Back to 2025 Abstracts
Nerve AI - A Machine Learning Algorithm for Detection of Nerve Pain in the Head and Neck
Giulia L. Mönnink
*1, Merel H. Hazewinkel
1, Weishen Pan
2, Siddharth Simon
2, Thomas R. Champion
3, William G. Austen
4, Darryl Sneag
5, Fei Wang
2, Lisa Gfrerer
11Division of Plastic and Reconstructive Surgery, Weill Cornell Medicine, New York, NY; 2Department of Population Health Sciences, Weill Cornell Medicine, New York, NY; 3Clinical & Translational Science Center, Weill Cornell Medicine, New York, NY; 4Division of Plastic and Reconstructive Surgery, Massachusetts General Hospital, Boston, MA; 5Department of Radiology and Imaging, Hospital for Special Surgery, New York, NY
Background: Nerve pain screening in patients with headache disorders requires specialized clinical knowledge that is frequently not readily accessible at the initial point-of-care. This results in restricted access to early treatment, thereby increasing the risk of chronic pain refractory to therapy and predisposing patients to long-term disability and narcotic dependence. Therefore, there is a need for a simple screening tool that enables timely and accurate diagnosis of nerve pain.
Methods: NerveAI, a 3D head and neck model that allows for an Artificial Intelligence-driven pattern recognition of nerve pain, was developed. The algorithm was trained using 1,299 head and neck pain drawings based on anatomic nerve paths and radiation patterns. Input features were constructed using coordinates of the pain drawings. We evaluated the model performance using 5-fold nested cross-validation with AUROC as the primary metric.
Results The best performing model to detect nerve pain was the multilayer perceptron model (AUROC: 0.879±0.044). When identifying specific nerve pain types, the model demonstrated high AUROC values of 0.928 (±0.025), 0.930 (±0.017), 0.884 (±0.031), and 0.954 (±0.025) for occipital, frontal, temporal, and trigeminal neuralgia, respectively.
Conclusion Our findings demonstrate NerveAI's potential to enable broader screening by non-specialized providers, allowing for early diagnosis and treatment to improve patient outcomes regardless of geographic, socioeconomic, or healthcare literacy barriers.
Figure 5. Performance of machine learning models in predicting nerve pain across four different types of nerve pain. The graphs show AUROC values for four different machine learning models: XGBoost (extreme gradient boosting), LR (Logistic Regression), RF (Random Forest), and MLP (multilayer perceptron). Results are presented for four affected regions: Occipital, Frontal, Temporal, and Trigeminal Nerve.
Back to 2025 Abstracts
