Dosimetric Predictors in Overall Survival Prediction for Patients with Mesothelioma through an Interpretable Cox Regression Model

Abstract

<h3>Purpose/Objective(s)</h3> Malignant Pleural Mesothelioma (MPM) is a type of rare but fatal cancer occurring in the pleural of the lung, with about 3000 new cases diagnosed in the US every year. The five-year survival rate of MPM is around 10%. Individualized radiation treatment (RT) planning has always been challenging due to the complex geometry of MPM. Our goal is to assess the clinical and dosimetric predictors that are correlated with overall survival (OS) for MPM patients following radiotherapy, via explainable algorithms. <h3>Materials/Methods</h3> A cohort of 59 MPM patients treated between 2013-2018 were included in the study. All patients underwent adjuvant radiotherapy at a helical IMRT unit with a prescription of 45 Gy (25 fx). Patient characteristics and the planning dosimetry were collected for each MPM patient. Cox Proportional Hazard Regression Model (Cox) and Survival Support Vector Machine (sSVM) were used for building models based on patient characteristics, dosimetric variables and the combined variables. C-index for each model was calculated after 10-fold cross validation for evaluating the performance. Shapley Additive exPlanations (SHAP), an algorithm mathematically based on game theory for interpreting the model, was utilized to select the top predictors and unveil their impacts on OS predictions from the best-performing model. <h3>Results</h3> No statistically significant differences were found in overall survival according to patient characteristics: disease laterality, age, histology, except gender. Female patients lived significantly longer than male patients (p-value=0.02). No significant survival differences were found between patients with/without post-RT pneumonitis. 23.3% of MPM patients were diagnosed with post-RT pneumonitis, which is significantly lower than 84.6% (a previous study of 13 MPM patients by Allen AM, et al 2006). The interpretable Cox models, based on dosimetric predictors, yielded the highest c-index of 0.76 to predict OS for patients with MPM. Top dosimetric predictors, chosen by SHAP from Cox models, are the ratio of PTV/Ipsilateral - PTV volume (negative impact), ipsilateral lung – PTV V20 (positive impact), and total lung – PTV mean dose (negative impact) in predicting OS. <h3>Conclusion</h3> Patient-specific dosimetry could predict overall survival of MPM patients. The explainable Cox and ML models provide additional prognostic power of the prediction models, which could possibly be applied in clinics to advance individualized radiation treatment planning for MPM patients.