Validating an evidence-based reliability engine for systematic MASLD mortality review: moving beyond clinical assertion

Background: Medical review articles often summarize claims without showing the strength of supporting evidence. This is important in metabolic dysfunction-associated steatotic liver disease (MASLD), where mortality risk reflects hepatic, cardiovascular, metabolic, renal, and emerging predictors. A claim-to-evidence method may improve transparency and reliability.

Aim: To develop and evaluate a claim-to-evidence reliability engine that links MASLD mortality claims to studies, separates verified from candidate evidence, assigns claim-level confidence, and recommends evidence-calibrated wording.

Methods: A curated MASLD mortality corpus was used as a case study. From 5,770 PubMed/PMC records, 250 were screened in detail, and 131 studies were retained after staged screening and evidence prioritization. Ten claim themes were assessed, including fibrosis mortality, FIB-4, non-invasive fibrosis markers, cardiovascular mortality, type 2 diabetes, CKD, lean MASLD, cardiometabolic burden, temporal trends, and emerging predictors. Claim-study links were generated using theme mapping, predictor-outcome labels, and keyword matching. Links were verified when they showed direct theme alignment or strict predictor and mortality-outcome label matching. Candidate links were broader clinically relevant matches requiring manual review. Confidence was assigned using verified support count, study design, evidence priority, effect-size availability, and outcome relevance. This was an internal audit assessment, not an external expert gold-standard validation.

Results: The engine generated 391 claim-study links: 274 verified and 117 candidate, giving a verified-link yield of 70.1%. High confidence was assigned to fibrosis mortality, cardiovascular mortality burden, and cardiometabolic burden. Moderate-high confidence was assigned to FIB-4, non-invasive fibrosis markers, type 2 diabetes, and lean MASLD. Emerging predictors were moderate. Type 2 diabetes, CKD, temporal trends, and emerging predictors were flagged for manual review. The engine recommended cautious wording for less certain claims, such as CKD-related mortality risk.

Conclusion: This engine offers a reproducible approach for transparent MASLD mortality reviews by separating verified from candidate evidence, assigning confidence levels, and guiding clinically cautious wording.