Purpose: Numerous studies have shown that severe radiation-induced lymphopenia (RIL - depletion of absolute lymphocyte count (ALC)) adversely impacts patient survival after radiation therapy (RT) of most solid tumors. In this study, we aimed to identify pre-RT patient characteristics and dosimetric features that could predict radiation-induced lymphocyte depletion in esophageal cancer patients during and after the course of RT to identify patients who may benefit from treatment plan reoptimization, use of a different modality, or a pharmacological intervention and ultimately improve outcomes.
Methods: Data of 860 esophageal cancer patients who received concurrent chemoradiotherapy were used for this study. All patients were treated with protons or photons radiation modalities with a dose of 50.4 Gy. Pre-, weekly during RT and post-RT ALCs, clinical characteristics and dosimetric parameters were extracted from data bases. A hybrid deep learning model using deep neural network and long short-term memory (LSTM) network was applied to predict ALC depletion trend using baseline value of ALCs measured before or during early stages of RT course.
Results: Several important prediction metrics were employed to evaluate the performance of the proposed model compared to other commonly used prediction methods. The results showed that the proposed model achieved predictions with the mean squared error (MSE) of 0.046 for 258 patients in test set. Inclusion of ALC after the first week reduced the MSE to 0.014. Moreover, this model outperformed off-the-shelve prediction methods by at least 30% reduction in MSE of weekly ALC predictions.
Conclusion: The proposed model performed well in predicting radiation-induced lymphocyte depletion. The ALC prediction would enable physician to identify patients at high risk of severe radiation-induced lymphopenia and potentially benefit from modified treatment approaches. Furthermore, if measured ALC were to be available after the first few fractions, it could be used for improved confidence in risk prediction.
Funding Support, Disclosures, and Conflict of Interest: The research was supported by NIH U19 CA021239.
Not Applicable / None Entered.