Demo · Brain strain results

See the impact, region by region.

We developed a machine learning model using the extreme gradient boosting (XGBoost) algorithm to rapidly predict brain deformation and provide brain strain estimates for critical brain regions of interest. Choose a preloaded archetype impact below — or stream impacts live from an Arduino-instrumented mouthguard. The model returns a max-principal-strain (MPS) value per white-matter region.

Each entry is the medoid of a cluster of real instrumented-mouthguard recordings, grouped by sport-dataset. Cluster sizes range from a few to several hundred impacts.

Anatomy

Polished voxel clouds · tonemapped · bloom + SMAA

Loading anatomy…
Render

iMG kinematics

Drag along the time axis to zoom · hover for per-axis values · double-click to reset.

Select an impact above to load its kinematics.

Brain strain predictions

90th-percentile max-principal-strain (MPS90) per region · impact 1

Whole brainoverall0.293± 0.020
  • Corticospinal Tract0.333± 0.023
  • Mean Skeleton0.290± 0.024
  • Superior Longitudinal Fasciculus0.281± 0.025
  • Superior Longitudinal Fasciculi Temporal0.275± 0.026
  • Inferior Longitudinal Fasciculus0.268± 0.024
  • Corpus Callosum Body0.265± 0.023
  • Uncinate Fasciculus0.255± 0.023
  • Cingulum Cingulate Hippocampus0.250± 0.021
  • Brain Stem0.242± 0.021
  • Inferior Frontal-Occipital Fasciculus0.239± 0.023
  • Cingulum Cingulate0.236± 0.023
  • Anterior Thalamic Radiation0.231± 0.022
  • Forceps Major0.227± 0.023
  • Corpus Callosum Splenium0.204± 0.023
  • Corpus Callosum Genu0.195± 0.024
  • Forceps Minor0.192± 0.023

Extrapolation — outside the validated envelope. The model was validated on elite-male rugby only; this is Boxing. Treat predictions as indicative, not validated. (R² 0.76–0.85 on rugby; ± RMSE shown.)

MPS90 from the XGBoost regressors (TWP/model/model_XGB_finetuned_final/). Accuracy & envelope: Chan et al. 2025, Table 2 (validation_chan2025.json). No universally accepted strain injury threshold exists — colours are indicative.

Methodology

Data capture

Instrumented mouthguards record translational and rotational head kinematics during head acceleration events. Sensors sample at 1 kHz; each impact window spans ~100 ms.

Feature extraction

Per-axis kinematic time series are reduced to four scalar features — range andsqrt(|max|) of resultant rotational velocity and rotational acceleration — and fed to the model.

Inference

A fine-tuned XGBoost regressor returns a max-principal-strain estimate per white-matter region. On iOS the models run on-device via Core ML; on the web the archetype impacts are pre-computed with the same Python pipeline, while live-mouthguard mode runs the models in your browser via ONNX Runtime Web.

Visualisation

Each region is rendered as a dense voxel cloud in its anatomical location (JHU white-matter atlas, MNI152 1 mm space) inside a translucent brain shell, and can be isolated individually. Per-region strain is encoded in the same colourmap the iOS app uses — green to yellow through red, with elevated regions in violet.