Back to 2026 ePosters
Prognostication of spontaneous recovery following Brachial Plexus Injury using machine learning.
Matthew Uzelac, BS
1, David Cholok, MD
1; Anna Luan, MD, MS
2(1)Stanford University, Palo Alto, CA, (2)Plastic Surgery, Stanford University, Stanford, CA
Introduction: The optimal timing of surgical intervention remains indeterminate following traumatic brachial plexus injuries (BPI) due to the uncertain degrees and rates of spontaneous functional recovery. Recovery is dependent on a multitude of injury- and patient-specific factors and is thereby difficult to accurately predict. This study was conducted to leverage machine learning to identify demographic, comorbid, and electrodiagnostic characteristics associated with spontaneous recovery of upper limb motor function.
Materials and Methods: The authors developed a gradient-boosting model using de-identified data from 120 adult patients with traumatic BPI seen at a single institution from 2006-2024. This was done using the lightgbm package in RStudio (Posit team, 2025). Predictors included demographic characteristics (age, sex, and race), electrodiagnostic findings, and comorbidities. Patients were included if electrodiagnostic studies were performed and medical research council (MRC) graded motor function was documented longitudinally for shoulder abduction, external rotation, and flexion, elbow flexion and extension, and wrist flexion and extension. Absence of meaningful recovery was defined as progression to operative intervention, or persistent MRC grades <4 after 6 months post-injury. The dataset was partitioned into training and testing cohorts using an 80:20 split. Leave-one-out cross validation was used to refine the model's hyperparameters on the training cohort. Performance metrics were then assessed, and learning curve analyses were performed to assess for overfitting.
Results: The areas under the curves (AUC) were 0.991 and 0.975 for the training and testing sets respectively. The sensitivities and specificities were 100% and 92%, and 100% and 95% for training and testing sets. Index electrodiagnostic metrics including fibrillations and positive sharp waves were significantly inversely correlated with rates of spontaneous recovery.
Conclusions: The authors demonstrate the capability of advanced machine learning to model spontaneous motor recovery following traumatic BPI. Early identification of patients unlikely to achieve spontaneous recovery could enable clinicians to expedite nerve transfers or other related procedures before the window of opportunity closes, thereby improving functional outcomes and reducing the need for salvage procedures.
Back to 2026 ePosters
