Crash Victims as Validation Data for Patient-Specific Leg Injury Models

When a car strikes a pedestrian's leg, whether the tibia (the shin bone) breaks depends on the impact and on the leg. Road safety models handle the impact well: vehicle speed, bumper height, and velocity change. They treat the leg as a standard object described at most by an age bracket. Surgical planning goes the other way. It builds detailed computer models of one patient's bone from their CT scan, but those models are rarely checked against real fractures, and injury biomechanics studies usually rest on a single reconstructed limb. We propose to join the two by treating every imaged crash victim as a validation experiment that nobody had to run on purpose. The vehicle's event data recorder, a black box that logs speed and deceleration, supplies an estimate of the load. The trauma CT supplies both the leg's anatomy and the actual outcome, fractured or not, and where. A patient-specific finite element model is a physics simulation that divides the bone into small elements whose stiffness comes from the patient's CT brightness. Such a model can be built for each victim, driven with the reconstructed load, and scored on whether it predicts the injury that actually happened.

The central question is whether individual anatomy adds real predictive power for injury beyond crash-level variables and simple bone-density numbers. If it does, the result is fracture risk curves that are specific to a person and validated on a cohort, where current work offers single-specimen demonstrations. If it does not, the finding is still useful: it would tell the field that expensive per-patient simulation is not needed for injury prediction.

Core Claim

We claim that per-patient finite element models of the lower leg, built from clinical CT and driven by recorder-derived impact loads with honest uncertainty, will predict which pedestrians fracture their tibia, and in which region, better than crash-level machine learning and better than regressions on bone density and bone dimensions. We further claim that the advantage will show most clearly for fracture location. A single number such as bone density cannot encode location, while a stress field can. The new element is not the finite element model itself. It is the validation design: a multi-site cohort of real injuries used as the ground truth that destructive human testing can never provide ethically.

Grounding in Prior Work

Four lines of work set up the proposal. Two motivate it, one supplies the method, and one constrains it.

On the road safety side, Qiu et al. combine crash simulations with recorder data and real occupant injuries. They show that velocity change, belt status, and airbag timing drive head and chest trauma. They also report recorder velocity-change errors of 3.5–6.6% and note that occupants are modeled as standardized bodies. We reuse their idea of coupling recorder data with injury outcomes and their error range for propagating uncertainty. We replace the standardized body with each victim's own anatomy.

Large-scale severity models motivate the gap from the opposite direction. AutoML with SHapley Additive exPlanations (SHAP, a method that attributes a prediction to its input features) on 2.3 million records finds that systemic predictors such as speed, location, and occupant age dominate (Predicting and Explaining Traffic Crash Severity). Ordinal classifiers with minority oversampling and calibrated probabilities improve detection of severe injury (Crash Injury Severity Prediction Using an Ordinal Classification Machine Learning Approach). Both work at the level of the person record and cannot see bone. Age appears in them only as a proxy for anatomy that we propose to measure directly.

The patient-specific simulation line supplies the method. The tibial impact study assigns five-zone graded material properties from CT grayscale and analyzes tibial stress under different postures and lateral impacts. It does this on a single reconstructed model and names simplified boundary conditions as a limitation. The lumbar spine pilot shows that routine clinical CT can support finite element estimates of failure load.

That same lumbar result also constrains us. Its predicted failure load correlates with quantitative CT bone mineral density at $R^2 = 0.8$, which suggests that much of what finite element analysis captures may already sit in a bone density scalar. We take that threat seriously in the baselines below.

The experimental validation literature sets the other constraint. Destructive testing cannot be run on patients (TAVI durability review). Scaled impact tests can validate explicit finite element models to within 5% force error when dimensionless similarity laws are used (offshore stiffened plate study). We borrow that scaling logic for a small laboratory anchor. We do not take it as proof that bone behaves like steel plate.

Finally, the single-case literature explains why the existing tibial evidence is insufficient. One subject shows that an injury mechanism is possible, not how common it is (vesselplasty case). Systems tuned to one person fail on others (brain-computer interface study).

Research Program

The cohort will be drawn retrospectively from multiple trauma centers. It will include pedestrians struck by vehicles whose recorders logged speed at impact, and whose trauma workup included CT covering the tibia. Cases are those with a tibial fracture, labeled by region (proximal, shaft, distal) using the standard orthopaedic trauma classification. Controls are struck pedestrians with lower-leg CT and no tibial fracture. We target several hundred of each and will report the achieved counts honestly.

For fractured legs, the intact geometry is approximated by mirroring the contralateral tibia. That substitution is itself validated on controls by mirroring one leg and comparing it to the other.

Each leg gets a segmented mesh with CT-derived graded material properties, following the tibial study's approach. Loading comes from recorder impact speed, vehicle front geometry, and bumper height. Unknown quantities, namely stance posture, which leg was struck first, and recorder error drawn from the 3.5–6.6% range, are sampled rather than fixed. Each patient's model therefore outputs a probability of fracture and a probability for each region, not a single yes or no. These per-patient outputs feed a calibrated injury model alongside crash-level features.

A small laboratory anchor checks the constitutive assumptions before cohort evaluation. It consists of drop-weight impacts on composite bone surrogates of varied density, compared using dimensionless force–displacement scaling.

Baselines

The naive reference is a logistic injury risk curve in impact speed, age, and sex, fitted on the same cohort. This is the standardized-body assumption that current road safety practice makes. Beating it only shows that we used more information, so it is a floor, not the target.

The strongest crash-level competitor is the pipeline of Castellani et al.: an AutoML ensemble explained by SHAP. We combine it with the ordinal treatment, minority oversampling, and isotonic calibration of the ordinal classification paper. It is run on exactly our cohort's crash-level features, with the same cross-validation folds and a hyperparameter search budget equal in trials to the one given to our combined model. The file contains no reported result on Lacuna for tibial fracture prediction by these methods, so its score will be established by us on our own splits. Beating it means a statistically significant improvement in Brier score and area under the ROC curve on held-out sites, with no loss of calibration.

The strongest anatomy competitor is a regression on CT-derived scalars: mean cortical density, cortical thickness, tibial length and cross-sectional area, and impact speed. It gets the same tuning budget. This baseline tests the lumbar study's hint that density explains most of what finite element models add. Beating it is the result that justifies simulation at all.

Two ablations isolate what personalization contributes. The first is a single template tibia scaled to each patient's height and mass, which amounts to the one-subject model applied to everyone. The second is each patient's own geometry assigned homogeneous material properties. These separate how much comes from shape and how much from material distribution.

Why not just…?

Why not just regress on bone density and dimensions? This is the most serious threat. The lumbar spine pilot found an $R^2$ of 0.8 between finite element failure load and density, and fracture occurrence could turn out to be largely a density-times-speed problem. The scalar baseline is designed to settle this. If it matches finite element analysis on occurrence, we will say so and narrow the claim to fracture location, which scalars cannot localize. If finite element analysis loses on location too, the proposal's thesis fails.

Why not just add more crash records to the machine learning model? Even 2.3 million records (Castellani et al.) carry no bone information, so age remains a loose stand-in for anatomy. Those models also say nothing about where a fracture occurs, which is the information surgeons need. The head-to-head comparison on held-out sites will show whether anatomy adds signal beyond scale of data.

Why not just calibrate one high-fidelity human body model against cadaver tests? That approach is the single-case problem written into engineering practice: one validated geometry then gets applied to everyone. Cadaver donors are also skewed toward the elderly. The scaled-template ablation measures directly what is lost when one geometry stands in for the population.

Proposed Experiments and Decisive Checks

The first check is fidelity of the mirroring step. Mirrored-contralateral geometry must reproduce the true ipsilateral tibia on controls with surface error small relative to cortical thickness. If it does not, fractured cases must be dropped or modeled differently.

The second check is the laboratory anchor. Surrogate impacts should collapse onto a single dimensionless curve, and the finite element model should reproduce peak force at a tolerance we fix in advance of the impact tests.

The decisive cohort test is leave-one-site-out evaluation of fracture occurrence. It uses Brier score, ROC area, and calibration slope for the naive, crash-ML, scalar, template, homogeneous, and full models. It is paired with top-region accuracy for fracture location. Success requires the full model to beat both the crash-ML and scalar baselines on held-out sites.

We will also report a variance decomposition across posture, load, and anatomy samples. This quantifies how much injury uncertainty comes from the person rather than the crash, and it would itself be a new number for road safety.

A negative result is informative whichever way it falls. If the scalar regression wins, clinical CT density should enter crash severity models directly, and per-patient simulation is not needed for this purpose.