Advancing post-traumatic seizure classification and biomarker identification: Information decomposition based...

Area of Science:

• Neuroscience
• Medical Imaging
• Machine Learning

Background:

• Traumatic brain injury (TBI) can lead to late post-traumatic seizures (LPTS), potentially causing lifelong post-traumatic epilepsy (PTE).
• Predicting PTE development and identifying biomarkers in TBI patients is crucial but challenging due to elusive epileptogenesis mechanisms.

Purpose of the Study:

• To develop a predictive model for identifying TBI patients at risk of developing LPTS.
• To identify reliable neuroimaging biomarkers for predicting LPTS in TBI patients.

Main Methods:

• Collected longitudinal multimodal data from 48 TBI patients.
• Employed a supervised binary classification approach comparing LPTS and non-LPTS groups.
• Developed novel fusion algorithms (RECC, IDSF) and utilized imputation techniques for handling missing data modalities.
• Applied kernel- or tree-based classifiers and interpretable machine learning (Shapley values).

Main Results:

• The proposed Information Decomposition and Selective Fusion (IDSF) algorithm demonstrated superior performance (AUC).
• Identified specific MRI abnormalities as potential biomarkers: left anterior limb of internal capsule (dMRI) and right middle temporal gyrus (fMRI).

Conclusions:

• The study successfully identified potential neuroimaging biomarkers for predicting LPTS in TBI patients.
• The IDSF algorithm offers an effective method for multimodal data fusion in clinical prediction tasks.