Machine Learning Prediction of Treatment Failure After Middle Meningeal Artery Embolization for Subdural Hematoma
Bluyé DeMessie, Muhammed Amir Essibayi, Hamza Adel Salim, Alireza Karandish, Ahmed Y. Azzam, Deepak Khatri, Neil Haranhalli, Amanda Baker, Richard Zampolin, Allan L. Brook, Seon-Kyu Lee, Nimer Adeeb, Dhairya A. Lakhani, James Grist, Fulvio Zaccagna, Yan-Lin Li, Davide Simonato, Diego Alejandro Ortega, Nicole Cancelliere, Jose Danilo Diestro, Joseph Carnevale, Craig Schreiber, Atakan Orscelik, Zachary Abecassis, Spencer Raub, Georgios S. Sioutas, Abdulrhman Alsalama, Clairice Pearce, Giancarlo Salsano, Rahim Abo Kasem, Svetlana Kvint, Andrew Falzon, Vance Cantrell, Brian Holliday, Jefferson O. Abaricia, Pedram D. Maleknia, Leonardo Cruz-Criollo, Samantha Schimmel, Basel Musmar, Matthew Alexander, Jorge Rios Zermeno, Prateeka Koul, Ahmed Aljuboori, Dominik F. Vollherbst, Domagoj Gajski, Jared Cooper, Omar Alwakaa, Mohamad Ezzeldin, Christopher S. Ogilvy, Fawaz Al-Mufti, Vladimir Kalousek, Markus A. Möhlenbruch, Luca Scarcia, William W. Wroe, Hussein A. Zeineddine, Tang Si Zhao, Peter B. Sporns, Anil Gopinathan, Robert W. Regenhardt, Spiros L. Blackburn, Clemens M. Schirmer, Thien Huynh, Rabih G. Tawk, Fabio Settecase, Stavropoula Tjoumakaris, Pascal Jabbour, Kunal Vakharia, Mario Zanaty, Santiago Ortega‐Gutierrez, Jesse G. Jones, Marco Colasurdo, Hussein H. Nasser, Sri Hari Sundararajan, Pascal J. Mosimann, Erez Nossek, Eytan Raz, Sami Al Kasab, Alejandro M. Spiotta, Lucio Castellan, Bruno Del Sette, Ali Alaraj, Caterina Michelozzi, Davide Saraceno, Pietro Panni, Visish M. Srinivasan, Jan-Karl Burkhardt, Gaultier Marnat, Pietro Mario Santini, Michael R. Levitt, Giuseppe Lanzino, Waleed Brinjikji, Jared Knopman, Tareq Kass-Hout, Julian Spears, Thomas Marotta, Vitor Mendes Pereira, Maurizio Fuschi, Adam A. Dmytriw, David J. Altschul,BACKGROUND AND OBJECTIVES:
Middle meningeal artery embolization (MMAE) is effective for chronic subdural hematoma (SDH), but failure rates vary and risk prediction tools are lacking. This study developed and validated a machine learning (ML) model to predict MMAE treatment failure for SDH.
METHODS:
We conducted an international, multi-institutional retrospective cohort study of patients who underwent MMAE for SDH between 2019 and 2025. The cohort was split into training and test sets, stratified by treatment failure, with all hematomas from each patient assigned to a single cohort to prevent data leakage. Multiple ML algorithms were evaluated using 5-fold cross-validation, with gradient boosting selected as the final model. The primary outcome was treatment failure, defined as rescue surgical evacuation or persistent/recurrent SDH requiring retreatment. Model performance was assessed using area under the curve-receiver operating characteristic. Patients were stratified into low-risk (<5%), medium-risk (5%-15%), and high-risk (>15%) groups using a priori clinically defined thresholds.
RESULTS:
Among 692 treated hematomas in the test cohort, 64 (9.2%) experienced treatment failure. The gradient boosting model achieved an area under the curve-receiver operating characteristic of 0.856 (95% CI, 0.803-0.902) before calibration and 0.832 (95% CI, 0.766-0.889) after isotonic calibration, with excellent calibration (Brier score, 0.053). The most important predictors were age-platelet interaction, platelet-hemoglobin ratio, platelet count, SDH thickness, and age. Observed treatment failure was 1.6% (95% CI, 0.4%-5.7%) in the low-risk group (n = 124; 17.9%), 6.0% (95% CI, 4.3%-8.4%) in the medium-risk group (n = 516; 74.6%), and 59.6% (95% CI, 46.1%-71.8%) in the high-risk group (n = 52; 7.5%) (
CONCLUSION:
This ML model stratified MMAE patients into risk groups with a 37.2-fold difference in treatment failure. High-risk patients had 59.6% failure, whereas low-risk patients had 1.6% failure. Risk stratification may improve patient selection and monitoring.