Artificial intelligence for short-term post-mortem interval estimation: a systematic review with PROBAST and TRIPOD+AI appraisal.
Authors
Affiliations (2)
Affiliations (2)
- Section of Legal Medicine, University of Bari Aldo Moro, Piazza Giulio Cesare 11, 70124, Italy. Electronic address: [email protected].
- Section of Legal Medicine, University of Bari Aldo Moro, Piazza Giulio Cesare 11, 70124, Italy.
Abstract
Estimating the post-mortem interval (PMI) is pivotal in forensic casework, yet traditional approaches are imprecise and context-dependent. Artificial intelligence and machine-learning (AI/ML) models aim to leverage imaging, biochemical and omic data, but their credibility depends on methodological rigor and transparent reporting. We systematically reviewed human studies developing or validating AI/ML models for short-term PMI (≤72 h). Searches to 19 December 2025 covered PubMed/MEDLINE, Web of Science, Scopus, IEEE Xplore and the Cochrane Library. Data were extracted using CHARMS; risk of bias/applicability were assessed with PROBAST; reporting completeness was appraised with TRIPOD+AI. Owing to heterogeneity, results were synthesised narratively. Only 8 studies met the stringent inclusion criteria for short-term PMI estimation in humans: corneal opacity imaging (n = 10; two studies), vitreous biochemistry (n = 174 and n = 201), ATR-FTIR spectroscopy of vitreous humour (n = 70), post-mortem CT radiomics (n = 51 with validation n = 80; and n = 97), and post-mortem microbiomics (n = 188). Datasets were usually small, single-centre and vulnerable to unit-of-analysis errors (multiple observations per decedent). Risk of bias was often high or unclear in the PROBAST analysis domain, driven by limited sample sizes, overfitting-prone analytic workflows and scarce external validation. Reporting gaps were common for sample size justification, missing-data handling, calibration, and model/code availability. AI/ML methods for PMI estimation are promising, but current human evidence is constrained by small, heterogeneous datasets, limited transportability testing, and incomplete reporting. Future research should prioritise adequately powered multi-centre studies with subject-level validation, calibration assessment and open, TRIPOD+AI-aligned reporting.