A novel deep learning model using dosimetric and clinical information for grade 4 radiotherapy-induced lymphopenia prediction

Radiotherapy-induced lymphopenia has increasingly been shown to reduce cancer survivorship. We developed a novel hybrid deep learning model to efficiently integrate an entire set of dosimetric parameters of a radiation treatment plan with a patient's pre- and mid-treatment information to improve the prediction of grade 4 radiotherapy-induced lymphopenia.

We proposed a two-input channel hybrid deep learning model to process dosimetric information using a stacked bi-directional long-short term memory structure and non-dosimetric information using a multilayer perceptron structure independently before integrating the dosimetric and non-dosimetric information for final prediction. The model was trained from 505 patients and tested in 216 patients. We compared our model with other popular predictive models, including logistic regression (with and without elastic-net regularization) random forest, support vector machines, and artificial neural network.

Our hybrid deep learning model out-performed other predictive models in various evaluation metrics. It achieved the highest area under the curve at 0.831, accuracy at 0.769, F1 score at 0.631, precision at 0.670, and recall at 0.610. The hybrid deep learning model also demonstrated robustness in exploiting the value of dosimetric parameters in predictive modeling.

We demonstrated that our hybrid deep learning model with a two-input channel structure, which addressed the sequential and inter-correlated nature of dosimetric parameters, could potentially improve the prediction of radiotherapy-induced lymphopenia. Our proposed deep learning framework is flexible and transferable to other related radiotherapy-induced toxicities. (paper)