JMIR Publications

ABSTRACT

Background:

Machine learning models for ischemic stroke prediction have achieved strong performance, yet clinical adoption remains limited by insufficient interpretability and a lack of integration into collaborative decision-making workflows. Explainable AI (XAI) methods such as SHAP and LIME improve transparency, and counterfactual explanations support "what-if" clinical reasoning, but these techniques are typically studied in isolation from one another and from the human-machine teaming (HMT) context in which clinical decisions are made. No prior review has systematically examined how these three elements interact within stroke decision support.

Objective:

This review aimed to (1) identify the XAI and counterfactual explanation methods used in ischemic stroke prediction and evaluate how they are implemented and assessed, (2) examine how the literature evaluates the interpretability and clinical value of these methods, (3) determine how integrating XAI with HMT affects clinicians' trust and decision-making, and (4) characterize the limitations preventing full integration of these elements.

Methods:

A systematic review was conducted following PRISMA 2020 guidelines. PubMed, IEEE Xplore, and Scopus were searched for studies published between January 2019 and April 2026, with Google Scholar used for supplementary identification. The main search was completed on March 15, 2026, with a targeted update on April 30, 2026. Of 4,027 records identified, 2,871 were screened after deduplication, 288 full texts were assessed, and 74 studies were included spanning prediction model development, XAI evaluation, interface usability testing, and HMT experiments. Methodological quality was appraised within each thematic section using dimensions appropriate to each study type.

Results:

Feature attribution methods (SHAP, LIME, Grad-CAM) dominate stroke XAI; counterfactual approaches remain rare and largely absent from deployed systems. Evaluation relies primarily on technical metrics rather than clinical outcomes. Explanation stability is scale-dependent, converging at the population level but unstable for individual patients. No deployed counterfactual interface was identified within an acute stroke workflow. Human-AI teaming on average outperforms clinicians alone but rarely achieves full complementarity, and explanation does not automatically improve performance. A recurring structural bottleneck, termed the causal-feasibility gap, was identified: counterfactual explanations can be mathematically valid yet clinically unusable when they do not account for patient-specific constraints such as treatment windows and modifiable variable ranges. An integrated conceptual framework is proposed in which XAI, counterfactual reasoning, and HMT operate across four stages of the stroke clinical pathway through bidirectional clinician-AI exchange.

Conclusions:

Transparency alone is insufficient for clinical value in stroke AI. Closing the causal-feasibility gap requires systems that receive feasibility constraints from clinicians rather than only delivering model outputs. The proposed bidirectional framework provides a basis for designing decision-support systems suited to time-critical stroke care. Priority directions include stroke-specific counterfactual benchmarks, case-level explanation stability methods, and prospective clinical outcome evaluation.