an autoencoder.

Results: IMPACT surpassed baselines in both datasets, achieving area-under-the-curve (AUC) values of 0.977 (95% CI: 0.931–1.000) and 0.973 (95% CI: 0.924–0.994). Brief connectivity breakdowns were highlighted by GradCAM and attention-weight analyses, suggesting alignment with clinical on–off medication states. [figure1], [figure2]

Conclusion: A transformer-based, multimodal fMRI pipeline that models transient connectivity states demonstrated high PD detection accuracy. By preserving temporal detail and integrating multiple data streams, this approach uncovers episodic network instabilities potentially tied to dopaminergic fluctuations, providing a rationale for earlier detection, individualized monitoring, and closed-loop neuromodulation strategies in Parkinson’s disease.

Model performance comparison across datasets.

Box plots show AUC scores

Enhanced visualization of ROI importance patterns.

Time series analysis with GradCAM

Radar plots showing the importance of brain lobes

« Back to 2025 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/multimodal-transformer-based-detection-of-parkinsons-disease-in-fmri-extracted-time-series/