Click here to

Session: Machine Learning for Toxicity Prediction [Return to Session]

Novel Neural Network Approach to Predict Radiation-Induced Lymphopenia in Lung Cancer Patients After Radiotherapy

Y Kim1,2,3*, I Chamseddine2, R Mohan4, T Xu4, S Lin4, Z Liao4, H Paganetti2, A Saraf2, D McClatchy2, J Kim3, Y Cho3, H Yoon3, S Cho1, C Grassberger2, (1) KAIST, Daejeon, South Korea (2) Harvard Medical School/Massachusetts General Hospital, Boston, MA (3) Yonsei Cancer Center, Seoul, South Korea (4) MD Anderson Cancer Center, Houston, TX


MO-C930-IePD-F5-4 (Monday, 7/11/2022) 9:30 AM - 10:00 AM [Eastern Time (GMT-4)]

Exhibit Hall | Forum 5

Purpose: Radiation-induced lymphopenia (RIL) is a complex phenomenon involving multiple critical structures and strongly dependent on patient-specific baseline features. This study aims to develop a novel model structure to predict RIL after Radiotherapy (RT) for stage III non-small cell lung cancer and compare the predictive power to Logistic Regression (LR).

Methods: In total, 162 patients from three institutions were collected. Differential Dose Volume Histograms (DVHs) of heart and normal lungs, mean doses, organ volumes, and baseline lymphocyte count before RT are used as inputs to predict grade (G)4 lymphopenia at the end of RT. The proposed model is a small, robust neural network (NN), in which DVHs are used as input for 1-D convolutional layers, whereas other patient characteristics are considered in a second path via fully-connected layers. The intermediate output features of the two paths are combined and processed using a second fully-connected layer to obtain the final prediction. The model is validated internally using 5-fold cross-validation for robustness and the performance is compared to LR using the area under the receiver operating characteristic curve (AUC).

Results: The average AUCs from LR and our suggested NN are 0.67 and 0.72 respectively. Model robustness quantified via the standard deviation of the prediction is 0.12 and 0.08 respectively. Detailed analysis of the data revealed substantial heterogeneity in the patient cohort, particularly intra-institutional differences in RIL frequency.

Conclusion: Despite interactions across multiple organs and heterogeneity in response among patients, we were able to benchmark a novel RIL model using multiple organ DVHs and blood-based lymphocyte counts from a multi-institutional cohort. Our suggested model structure has the potential to robustly abstract differential DVH data for multiple critical structures to predict toxicity, though stronger predictive factors are needed to improve overall performance of RIL prediction.


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email