AI Estimates Time Of Death More Accurately Than Standard Methods Used Today

When a body is discovered, one of the first questions investigators ask is how long ago the person died. It sounds like something forensic science should have nailed by now. It hasn’t. The stiffness of a body, the temperature of the skin, the potassium level in the eye. These measures work best in the first one to three days after death, but their precision drops off after that. Now, however, a new AI breakthrough may offer a much less time-restrictive option.

Researchers at Linköping University trained an AI model to read the chemical fingerprints left in blood as a body breaks down after death. Tested on nearly 5,000 real forensic cases, the model estimated how long someone had been dead with a mean absolute error of 1.45 days, primarily across cases within the first two weeks after death, where most samples fell. For investigators reconstructing timelines, narrowing that window to a day or two could be meaningful. The study was published in Nature Communications.

Current standard methods only hold up reliably for one to three days after death. The model was trained on cases spanning up to 67 days after death.

Why Blood Holds the Answer

When a person dies, the body begins a predictable chemical unraveling. Cells stop receiving oxygen. Fats break down. Proteins fall apart into their building blocks. These processes leave measurable traces in the blood, and they follow a rough timeline tied to how long ago death occurred.

Researchers focused on molecules called metabolites, the small chemical byproducts of biological activity. After death, certain fat-related compounds tied to cell membrane function declined steadily, while amino acids and small protein fragments rose as tissue broke down. Together, those shifts created a chemical signature the AI could learn to match against a timeline.

What made the approach practical was where the blood came from. Rather than specially collected research samples, the team used blood already drawn for routine toxicology screenings, the standard tests forensic labs perform to check for drugs and poisons. The raw material for this kind of analysis already exists in forensic labs around the world, at no extra cost.

How the Model Was Built

Researchers trained the model on blood samples from 4,876 individuals whose deaths were investigated by Sweden’s National Board of Forensic Medicine between 2017 and 2019. Each sample came with a recorded time since death, based on forensic investigation records, giving the AI a verified answer to learn from. Though the post-mortem intervals in the dataset ranged from 1 to 67 days, 97% of cases fell within the first 13 days, and that is where the model’s performance was primarily evaluated.

The model analyzed thousands of chemical signals from each blood sample. When tested on cases it had never seen, it returned a mean absolute error of 1.45 days and a median error of just over one day.

For comparison, measuring potassium in the eye’s vitreous fluid is widely considered the current gold standard for timing a death, but it becomes unreliable after about 48 hours. The AI model maintained predictive performance beyond the first few days, where potassium-based methods typically widen substantially. The team also compared the neural network against several other machine learning approaches. All showed some ability to estimate time of death, but none matched the neural network’s accuracy.

The Model Survived a Real-World Stress Test

The most telling result came when researchers applied the model to a completely separate set of 512 cases collected in 2021, run on different lab equipment in a different setting. Models that perform well in one lab often fall apart in another. This one didn’t. Its error on the independent dataset came to 1.78 days, only slightly higher than on the original test cases.

Smaller forensic institutes also have reason to pay attention. Using a simpler model type called LASSO regression, the researchers found that as few as 256 to 512 blood samples were enough to produce a working model, with average errors of around 2.05 and 1.91 days respectively. Labs don’t need thousands of cases to get started.

One consistent weakness appeared at the edges. All models tended to overestimate how long someone had been dead when the answer was just one day, and underestimate at longer intervals beyond 11 days. The researchers tied this to the fact that most training cases clustered in the middle of the time range, a problem that more balanced future datasets could address.

What This Means for Investigators

For anyone working a suspicious death, when someone died can be just as consequential as how or why. Reconstructing a timeline, establishing the sequence of events around a death, confirming whether key facts line up, these are questions where a margin of one or two days is workable. A margin of seven days is not.

The study was conducted entirely on Swedish cases, where bodies are routinely refrigerated within 48 hours of death. Different conditions elsewhere, outdoor exposure, warmer climates, inconsistent storage, could affect results. The authors also noted that incorporating temperature data could sharpen accuracy further, since temperature affects how quickly a body breaks down after death.

Those are real constraints, and further validation across different countries and forensic settings will be needed before this becomes standard practice. But the core finding is hard to dismiss. A model trained on blood already collected for other reasons, holding its accuracy across a range stretching weeks, surviving a test on entirely different equipment in a different year. Forensic science now has a tool that deserves serious attention.

---