AI Models Predict Time of Death With One-Day Accuracy Using Blood Chemistry
Researchers at Linköping University and the Swedish National Board of Forensic Medicine have trained artificial intelligence models to estimate post-mortem intervals by analyzing blood metabolites, achieving a mean error of 1.45 days. The method extends forensic investigators' ability to determine when death occurred days after the fact, filling a gap that traditional approaches cannot bridge.
Current methods-measuring body temperature, assessing rigor mortis, or analyzing potassium in eye fluid-lose accuracy within one or two days. The AI approach works by tracking predictable chemical changes in blood that continue long after death.
How the Method Works
After death, tissues break down in measurable ways. Proteins degrade, fats break down, and cells fail energetically. These processes leave specific chemical signatures in blood called metabolites. The AI model learned to correlate these chemical patterns with elapsed time since death.
Rasmus Magnusson, postdoctoral fellow at Linköping University's Department of Biomedical Engineering, led the study. "Death is a strong biological signal," he said. The researchers used neural network techniques to identify patterns among hundreds of metabolites in blood samples.
Training and Testing the Model
The team accessed over 45,000 blood samples collected by Sweden's National Board of Forensic Medicine over nearly a decade. Of these, 4,876 samples with documented post-mortem intervals trained the AI model.
When tested on unseen cases, the model achieved a median error of 1.03 days-meaning more than half of predictions fell within one day of the actual post-mortem interval. A separate validation on 512 samples collected in a different year and analyzed on different equipment produced a median error of 1.29 days.
Magnusson noted the practical advantage: "A few hundred individuals are enough to build corresponding models, which makes our method useful even in laboratories worldwide that don't have access to as much data."
What Metabolites Reveal
The study identified three key metabolic processes that signal elapsed time: lipid breakdown, mitochondrial dysfunction, and proteolysis-the natural degradation of fats, energy-producing cell components, and proteins. Together, these signals create a molecular map of how long death has occurred.
Carl Söderberg, forensic pathologist at the National Board of Forensic Medicine, said the approach helps solve cases where traditional methods fail. "This new tool gives us better opportunities to assess how long someone has been deceased even when a long time has passed since their death, which is of great importance especially in more complex cases."
Practical Applications for Investigation
Accurate post-mortem estimates help police allocate resources more effectively. Henrik Green, professor of forensic sciences at Linköping University, explained the stakes: investigators can focus on witnesses and evidence from the correct period in the deceased person's life rather than casting a wider net.
Unlike physical methods, the AI approach is less affected by environmental variables like temperature or humidity. This consistency makes it more reliable across different climates and storage conditions.
Limitations and Next Steps
The method requires blood samples from autopsy, which are not always available. Environmental exposure, cause of death, and individual variation can affect metabolite levels, so researchers caution that the method needs testing across diverse circumstances.
Future work aims to improve temporal precision further. Researchers plan to incorporate more exact timestamps for death to estimate not just the day but the approximate time of day when death occurred. Expanding datasets to include tissue samples or other bodily fluids may also strengthen predictions.
The findings appear in Nature.
Professionals working with forensic data or developing analytical methods may benefit from understanding how machine learning processes complex biological signals. AI Data Analysis Courses and AI Research Courses cover the methodologies used in this type of analysis.
Your membership also unlocks:
