Related Experiment Video
Updated: Sep 14, 2025
Author Spotlight: Automated Lifespan Monitoring – Discovering Aging Dynamics with the Lifespan Machine
Published on: January 26, 2024
Advances in postmortem interval estimation: A systematic review of machine learning and metabolomics across various
Abdulkreem Abdullah AlJuhani1, Rodan Mahmoud Desoky2, Abdulaziz A Binshalhoub3
1Department of Surgery King, Abdulaziz University Hospital, Riyadh, SA, Saudi Arabia. Kroomx@outlook.com.
Background:
Traditional postmortem interval (PMI) estimation methods rely on observable changes such as rigor mortis, livor mortis, and algor mortis but are often affected by environmental factors. Metabolomics, combined with techniques like nuclear magnetic resonance (NMR) and mass spectrometry, improves accuracy by identifying biochemical changes postmortem. Machine learning methods such as Principal Component Analysis (PCA), Partial Least Squares (PLS), and Support Vector Machines (SVMs), enhance PMI predictions by analyzing metabolite data. This review aims to summarize advances in using machine learning for PMI estimation and identify the optimal combination of tissue samples and algorithms for accurate predictions.
Methods:
We retrieved relevant articles up to September 2024 from PubMed, Scopus, Web of Science, IEEE, and Cochrane Library. Data were extracted from eligible studies by two independent reviewers. This included the number and species of subjects, tissue sample used, PMI range in the study, metabolic profiling technique, machine learning algorithms, potential PMI markers, and model performance.
Results:
We compared machine learning models for PMI estimation across various tissues. Zhang et al. (2022) had the best performance with a random forest (RF) model using cardiac blood, achieving a mean absolute error (MAE) of 1.067 h by selecting key metabolites. Wu et al. (2017) followed with an orthogonal signal-corrected PLS model (R2 > 0.99, MAE 1.18-2.37 h). Lu et al. (2022) achieved 93% accuracy with a multi-organ stacking model. Other promising models include Zhang et al.'s (2017) nu-SVM on pericardial fluid (RMSE = 2.38 h) and Sato et al.'s (2015) PLS model on cardiac blood (MAE = 5.73 h).
Conclusion:
Cardiac blood is best for short PMIs with random forest models, while skeletal muscle and stacking models excel for longer PMIs. Future studies should refine and validate these findings as well as extend the findings to human subjects.
