Multimodal ML with CT habitat heterogeneity predicts radiation proctitis in cervical cancer

A multimodal machine learning model integrating CT habitat heterogeneity predicted radiation proctitis with an AUC of 0.87 on an independent test set (retrospective, n=100). All high-risk patients per the nomogram developed radiation proctitis. External validation is needed.

• Retrospective study of 100 cervical cancer patients (70 training, 30 test); logistic regression performed best among four models.
• SHAP analysis: Rectum-V50 and habitat features (Habitat2_Flatness) were top predictors; high-risk group had 100% radiation proctitis in test set.
• Limitations: single-center, small test set, no external validation; model not yet demonstrated to generalize.

RadPigeon summaries are original and for information only. They are not clinical advice.