Contrastive explainable AI in healthcare: a systematic review of trends, benefits, research gaps, and future directions

Abstract

Abstract Background Contrastive eXplainable Artificial Intelligence (cXAI) has emerged as a promising approach to align machine learning explanations with human reasoning. Unlike classical XAI, which addresses questions such as “Why P?”, contrastive explanations explicitly consider alternatives, answering “Why P, rather than Q?”. By making the implicit comparison explicit and focusing only on features that distinguish between outcomes, cXAI has been argued to provide more cognitively aligned, less ambiguous, and more actionable explanations than classical XAI. Moreover, in the context of healthcare, this comparative structure closely mirrors clinicians’ abductive reasoning in differential diagnosis and treatment selection. Our purpose is to systematically examine how contrastive explanations have been developed and applied in healthcare; emphasize the related benefits; identify current trends, research gaps, and future directions. Methods We conducted a systematic literature review of studies published from 2018 onward, following the PRISMA 2020 guidelines. Five scientific databases (IEEE Xplore, Scopus, PubMed, ACM Digital Library, and SpringerLink) were queried using structured search strings, complemented by an AI-assisted search using the Elicit tool. Eligible studies focused on the development or application of contrastive explanation techniques in healthcare-related machine learning tasks. Extracted information included data modalities, AI models, contrastive methods, visualization techniques, and validation strategies. Results Only 18 studies met the eligibility criteria, reflecting the relative novelty of the field, while also indicating that the translation of cXAI into healthcare applications is still at an early stage. Most works focused on tabular data and medical imaging, while physiological signals and video/volumetric data were notably underrepresented. Contrastive Explanation Method (CEM) was the most frequently employed approach. Visualization methods for explanations predominantly relied on heatmaps or feature-importance displays, with textual explanations mainly used in studies involving patients or lay users. Importantly, only a small number of studies reported empirical evaluations involving end users, indicating a lack of rigorous validation. Conclusions cXAI offers a clinically aligned and cognitively coherent approach to explainability in healthcare by explicitly framing decisions in relation to meaningful alternatives. Despite its strong conceptual suitability, its adoption in real-world medical settings remains limited, with underexplored modalities and scarce user-centered validation. Advancing cXAI in healthcare will require (i) rigorous clinical assessment of explanations and systematic collection of user feedback and (ii) the development of multimodal, human-centered explanation frameworks. Through these efforts, cXAI can evolve from a promising theoretical paradigm into a practical tool for clinically aligned decision support.