Related Experiment Video
Updated: Mar 27, 2026
Automated Joint Space Detection Improves Bone Segmentation Accuracy
Published on: November 28, 2025
Joint Damage Prediction in Non-Severe Hemophilia A with Artificial Intelligence
Ana Marco-Rico1, José Manuel Calvo-Villas2, Francisco-Jose Lopez-Jaime3
1Department of Hematology and Hemotherapy, Dr. Balmis General University Hospital-ISABIAL, Alicante, Spain.
Purpose:
Patients with non-severe hemophilia A (PwnSHA) can develop joint damage (JD). The objective was to identify a machine learning model based on routinely collected variables to predict the presence of JD in PwnSHA.
Patients And Methods:
A nationwide, multicenter, cross-sectional study was conducted. Clinical and laboratory variables to assess joint health were included. Predictors were age, target joint history, thrombin generation capacity, baseline factor VIII (FVIII) measured by one-stage clotting (FVIII-CLOT) and chromogenic (FVIII-CHR) assays, and the FVIII-CLOT/FVIII-CHR ratio. The joint condition was described using the HEAD-US score. JD was defined as HEAD-US >0. A Random Forest (RF) ensemble was trained with regression-based multiple imputation, z-scaling, and Synthetic Minority Oversampling within a stratified five-fold stratified cross-validation repeated 100 times. Support Vector Machine, Decision Tree, Gaussian Naïve Bayes and k-Nearest Neighbors were used as comparators. Model performance was assessed on held-out test folds, and 95% confidence intervals (CIs) were obtained by bootstrap resampling with 10,000 repetitions.
Results:
Eighty-four Spanish males ≥12 years old were enrolled. Forty-two percent (35/84) had JD. JD was present in 30% (3/10) of patients with moderate hemophilia and 43% (32/74) with mild hemophilia. The RF achieved an accuracy of 92.0% (95% CI: 90.72-93.31), a recall of 92.1% (95% CI: 90.87-93.41), a specificity of 91.9% (95% CI: 90.58-93.27), and an AUC-ROC of 0.92 (95% CI: 0.907-0.938), outperforming all alternative classifiers. Permutation-based feature importance identified age, target joint history, thrombin generation and the FVIII-CLOT/FVIII-CHR ratio as the most influential variables.
Conclusion:
The RF model identifies PwnSHA more likely to have prevalent, occult JD in a cross-sectional setting, enabling rapid triage for targeted HEAD-US evaluation. External and prospective validation in larger cohorts is now warranted to confirm generalizability and to facilitate integration into electronic health-record decision-support systems aimed at preserving long-term joint health in PwnSHA.
