HUMBLE

Hamilton Uveal Melanoma Brachytherapy Long-term Estimator

An AI-Driven Prognostic Tool from the Hamilton Eye Institute


About the Tool

HUMBLE (Hamilton Uveal Melanoma Brachytherapy Long-term Estimator) is an AI-driven platform developed to generate individualized survival predictions for patients with uveal melanoma treated by low-dose-rate Iodine-125 episcleral plaque brachytherapy (EPBT). By integrating a suite of advanced analytic methods—spanning Random Survival Forests to deep neural network models—HUMBLE offers a multi-faceted picture of long-term outcomes, empowering ophthalmologists and researchers to refine patient counseling and clinical decision-making.

HUMBLE was trained on a retrospective cohort of 1,142 patients, all of whom underwent LDR I-125 plaque brachytherapy from 1984 to 2022. The dataset encompasses a broad spectrum of clinical and tumor variables, including patient age, best-corrected visual acuity (BCVA), tumor geometry (apex height, basal diameters, and ellipsoid-derived volume), tumor proximity to the optic nerve head, incidence of vitreous hemorrhage or bullous retinal detachment, and AJCC T staging. By modeling these features in tandem, HUMBLE helps clinicians recognize patient-specific risk trajectories to optimize long-term management strategies.


Methods

  1. Data Collection & Cleaning: From an initial dataset of 1,807 patients, we retained 1,142 with standardized, complete records relevant to primary uveal melanoma treated with LDR I-125 plaque brachytherapy. Data integrity checks ensured consistency in key clinical variables.
  2. Feature Engineering: We primarily focused on nine core features reported in the literature to influence uveal melanoma outcomes:
    • Age at Surgery
    • BCVA (LogMAR)
    • Distance to Optic Nerve Head (≤1.5 mm)
    • Vitreous Hemorrhage (VH)
    • Bullous Retinal Detachment
    • Computed Tumor Volume (approximate ellipsoid)
    • AJCC Cancer Stage (T category)
    • Apex Height (z_apex_mm)
    • Maximal Basal Diameter (max_basal_mm)
    Additional fields, such as x_diam_mm and y_diam_mm, supported automatic calculations of tumor volume and maximal basal dimension.
  3. Training & Validation (5-Fold Cross-Validation): The dataset was divided into five folds, each representing ~20% of the patients. In each iteration, one fold was used for testing while the remaining four (~80%) were used for training and internal hyperparameter tuning (e.g., number of trees, minimum leaf size, etc.). We averaged performance metrics across the five folds to estimate generalization.
  4. Model Construction: Random Survival Forest (RSF): We selected the best hyperparameters following cross-validation, then retrained on the full dataset. The final RSF configuration was:
    • n_estimators = 2000
    • min_samples_split = 5
    • min_samples_leaf = 15
    • random_state = 42
    • oob_score = True
    We trained:
    • Overall Survival RSF: all-cause mortality.
    • Recurrence-Free Survival RSF: local or distant tumor recurrence.
    • Enucleation-Free Survival RSF: enucleation events.
  5. Evaluation Metrics: We assessed predictive accuracy using:
    • Harrell’s Concordance Index (C-index)
    • Integrated Brier Score (IBS)
    • Time-Dependent AUC (cumulative dynamic AUC)
    Calibration curves (bin-based and isotonic regression) were employed to check reliability of predicted survival probabilities.
  6. Model Explainability (SHAP): We adopted SHapley Additive exPlanations (SHAP) to illuminate how each clinical feature influences RSF predictions. SHAP decomposes the prediction for an individual patient, highlighting which factors most raise or lower their estimated risk relative to a baseline population.
  7. Competing Risks (DeepHit): Beyond RSF, we introduced a DeepHit neural network to model multiple event types in a single unified framework—specifically, metastasis-associated mortality and non-cancer (other cause) mortality. DeepHit directly learns survival time distributions under competing risks, enabling more granular estimates of cause-specific incidence. These probabilities are visualized in stacked bar charts, complementing the RSF-based analyses.

Feature Importance

Feature Importance in Uveal Melanoma

Feature Importance in Predicting Survival Outcomes
Relative contribution of each variable in the RSF, measured via permutation importance.


Model Performance Metrics

Overall Survival RSF (Test Set):
C-index = 0.8339; IBS = 0.0971

Recurrence-Free Survival RSF (Test Set):
C-index = 0.701; IBS = 0.044

Enucleation-Free Survival RSF (Test Set):
C-index = 0.768; IBS = 0.062

DeepHit (Competing Mortality Risk):
Metastasis-Associated Mortality C-index ≈ 0.82; Other-Cause Mortality C-index ≈ 0.80

Cumulative Dynamic AUC (1–10 years)

Cumulative Dynamic AUC at Yearly Intervals

Demonstrates strong performance with time-dependent AUC values of ~0.947 at 1 year, ~0.894 at 5 years, and ~0.869 at 10 years.

Averaged Time-Dependent ROC Curve

Averaged Time-Dependent ROC Curve

Summarizes the model’s overall discriminative capacity, with a mean AUC approaching 0.88 over multiple time horizons.


Limitations & Disclaimer

While HUMBLE offers quantitative estimates, it is not intended to replace comprehensive clinical judgment. The training data reflect particular patient populations, protocols, and time periods, and certain risk variables (e.g., genetics, comorbidities) may not be fully captured. Accordingly, real-world outcomes may differ. We recommend interpreting predictions as probabilistic guidance, integrated into standard clinical decision-making.

By using HUMBLE, you acknowledge that ultimate treatment decisions rest on physician expertise and the evolving landscape of ocular oncology. HUMBLE is provided for research and educational purposes and is not an FDA-approved diagnostic device.


References

AI-Driven Survival Prediction in Uveal Melanoma Patients Treated with Low-Dose-Rate Iodine-125 Brachytherapy

Authors: L.A. Cernichiaro-Espinosa, D.J. Taylor Gonzalez, B.A. King, S. Choi, A. Nabavi, M. Delsoz, L. Pfeffer, S. Yousefi, M.W. Wilson, et al.

(Abstract accepted to ARVO 2025)

Efficacy of Low-Dose-Rate Iodine-125 Plaque Brachytherapy in the Treatment of Uveal Melanoma

Authors: L.A. Cernichiaro-Espinosa, S.L. Choi, D.J. Taylor Gonzalez, T. Hayes, J. Mastellone, M. Stinson, L.M. Pfeffer, L.H. Rinker, H.Y. Choi, B.A. King, M.W. Wilson

Submitted to Ophthalmology

A forthcoming manuscript will detail further expansions, including multi-institutional validations.