An Interpretable Machine Learning Model for Predicting Early Neurological Deterioration Following Posttraumatic Acute Diffuse Brain Swelling.

backgroundDiffuse brain swelling (DBS) significantly contributes to intracranial hypertension and poses a substantial risk of early neurological deterioration (END). This study aimed to develop and validate various machine learning (ML) models for predicting END in patients with traumatic DBS.

methodsClinical data were retrospectively collected from 208 consecutive adult patients diagnosed with traumatic DBS on admission (within 6 h after injury) at two centers. END was assessed within 72 h of admission, and predictors for END were identified using least absolute shrinkage and selection operator regression and multivariate logistic regression analysis. Six ML algorithms were trained to develop prediction models. The performance of the ML models was evaluated by the area under the receiver operating characteristic curve (AUROC), Brier score, and decision curve analysis and was externally verified in the validation cohort. The optimal model was internally cross-validated, interpreted using Shapley Additive Explanations, and ultimately deployed as a Web-based risk calculator.

resultsA total of 79 patients experienced END, with an incidence of 38.0%. The four confirmed predictors of END were subdural hemorrhage, severe traumatic subarachnoid hemorrhage, hemoglobin levels, and fibrinogen levels. The extreme gradient boosting model outperformed the other five models in discrimination, achieving an AUROC of 0.879, and had better calibration and clinical utility. This model had an acceptable generalizability, achieving mean AUROCs of 0.762 ± 0.033 and 0.770 ± 0.109 in fivefold and tenfold cross-validations, respectively, and an AUROC of 0.862 in the validation cohort.

conclusionsThe developed ML model shows clinical promise in accurately predicting END following traumatic DBS. However, multicenter external validation remains essential before its widespread clinical application.