Radiomics-based machine learning model for predicting secondary decompressive craniectomy in TBI patients after...

Secondary decompressive craniectomy (DC) is commonly integrated into tiered therapeutic protocols in the intensive care unit (ICU) to manage elevated intracranial pressure following traumatic brain injury (TBI). Identifying high-risk patients in advance could enable early intervention and help prevent further deterioration. This study aims to develop a machine learning-based predictive model using radiomics to assess the likelihood of secondary DC in TBI patients.

A total of 65 patients were enrolled and divided into training and test cohorts through stratified random sampling with a 7:3 ratio. Radiomic features were extracted from pre-evacuation CT data. The most relevant features were identified through importance score computation, and various predictive models were assessed using distinct machine learning algorithms and data sources. Model performance was benchmarked to construct an optimal predictive model.

No statistically significant differences were observed in demographic and clinical characteristics between the DC and non-DC groups. The model based solely on demographic and clinical data did not achieve satisfactory performance, with an AUC below 0.5 in the test cohort. In radiomic modeling, the randomForest model demonstrated consistent performance, achieving an AUC of 0.83 in the test cohort. The multiomic model, which incorporated demographic, clinical, and radiomic features, showed improved predictive performance, with the cforest model achieving an AUC of 0.87 in the training cohort and 0.86 in the test cohort.

We developed radiomics-based predictive models to assess the likelihood of progressively refractory intracranial hypertension leading to secondary DC in a selected cohort of TBI patients who had undergone emergent craniotomy for hematoma evacuation with bone flap replacement. The model relying solely on radiomic features extracted from the lesion demonstrated satisfactory performance. When these features were integrated with demographic and clinical data to create a multiomic model, predictive performance further improved. These findings highlight the model's potential to identify high-risk patients, enabling early intervention to prevent further deterioration.