Transcriptome-based postmortem interval prediction in the human dorsolateral prefrontal cortex using covariate-aware machine learning

Accurate estimation of the postmortem interval (PMI) remains a persistent challenge in forensic medicine, as it provides essential temporal evidence for death investigations and judicial proceedings. While messenger RNA (mRNA) degradation kinetics represents a promising molecular strategy for PMI prediction, most prior studies have been confined to small candidate gene sets and a limited range of human tissues. Here, we developed a covariate-aware transcriptomic framework for PMI estimation using RNA-seq data from 270 human dorsolateral prefrontal cortex (DLPFC) samples. Following expression filtering and trimmed mean of M-values (TMM) normalization, we prioritized candidate genes via covariate-adjusted gene-PMI association analysis and bootstrap elastic-net stability ranking within a repeated cross-validation framework. Panel-size and feature-set optimization identified a parsimonious 400-gene “Gene-only” configuration for downstream model comparison. Among all evaluated supervised learning models, the XGBoost algorithm achieved the best predictive performance, with a mean absolute error (MAE) of 9.48 h, root mean squared error (RMSE) of 12.57 h, Pearson correlation coefficient (r = 0.535), and predictive R 2 (0.282). This model outperformed both null and covariate-only baseline models, although prediction errors varied across PMI intervals. Functional enrichment analysis of the final 400-gene signature revealed significant over-representation of immune effector processes, antigen processing and presentation, MHC class II-related pathways, and phagocytosis-associated biological processes.These findings demonstrate that postmortem brain transcriptomes contain reproducible PMI-associated signals and provide a candidate marker resource and standardized analytical framework for transcriptomic PMI prediction in forensic practice.