Scholarlink

Background <p>Artificial intelligence models for acute kidney injury (AKI) prediction achieve strong discriminative accuracy, yet clinical adoption remains constrained by model opacity and alarm fatigue. Explainable artificial intelligence (XAI) methods may enhance clinician trust and alert acceptance; however, the extent of their clinical validation and implementation remains unclear.</p> Methods <p>We performed a scoping review following PRISMA–ScR guidelines. PubMed, Embase, Web of Science, CINAHL, IEEE Xplore, and ACM Digital Library were searched from January 2012 through February 2026. Studies applying XAI methods to AKI prediction or management in adult inpatients were eligible. Prediction model studies were appraised using PROBAST. We additionally propose the GLASS–AKI conceptual framework as a structured research agenda derived from identified gaps.</p> Results <p>Thirty-four studies met inclusion criteria (total patient population &gt; 2.1 million). SHAP-based attribution was the most frequently reported XAI technique (26/34; 76.5%), followed by LIME (9; 26.5%), attention-weight visualization (7; 20.6%), and rule-based surrogates (5; 14.7%). Reported AUROC values ranged from 0.71 to 0.95 (median 0.84). Critically, XAI was applied as a post hoc explanation method in 88.2% of studies and did not inherently alter model discriminative performance. Prospective clinical deployment data were limited to 4 studies (11.8%). PROBAST assessment identified high overall risk of bias in 79.4% of studies, predominantly in the Analysis domain. No study integrated multi-omic biomarker signals with real-time XAI reasoning at the point of care.</p> Conclusions <p>XAI methods are increasingly applied to AKI prediction models, but prospective evidence linking explainability to improved clinician behavior or patient outcomes remains limited. We propose the GLASS–AKI framework as a conceptual research agenda—not a validated system—to guide future multicenter prospective evaluation of integrated XAI-biomarker approaches.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

From black box to glass box: explainable artificial intelligence for acute kidney injury prediction—a scoping review and the GLASS–AKI translational framework proposal

  • 期刊论文

  • 发表日期：

    2026年6月20日

查看全文
  • Wan-Ling Lin,
  • Shih-Shuan Fang,
  • Sheng-Han Chen

摘要

Background

Artificial intelligence models for acute kidney injury (AKI) prediction achieve strong discriminative accuracy, yet clinical adoption remains constrained by model opacity and alarm fatigue. Explainable artificial intelligence (XAI) methods may enhance clinician trust and alert acceptance; however, the extent of their clinical validation and implementation remains unclear.

Methods

We performed a scoping review following PRISMA–ScR guidelines. PubMed, Embase, Web of Science, CINAHL, IEEE Xplore, and ACM Digital Library were searched from January 2012 through February 2026. Studies applying XAI methods to AKI prediction or management in adult inpatients were eligible. Prediction model studies were appraised using PROBAST. We additionally propose the GLASS–AKI conceptual framework as a structured research agenda derived from identified gaps.

Results

Thirty-four studies met inclusion criteria (total patient population > 2.1 million). SHAP-based attribution was the most frequently reported XAI technique (26/34; 76.5%), followed by LIME (9; 26.5%), attention-weight visualization (7; 20.6%), and rule-based surrogates (5; 14.7%). Reported AUROC values ranged from 0.71 to 0.95 (median 0.84). Critically, XAI was applied as a post hoc explanation method in 88.2% of studies and did not inherently alter model discriminative performance. Prospective clinical deployment data were limited to 4 studies (11.8%). PROBAST assessment identified high overall risk of bias in 79.4% of studies, predominantly in the Analysis domain. No study integrated multi-omic biomarker signals with real-time XAI reasoning at the point of care.

Conclusions

XAI methods are increasingly applied to AKI prediction models, but prospective evidence linking explainability to improved clinician behavior or patient outcomes remains limited. We propose the GLASS–AKI framework as a conceptual research agenda—not a validated system—to guide future multicenter prospective evaluation of integrated XAI-biomarker approaches.