Establishment and validation of a prediction model for self-absorption probability of chronic subdural hematoma

Abstract

Background: Chronic subdural hematoma (CSDH) is common in elderly people with a clear or occult traumatic brain injury history. Surgery is a traditional method to remove the hematomas, but it carries a significant risk of recurrence and poor outcomes. Non-surgical treatment has been recently considered effective and safe for some patients with CSDH. However, it is a challenge to speculate which part of patients could obtain benefits from non-surgical treatment. Objective: To establish and validate a new prediction model of self-absorption probability with chronic subdural hematoma. Method: The prediction model was established based on the data from a randomized clinical trial, which enrolled 196 patients with CSDH from February 2014 to November 2015. The following subjects were extracted: demographic characteristics, medical history, hematoma characters in imaging at admission, and clinical assessments. The outcome was self-absorption at the 8th week after admission. A least absolute shrinkage and selection operator (LASSO) regression model was implemented for data dimensionality reduction and feature selection. Multivariable logistic regression was adopted to establish the model, while the experimental results were presented by nomogram. Discrimination, calibration, and clinical usefulness were used to evaluate the performance of the nomogram. A total of 60 consecutive patients were involved in the external validation, which enrolled in a proof-of-concept clinical trial from July 2014 to December 2018. Results: Diabetes mellitus history, hematoma volume at admission, presence of basal ganglia suppression, presence of septate hematoma, and usage of atorvastatin were the strongest predictors of self-absorption. The model had good discrimination [area under the curve (AUC), 0.713 (95% CI, 0.637–0.788)] and good calibration (p = 0.986). The nomogram in the validation cohort still had good discrimination [AUC, 0.709 (95% CI, 0.574–0.844)] and good calibration (p = 0.441). A decision curve analysis proved that the nomogram was clinically effective. Conclusions: This prediction model can be used to obtain self-absorption probability in patients with CSDH, assisting in guiding the choice of therapy, whether they undergo non-surgical treatment or surgery.