A machine learning approach to MRI screening for brain arteriovenous malformations in hereditary hemorrhagic telangiectasia.
Authors
Affiliations (6)
Affiliations (6)
- Department of Neurosurgery, Fondazione Policlinico Universitario A. Gemelli IRCCS, Università Cattolica del Sacro Cuore, Rome, Italy.
- Department of Computer Science, University of Bari Aldo Moro, Bari, Italy.
- Department of Genetics, Fondazione Policlinico Universitario A. Gemelli IRCCS, Università Cattolica del Sacro Cuore, Rome, Italy.
- Department of Translational Medicine and Surgery, Fondazione Policlinico Universitario A. Gemelli IRCCS, Università Cattolica del Sacro Cuore, Rome, Italy.
- Department of Aging, Orthopedic, and Rheumatologic Sciences, Fondazione Policlinico Universitario A. Gemelli IRCCS, Università Cattolica del Sacro Cuore, Rome, Italy.
- Department of Neurosurgery, Fondazione Policlinico Universitario A. Gemelli IRCCS, Università Cattolica del Sacro Cuore, Rome, Italy. [email protected].
Abstract
Hereditary hemorrhagic telangiectasia (HHT) is an inherited vascular disorder characterized by telangiectasias and arteriovenous malformations (AVMs) in multiple organs. Although brain MRI screening is recommended in all patients with HHT, in many centers, this may be difficult to implement due to organizational constraints. The aim of this study was to develop a machine learning model to predict the presence of bAVMs and other cerebrovascular malformations in patients with HHT, with the goal of supporting risk stratification in centers where genotype information is already available and identifying patients who may benefit from closer and more conservative surveillance. All patients with confirmed diagnosis of HHT, complete genetic work-up, and radiological screening (2018-2026) were included for the analysis. Two end points were defined: probability of detecting bAVMs, and probability of detecting any cerebrovascular findings. Analysis was performed using Python and R. A total of 142 patients with confirmed HHT (meeting ≥ 3 Curaçao criteria) were included. For bAVM prediction, both models showed good discrimination, with an AUC of 0.84 (95% CI: 0.73-0.94) for Logistic Regression (LR) and 0.85 (95% CI: 0.74-0.96) for Random Forest (RF). Despite the slightly higher AUC, RF showed lower specificity, PPV, and F1 score, higher sensitivity (0.73 vs. 0.69) and same NPV. LR also showed better calibration, with an intercept of -0.30 (95% CI: -1.38-0.78) and slope of 0.87 (95% CI: 0.34-1.40), compared with 1.39 (95% CI: -0.46-3.23) and 2.39 (95% CI: 0.64-4.15), respectively, for RF. In contrast, models predicting overall cerebrovascular malformations showed limited discrimination and poor calibration with both LR and RF. For bAVM prediction, LR identified deletion mutations, ENG, family history, ACVRL1, and truncated variants as the most influential features, while RF SHAP analysis highlighted ACVRL1, ENG, age, pulmonary AVMs, and family history. Effective risk stratification was achieved by the phenotype-specific bAVM model, supporting its potential applicability. Upon external validation this tool may enable an efficient allocation of screening strategies.