Enterprise AI Analysis

Advanced Data Curation & Machine Learning Pipeline

This study leveraged a meticulously curated dataset of 74 patients and 311 fibroids from a cohort of 573 individuals who underwent UFE. For each patient, pre- and post-treatment contrast-enhanced MRI scans were used to manually annotate and extract key fibroid features, including FIGO classification, spatial location, 3D volume, vascularity, and tissue viability. These annotations were rigorously validated by board-certified radiologists.

The core of the predictive system is a Deep Set Network, a deep learning architecture designed to process variable numbers of fibroids per patient in an order-invariant manner. This model combines individual fibroid features with patient-specific clinical data (age, BMI, baseline symptoms) to predict overall clinical success (fibroid shrinkage >50%) and symptom relief.

Predicting Patient-Level Outcomes and Symptom Relief

The Deep Set Network model demonstrated strong performance in predicting both overall clinical success and the likelihood of symptom relief post-UFE. For overall procedure outcome (successful shrinkage), the model achieved an accuracy of 75% (AUC = 0.74). More impressively, the model showed high accuracy in predicting relief for specific symptoms:

• Frequent Urination: 88% Accuracy (F1=88%, AUC ≈ 0.87)
• Pelvic Pain: 82% Accuracy (F1=81%, AUC ≈ 0.81)
• Severe Back Pain: 82% Accuracy (F1=81%, AUC ≈ 0.83)
• Severe Bloating: 81% Accuracy (F1=83%, AUC ≈ 0.82)
• Heavy Bleeding: 81% Accuracy (F1=78%, AUC ≈ 0.78)

These results highlight the model's ability to capture complex patterns influencing treatment response, particularly for symptoms directly linked to fibroid size and mass effect.

Individual Fibroid Response Prediction & Key Factors

Beyond patient-level outcomes, the study also developed models to predict the success of treatment for each individual fibroid. This fibroid-level prediction aimed to identify which specific fibroids responded well to UFE, defined as >50% volume reduction. Ensemble tree methods, such as Random Forests and XGBoost, outperformed linear and instance-based classifiers:

• Random Forests: 76% Accuracy (F1=75%)
• XGBoost: 71% Accuracy (F1=67%)

A crucial aspect of this analysis was the identification of key features driving predictive performance. Fibroid volume and vascularity emerged as the strongest predictors of treatment response. High vascularity facilitates better delivery of embolic agents, leading to more substantial infarction and volume reduction. FIGO subtype and anatomic location also played a role, with hybrid/submucosal classifications tending to favor better outcomes.

Strategic Insights for Future AI in Healthcare

The study's findings underscore the potential of integrating fibroid-level imaging features into predictive models for pre-procedural treatment planning. While predicting overall fibroid shrinkage remains more challenging than symptom relief due to complex biological factors and follow-up variability, the models offer significant clinical utility.

Future directions include validating the model in larger, multi-center cohorts, incorporating additional clinical features (hormonal profiles, menstrual history), predicting longer-term outcomes, and automating feature extraction. Exploring advanced imaging features like dynamic contrast-enhanced MRI and diffusion metrics, alongside more complex models (e.g., convolutional neural nets for 3D fibroid shape), promises to further refine predictions and enable more personalized, patient-centered care.