262 - Machine Learning with Demographic and Procedural Data Predicts the Need for Additional Sclerotherapy for Lymphatic Malformations Following Initial Treatment with Doxycycline

Purpose: Lymphatic malformations (LMs) are congenital vascular anomalies arising from the abnormal development of lymphatic vessels. While percutaneous sclerotherapy (PS) is the first-line treatment for LMs, predicting the number of treatment sessions required for remission remains challenging. To address this, we developed a machine learning (ML) model that uses readily available clinical data—patient demographics and procedural details from initial treatment—to predict whether patients will require additional treatment sessions following initial doxycycline sclerotherapy.

Materials and Methods: We conducted a retrospective analysis of patients who received doxycycline PS for LMs at our institution between January 2014 and April 2025. We included patients with a confirmed LM diagnosis receiving their first PS at our institution. Patients were excluded if they had not received PS, had prior PS at an outside hospital, or lacked a confirmed LM diagnosis. We collected data on patient demographics (e.g., age, sex) and procedural details from initial treatment (e.g., number of access sites, volume of sclerosant administered). Gradient-boosted ML models were trained to classify patients as requiring single vs. multiple PS sessions, and assessed using five-fold cross-validation.

Results: A LightGBM model achieved a mean accuracy of 0.87, a mean AUC of 0.84, and a mean F1-score of 0.81 across five folds, demonstrating robust performance. The five most influential features were: volume of sclerosant used, age at time of treatment, reference air kerma, number of sites accessed, and total procedure time.

Conclusion: An ML model trained with readily available clinical data, such as patient demographics and procedural details from initial treatment, may accurately predict whether patients with lymphatic malformations will require additional treatment sessions following initial doxycycline sclerotherapy. Importantly, this approach achieved robust performance without relying on time-consuming tasks like imaging preprocessing or lesion segmentation, offering a practical and resource-efficient tool for treatment prognosis.