Purpose: Malignant pleural mesothelioma (MPM) is a rare but aggressive cancer arising from the cells of the thoracic pleura with poor prognosis. We aimed to develop a model, via an interpretable machine learning (ML) method, predicting overall survival for MPM rafter radiotherapy based on dosimetric metrics, as well as patient characteristics.
Methods: Fifty-nine MPM (36: right, 23: left) patients treated on a Tomotherapy unit between 2013-2018 were included. All patients underwent 45 Gy (25 fractions). The patient overall survival time, as well as patient characteristics, and achievable dosimetric parameters for target and critical OARs were collected. Survival SVM (sSVM) was applied to build predictive models of overall survival (OS) based on patient characteristics, achieved planning dosimetric variables, and combined parameters. SHAP, a unified interpretable ML algorithm based on game theory, was further employed to select most associated variables, as well as the contribution of each predictor in predicting OS.
Results: The OS was only found significantly different (p=0.02) between female and male MPM patients. The sSVM model based on dosimetric/combined predictors achieved the highest c-index=0.76, while the patient characteristic sSVM model achieved c-index=0.66. Among top three patient characteristics: N staging, gender, and T staging were mostly correlated with OS. The top dosimetric variables (mostly correlated with OS prediction) characterized by the SHAP values were total lung-PTV volume, ipsilateral lung-PTV volume and contra-lateral lung volume. Higher values of total lung-PTV volume may likely result in a smaller prediction of OS, while higher values of ipsilateral-PTV volume and contralateral lung volumes may likely result in predicting longer OS.
Conclusion: Patient-specific dosimetric variables may predict overall survival of mesothelioma patients (c-index=0.76), which could guide personalized treatment planning towards better treatment response. The identified predictors and their impact on survival, revealed by explainable ML model, offered additional values for translational application in clinical practice.