Artificial intelligence in postmortem interval estimation: emerging trends, predictive models, and forensic...

Accurate estimation of the postmortem interval (PMI) remains a critical yet challenging task in forensic science due to the inherent variability of biological and environmental factors. Conventional methods based on observable postmortem changes are often subjective and limited in temporal applicability. Recent advancements in artificial intelligence (AI), particularly machine learning (ML) and deep learning (DL), have provided promising approaches with the potential to improve the accuracy, objectivity, and reproducibility of PMI estimation. This systematic review aims to critically evaluate recent developments in AI-based approaches for PMI estimation, with emphasis on analytical techniques, biological matrices, and predictive performance.

A systematic literature search was conducted following PRISMA 2020 guidelines across PubMed/MEDLINE, Scopus, Web of Science Core Collection, IEEE Xplore, and ScienceDirect databases for studies published between January 2021 and April 2026. Eligible studies included original research evaluating artificial intelligence, machine-learning, chemometric, statistical, and other data-driven computational approaches for PMI estimation using human or animal postmortem samples. Data extraction focused on study design, biological matrices, biomarkers, AI models, PMI range, and performance metrics such as accuracy, mean absolute error (MAE), and coefficient of determination (R²). Quality assessment was performed using the Prediction Model Risk of Bias Assessment Tool (PROBAST). Owing to substantial methodological heterogeneity across AI models, datasets, outcome measures, and validation strategies, a structured narrative synthesis with descriptive summaries of reported performance metrics was conducted instead of a quantitative meta-analysis.

A total of 29 studies were included, comprising animal (n = 14), human (n = 13), and combined human-animal datasets (n = 2). The evidence encompassed controlled animal experiments and human studies, with substantial variation in PMI windows, prediction tasks, computational approaches, and validation strategies. Machine learning and deep learning models, together with chemometric and conventional statistical approaches, demonstrated promising predictive performance across metabolomic, proteomic, microbiome, spectroscopic, and imaging datasets, with prediction errors as low as approximately 1-5 h reported in some controlled experimental settings. However, predictive performance was generally more promising in controlled experimental settings, whereas human evidence was more limited and heterogeneous. External validation was also limited, restricting assessment of model generalizability and routine forensic applicability.

Computational and predictive approaches demonstrate considerable potential to improve the objectivity and predictive performance of PMI estimation; however, substantial methodological heterogeneity, limited external validation, small human datasets, and reliance on experimental models currently constrain their translation into routine forensic practice.