Decoding dynamic brain networks in Parkinson’s disease with temporal attention

Abstract

Detecting brief, clinically meaningful changes in brain activity is crucial for understanding neurological disorders. Conventional imaging analyses often overlook these subtle events due to computational demands. IMPACT (Integrative Multimodal Pipeline for Advanced Connectivity and Time-series) addresses this challenge by converting 3D/4D fMRI scans into time-series signals using a standardized brain atlas. This approach integrates regional signals, network patterns, and dynamic connectivity, and employs machine learning to detect subtle fluctuations. In Parkinson’s disease diagnosis across two independent cohorts (n=43 and n=40), it achieves high accuracy (area under the curve = 0.97-0.98), outperforming conventional methods. Analyses reveal transient connectivity disruptions that align with dopaminergic mechanisms, while interpretability highlights the critical time windows and regions driving classification. This reproducible, standardized pipeline is readily adaptable to other conditions where short-lived brain changes serve as key diagnostic markers.