Machine learning models using routinely available clinical and radiographic variables did not achieve clinically actionable risk stratification for cSDH recurrence, with discriminative capacity insufficient to identify a low-risk subgroup suitable for de-escalated surveillance.
Postoperative recurrence requiring reoperation occurred in 30.1% of patients in this cohort.
XGBoost achieved the highest cross-validated ROC AUC in the training set, outperforming logistic regression but not Random Forest.
The final XGBoost model achieved a ROC AUC of 0.688 on the held-out test set, indicating modest discriminative performance.
Hematoma volume, coagulation parameters, and disease severity markers were the most influential predictors, though effect sizes remained modest.
At a clinically relevant 90% sensitivity threshold, specificity was only 30.3%, allowing potential imaging reduction in roughly one-third of non-recurrence patients.
The study concluded that cSDH recurrence is likely driven by factors not captured in standard clinical assessment, supporting uniform or symptom-driven imaging strategies over risk-stratified approaches.
This was a retrospective, single-center study with internal validation only, limiting the generalizability of the findings.
This research examined whether artificial intelligence (machine learning) could predict which patients would need a second surgery after treatment for chronic subdural hematoma (cSDH) — a type of blood clot that collects on the surface of the brain. Using data from 564 patients treated at a single hospital over eight years, the researchers tested three different machine learning approaches using 31 variables routinely collected during clinical care, such as blood clot size, blood clotting test results, and measures of how sick the patient was at admission. About 30% of patients in this study experienced a recurrence requiring reoperation, which is consistent with rates reported in prior research. The best-performing model (XGBoost) was only modestly accurate at distinguishing patients who would need reoperation from those who would not, achieving an AUC of about 0.69 on an independent test group — a value where 1.0 would be perfect and 0.5 would be no better than chance. When the model was set to catch 90% of true recurrences (high sensitivity), it could only correctly rule out recurrence in about 30% of patients who did not actually recur (low specificity). This means the model could not reliably identify a 'low-risk' group that could safely skip follow-up imaging. The researchers noted that performance limitations appeared to reflect the information content of the available variables rather than technical issues with the models themselves. This research suggests that the factors driving cSDH recurrence may not be well captured by standard clinical measurements currently recorded in patient records, making accurate individual-level prediction very difficult with existing data. Rather than supporting risk-based imaging schedules — where low-risk patients get less monitoring — the findings suggest that uniform follow-up or symptom-driven imaging strategies may be more appropriate until better predictive variables or external validation studies are available. Future work may need to identify new biological or imaging markers associated with recurrence risk.
Hamou H, Kernbach J, Ridwan H, Fay-Rodrian K, Clusmann H, Hoellig A, et al.. (2026). Classification of recurrence status after surgical treatment of chronic subdural hemorrhage - A machine learning approach.. PloS one. https://doi.org/10.1371/journal.pone.0346756