Related Experiment Video
Updated: Jan 6, 2026
05:47
Evidence-based Knowledge Synthesis and Hypothesis Validation: Navigating Biomedical Knowledge Bases via Explainable AI and Agentic Systems
Published on: June 13, 2025
1.2K
eXplainable Artificial Intelligence (XAI): A Systematic Review for Unveiling the Black Box Models and Their Relevance
Nadeesha Hettikankanamage1, Niusha Shafiabady2,3, Fiona Chatteur4
1Design and Creative Technology, Torrens University Australia, 88 Wakefield St., Adelaide, SA 5000, Australia.
Sensors (Basel, Switzerland)
|November 13, 2025
Summary
eXplainable Artificial Intelligence (XAI) is crucial for AI in biomedical fields. This review identifies 13 XAI techniques for prediction, with SHAP being most popular, but highlights gaps in usability validation for clinical use.
Area of Science:
- Biomedical Imaging and Sensing
- Artificial Intelligence
- eXplainable Artificial Intelligence (XAI)
Background:
- AI adoption in safety-critical biomedical fields is hindered by complex model opacity.
- eXplainable Artificial Intelligence (XAI) is essential for transparency, fairness, accountability, and bias mitigation in AI decision-making.
- Quantitative prediction tasks are a key focus for AI applications in various domains.
Purpose of the Study:
- To systematically review eXplainable Artificial Intelligence (XAI) techniques applied to quantitative prediction tasks across different fields.
- To assess the methodological relevance and potential adaptation of these XAI techniques for biomedical imaging and sensing.
- To identify gaps and provide guidance for future research in interpretable AI for biomedical applications.
Main Methods:
- A systematic literature review following PRISMA guidelines was conducted.
- Analysis of 44 Q1 journal articles utilizing XAI techniques for prediction tasks with quantitative databases.
- Identification and categorization of XAI techniques based on their application and contribution to explaining predictions.
Main Results:
- 13 distinct eXplainable Artificial Intelligence (XAI) techniques were identified for quantitative prediction tasks.
- Shapley Additive eXPlanations (SHAP) was the most frequently used technique (35/44 articles), followed by LIME, PDPs, and PFI.
- Theoretical limitations of model-agnostic methods like SHAP (e.g., additive/causal assumptions) were noted, especially for heterogeneous biomedical data.
Conclusions:
- While computational evaluations of XAI are common, a significant research gap exists in structured human-subject usability validation for clinical translation.
- Methodological and usability gaps need addressing for effective adaptation of XAI in biomedical imaging and sensing.
- Future research should focus on developing and validating XAI techniques that ensure safe, interpretable, and clinically relevant AI deployment.
