The human metabolome and machine learning improves predictions of the post-mortem interval

Area of Science:

• Forensic Science
• Biochemistry
• Computational Biology

Background:

• Accurate post-mortem interval (PMI) estimation is crucial for forensic investigations.
• Current methods like rectal temperature and vitreous potassium are limited to 1-3 days post-death.
• There is a need for more reliable and longer-duration PMI estimation techniques.

Purpose of the Study:

• To develop and validate a machine learning model for predicting the post-mortem interval using metabolomic data.
• To explore metabolite dynamics related to post-mortem changes.
• To assess the generalizability and scalability of the developed model.

Main Methods:

• Utilized metabolomic data from routine toxicological screenings of femoral whole blood samples (n=4876) with known PMIs (1-67 days).
• Developed a neural network model and compared its performance against six other machine learning architectures.
• Applied pseudo-time series clustering to identify key metabolite dynamics and validated the model on independent datasets from different years and platforms.

Main Results:

• The neural network model achieved a mean absolute error of 1.45 days and a median absolute error of 1.03 days on unseen test data.
• The model outperformed six other machine learning architectures in predicting PMI.
• The model demonstrated generalizability on independent test data (mean absolute error 1.78 days, median absolute error 1.29 days) despite cross-platform variability.
• Scalability was demonstrated, with robust models trainable using a few hundred cases.

Conclusions:

• Post-mortem metabolomics from routine toxicological samples can accurately predict the post-mortem interval beyond 3 days.
• The developed neural network model offers a significant advancement over existing PMI estimation methods.
• This approach provides a transferable framework for future forensic applications, enhancing investigative capabilities.