A new tool to aid forensic investigations of traumatic brain injuries (TBI), has been developed by a team of researchers in collaboration with practitioners from law enforcement, healthcare and industry.
The advanced physics-based AI-driven technology introduces a mechanics-informed machine learning framework to help police and forensic teams accurately predict TBI outcomes based on described assault scenarios.
The new study, led by the University of Oxford, in collaboration with Cardiff University, Thames Valley Police, the National Crime Agency, the John Radcliffe Hospital and Lurtis Ltd, is published in Nature Communications Engineering.
TBI is a critical public health issue, with severe and long-term neurological consequences.
In forensic investigations, determining whether an impact could have caused a reported injury is crucial for legal proceedings, yet there is currently no standardised, quantifiable approach to do this.
These findings demonstrate how machine learning tools informed by mechanistic simulations could provide evidence-based injury predictions, to improve the accuracy and consistency of TBI investigations.
Lead researcher Antoine Jérusalem, Professor of Mechanical Engineering in the Department of Engineering Science at the University of Oxford, said: "This research represents a significant step forward in forensic biomechanics."
By leveraging AI and physics-based simulations, we can provide law enforcement with an unprecedented tool to assess TBIs objectively.
The study's AI framework, trained on real, anonymised police reports and forensic data, achieved remarkable prediction accuracy for TBI-related injuries:
- 94% accuracy for skull fractures
- 79% accuracy for loss of consciousness
- 79% accuracy for intracranial haemorrhage (bleeding within the skull)
In each case, the model showed high specificity and high sensitivity (a low rate of false positive and false negative results).
The framework uses a general computational mechanistic model of the head and neck, designed to simulate how different types of impacts—such as punches, slaps, or strikes against a flat surface—affect various regions.
This provides a basic prediction of whether an impact is likely to cause tissue deformation or stress. However, it does not predict on its own any risk of injury. This is done by an upper AI layer which incorporates this information with any additional relevant metadata, such as the victim's age and height before providing a prediction for a given injury.
The researchers trained the overall framework on 53 anonymised real police reports of assault cases. Each report included information about a range of factors which could affect the blow's severity including age, sex, body build of the victim/offender.
This resulted in a model capable of integrating mechanical biophysical data with forensic details to predict the likelihood of different TBIs occurring.
Co-author Dr Mike Jones is a Forensic Researcher and Consultant at Cardiff University's School of Engineering and member of the International Brain Mechanics and Trauma Laboratory at the University of Oxford. He routinely assists in the investigation of the causes of murder, assault, accident and suicide.
He said: "An Achilles heel of forensic medicine is the assessment of whether a stated witnessed or inferred mechanism of injury, often the force, matches the observed injuries."
With the application of machine learning to forensic investigation, each additional case contributes to improving the overall understanding of the association between the mechanism of cause, primary injury, pathophysiology and outcome.
When the researchers assessed which factors had the most influence on the predictive value for each type of injury, the results were remarkably consistent with medical findings. For instance, when predicting the likelihood of skull fracture, the most important factor was the highest amount of stress experienced by the scalp and skull during an impact. Similarly, the strongest predictor of loss of consciousness was the stress metrics for the brainstem.
The research team insists that the model is not intended to replace the involvement of human forensic and clinical experts in investigating assault cases. Rather, the intention is to provide an objective estimate of the probability that an assault was the true cause of a reported injury.
The model could also provide a tool to identify high-risk situations, improve risk assessments, and develop preventive strategies to reduce the occurrence and severity of head injuries.
Sonya Baylis, Senior Manager at the National Crime Agency, said: "When we're looking at brain injuries, every detail matters – and this tool helps us get those details right. It's not just about having better technology, it's about making sure we can piece important events together."
By helping forensic specialists and the police to work smarter, and ultimately making sure justice is served, we can make our communities safer for everyone.
The paper, ' A mechanics-informed machine learning framework for traumatic brain injury prediction in police and forensic investigations ', is published in Nature Communications Engineering.
