248 - Uncertainty Quantification-Incorporated Ensemble Model for Personalized Prediction of Severe Radiotherapy-Induced Lymphopenia among Esophagus Cancer Patients

01:20pm - 01:30pm ET

Room 254

Presenter(s)

Baode Gao, MS - University of Texas Health Science Center at Houston, Houston, TX

B. Gao^1,2, Y. Chu³, Z. Hu¹, M. Grayson¹, P. S. N. van Rossum⁴, C. Grassberger⁵, H. Enderling⁶, S. H. Lin⁶, B. P. Hobbs⁷, R. Mohan¹, and Y. Chen⁸; ¹Department of Radiation Physics, The University of Texas MD Anderson Cancer Center, Houston, TX, ²Department of Biostatistics and Data Science, University of Texas Health Science Center, Houston, TX, ³School of Biomedical Informatics, University of Texas Health Science Center, Houston, TX, ⁴Department of Radiation Oncology, Amsterdam UMC, Amsterdam, Netherlands, ⁵Department of Radiation Oncology, University of Washington/Fred Hutchinson Cancer Center, Seattle, WA, ⁶Department of Radiation Oncology, The University of Texas MD Anderson Cancer Center, Houston, TX, ⁷Telperian, Austin, TX, ⁸Department of Epidemiology and Biostatistics, Texas A&M University, College Station, TX

Purpose/Objective(s):

Radiation-induced lymphopenia (RIL) is associated with poor survival outcomes in esophageal cancer. While machine learning models have been developed to predict RIL, current approaches lack patient-specific uncertainty quantification, which is essential for determining confidence in clinical decision-making. To address this critical need, this study aims to develop and validate an uncertainty-incorporated framework to predict Absolute Lymphocyte Count (ALC) nadir and the probability of Grade 4 RIL,

Materials/Methods:

We analyzed 1395 EC patients treated with photons (n=956) or protons (n=439) therapy. The outcomes were absolute lymphocyte count (ALC) nadir and Grade 4 RIL (nadir <200 cells/). We utilized AutoGluon to create an ensemble model integrating clinical variables and a “Composite Dosimetric Score (CDS),” which condenses dose-volume indices via non-negative matrix factorization to reduce multicollinearity. We implemented three uncertainty quantification methods: standard Conformal Prediction (CP), Residual Conformal Prediction (RCP), and Cross-Residual Conformal Prediction (CRCP). Model performance was evaluated using Mean Absolute Error (MAE)/Root Mean Absolute Error (RMSE) for ALC nadir prediction and Area Under the Curve Receiver Operating Characteristic (AUC-ROC)/F1-score for G4RIL classification, while uncertainty quantification was assessed via coverage rates and interval lengths.

Results:

Proton therapy patients exhibited significantly higher average ALC nadir (0.33 ) compared to photon patients (0.25 ; p<0.001), despite similar baseline ALC levels. The ensemble model achieved robust performance with a test set MAE of 0.0926 (RMSE 0.1328 ) for ALC nadir and an AUC of 0.78 (F1-score 0.639) for G4RIL classification. SHAP analysis identified baseline ALC, PTV, and Body CDS as the most influential features. For uncertainty quantification, the CRCP method demonstrated the highest efficiency, providing the tightest prediction intervals while maintaining valid coverage across all confidence levels.

Conclusion:

This study successfully integrates rigorous uncertainty quantification into a high-performance ensemble model for RIL prediction. By providing reliable prediction intervals alongside point estimates, this approach enhances confidence in clinical decision-making, facilitating the future development of Digital Twins and personalized adaptive interventions to mitigate severe immunosuppression.