Deep learning-based classification of cortical and subcortical ischemic stroke infarction using fNIRS signals.

A model separated cortical from subcortical stroke groups with encouraging internally cross-validated performance across three fNIRS conditions. It is an early engineering result, not a validated diagnostic test.

Xu B, Xu H, Liu H, Chen Y.
Computer methods in biomechanics and biomedical engineering · Publication date: 2026-09-23
Added to site: 2026-09-26
DOI: 10.1080/10255842.2026.2736283

Electronic publication date verified in Europe PMC; first indexed 2026-09-24. This source date is separate from the site's addition date.

Modalities

fNIRS, Machine learning

Clinical focus

Ischemic stroke, Stroke subtype classification, Explainable AI

Population

People with cortical or subcortical ischemic stroke infarctions

Sample

175 participants; cortical and subcortical subgroup counts were not reported in the checked abstract

Intervention / method

A deep model combining a residual module, bidirectional LSTM and Transformer classified fNIRS recordings collected during verbal fluency, right-hand grasp and resting-state conditions

Comparator

Cortical versus subcortical infarction labels within patient-wise internal cross-validation; no external validation cohort was reported

Clinician summary

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, with SHAP highlighting channels that influenced predictions.

How to interpret the findings

Internal cross-validation cannot establish performance in a new hospital or clinical workflow. The checked abstract did not report subgroup denominators, demographics, time since stroke, reference-standard procedures, confidence intervals for individual folds, external calibration, adverse events, funding or conflicts.

Equipment and disclosures

fNIRS acquisition system and optode layout were not available in the checked abstract.

Funding and competing-interest details were not available in the checked indexed record.

Related evidence

Compare findings from the same clinical area, including mixed results.