Main Session
Sep
28
PQA 03 - Digital Health Innovation and Informatics, Patient Safety & Quality, and Radiation and Cancer Biology
2482 - Machine Learning-Based Identification of the Spleen as an Immune Organ-at-Risk and Model-Based Evaluation of Spleen-Sparing Esophageal Chemoradiotherapy
Presenter(s)
Zongsheng Hu, PhD - University of Pennsylvania, Philadelphia, PA
Z. Hu1, Y. Li1, C. R. Peeler1, B. Gao1, Y. Chen2, U. Titt1, S. H. Lin3, and R. Mohan1; 1Department of Radiation Physics, The University of Texas MD Anderson Cancer Center, Houston, TX, 2Department of Epidemiology and Biostatistics, Texas A&M University, College Station, TX, 3Department of Radiation Oncology, The University of Texas MD Anderson Cancer Center, Houston, TX
Purpose/Objective(s):
Radiation-induced lymphopenia (RIL) is common during esophageal chemoradiotherapy and is associated with adverse outcomes. The spleen is not routinely considered organ-at-risk (OAR) despite its immunological role. We hypothesized that splenic irradiation is a major dosimetric determinant of absolute lymphocyte count (ALC) nadir and that spleen sparing planning could improve lymphocyte preservation.Materials/Methods:
We analyzed 922 esophageal cancer patients treated with chemoradiotherapy using IMRT (n 412), VMAT (n 148), PSPT (n 300), or IMPT (n 62) at one institution. Clinical variables included age, BMI, baseline ALC, and PTV. Dose–volume histograms (DVHs) of the spleen, heart, lung, and body were condensed using non-negative matrix factorization into “composite dosimetric scores” (CDSs). An XGBoost model was trained using 8-fold cross-validation to predict ALC nadir, defined as the lowest ALC from RT start through 1-week post-treatment. Feature importance was assessed using SHapley Additive exPlanations (SHAP) values. Dose response relationships between mean spleen dose (MSD) and predicted ALC were evaluated. To assess feasibility, 20 high–spleen-dose VMAT patients were replanned with spleen-sparing constraints while maintaining target and OAR coverage.Results:
The model demonstrated good performance (mean absolute error = 78 cells/µL). SHAP analysis identified splenic irradiation as the most influential dosimetric predictors. Dose response analysis showed diminishing predicted immune benefit below mean spleen dose (MSD) of 9 Gy and substantially decreased ALC nadir at MSD > 17 Gy. Across the full cohort, 73.5% of patients had MSD over 9 Gy, while 46.2% exceeded 17 Gy, indicating a substantial proportion of patients may benefit from spleen-sparing strategies. In the replanned cohort, MSD decreased from 24.9 Gy to 7.9 Gy. Reduction of MSD to <17 Gy was achievable in 100% of replanned cases, and <9 Gy in 60%, without compromising target coverage or other OARs. Mean dose differences for all other OARs were <0.7 Gy, and spinal cord maximum dose differences were <0.5 Gy. Predicted average ALC nadir increased from 188 ± 15 to 250 ± 17 cells/µL, corresponding to a 36% improvement. Among 9 patients with grade 4 RIL (ALC <200 cells/µL), 4 were predicted to improve to lower-risk status.Conclusion:
Interpretable machine learning supports the spleen as a key immune organ-at-risk in esophageal chemoradiotherapy. Dose response analysis supports the spleen planning strategy of an ideal MSD goal of 9 Gy and a practical action threshold of 17 Gy. As 73.5% of patients had MSD >9 Gy, spleen-sparing objectives may be relevant for a large proportion of patients. A model-based replanning demonstrates clinically feasible spleen-sparing planning may substantially mitigate severe lymphopenia without compromising other OAR or target coverage. These findings support incorporating spleen-sparing objectives into radiotherapy planning and motivate prospective clinical validation.