Söderberg, C. | bioRxiv Science

bioRxiv2026

Detecting and Subtyping Ketoacidosis from Metabolomic Patterns in Forensic Casework

Subtyping of ketoacidosis, a metabolic state characterized by blood acidification due to various causes, remains challenging in forensic casework. Postmortem omics samples paired with machine learning offers an independent tool to address this challenge. However, such data, especially related to real forensic cases, are rare. In Sweden, high-resolution mass spectrometry data routinely collected in forensic toxicology, can be leveraged for metabolomic analysis. Here, we integrate postmortem metabolomics and machine learning models to detect and subtype ketoacidosis-related deaths using real forensic cases in Sweden. From femoral blood samples of 109 alcoholic ketoacidosis cases, 220 diabetic ketoacidosis cases, 140 hypothermia cases, and 1,229 controls (hanging cases), we developed and tested three machine learning models, which achieved over 90% accuracy in ketoacidosis detection and over 80% in subtyping. Validation with independent cohorts (21 starvation cases, 29 alcoholic controls, and 40 diabetic controls) confirmed robustness with over 80% of starvation cases classified as ketoacidosis-related. Feature clustering highlighted metabolites such as cortisol to be important for subtyping. In summary, our findings demonstrate that combining machine learning with postmortem metabolomics enables accurate detection and subtyping of ketoacidosis-related deaths, which is useful for forensic casework.

systems biology↗

bioRxiv2025

The Human Metabolome and Machine Learning Improves Predictions of the Post-Mortem Interval

An accurate prediction of the time since death, known as the post-mortem interval (PMI), remains a critical research question in forensic and police investigations. Current methods, such as rectal temperature or vitreous potassium levels, only provide reliable PMI estimations up to 48-72 hours. In this study, we utilized metabolomic data from femoral whole blood samples of 4,876 individuals with known PMIs ranging from 1 to 67 days. We developed a neural network model that predicted PMI with a mean/median absolute error of 1.45/1.03 days in unseen test cases, outperforming six other machine learning architectures. To further highlight the biological signal, we performed pseudo time-series clustering of metabolic features used by the model, revealing 158 decreasing, 254 increasing, and 398 features with more complex patterns over the pseudo-time scale. Our findings also indicate that metabolomic data from approximately 256 individuals is sufficient to train a machine learning model for PMI prediction, making this approach widely applicable for researchers and forensic institutes worldwide.

systems biology↗