Published September 23, 2026, in Computer Methods in Biomechanics and Biomedical Engineering, this study used fNIRS signals from 175 participants to distinguish cortical from subcortical ischemic stroke infarctions.
Three recording conditions The model combined residual, bidirectional recurrent and Transformer components. Accuracy was 0.886 during verbal fluency, 0.913 during right-hand grasp and 0.878 at rest. Corresponding AUCs were 0.920, 0.912 and 0.900. A SHAP analysis identified channels that contributed most to the predictions.
Why the approach is promising Results across task and resting conditions suggest that portable optical recordings may carry useful subtype information. Patient-wise cross-validation is stronger than mixing one person's samples across folds.
The important boundary This was internal five-fold validation, not testing in a separate hospital. The abstract did not report subgroup counts, demographics, time since stroke, clinical workflow comparisons, funding or conflicts. The model is not a validated diagnostic tool.
Source and disclosure Authoritative abstract checked. Prepared with AI assistance and source-checked September 26, 2026.