Purpose: To develop and validate treatment-specific prediction models for critical weight loss (CWL) in nasopharyngeal carcinoma (NPC) patients during radiotherapy (RT) course via a multi-centre setting.
Methods: A total of 303 and 146 NPC patients, retrospectively recruited from two medical centres, were allocated to a training group (n=303) and testing group (n=146), respectively, forming a Combined Cohort (CC) regardless of treatment regimens. From the CC, three treatment-specific sub-cohorts were formed according to the prescribed treatment regimen, including Concurrent Chemoradiotherapy (CRT), CRT plus Adjuvant Chemotherapy (ACR), and Induction Chemotherapy plus CRT (ICR), with a ratio of training to testing samples being 189:27, 64:59 and 51:60, respectively. Radiomics and dosiomics features were extracted from four organ structures using computed tomography images (GTVnp, GTVn, Parotids, and Larynx). Model development was performed in the training samples of each cohort separately. A filter-based recursive feature elimination method was employed for feature selection and feature dimensionality reduction. The remaining features were then passed into a Support Vector Classifier to predict CWL of >10% during the RT course. External model validation was performed in testing samples of respective cohorts. The Area under Receiver-Operating-Characteristics Curve (AUC) and Brier score (BS) were used to evaluate model discriminability and calibration, respectively.
Results: All treatment-specific models demonstrated an improved discriminability in both training and external testing samples compared to the CC model (Training AUC: CC=0.75, CRT=0.95; ACR=0.91, ICR=0.81; Testing AUC: CC=0.57, CRT=0.79, ACR=0.79, ICR=0.78), and also displayed better probabilistic prediction accuracy (calibration) (Testing BS: CC=0.24, CRT=0.18, ACR=0.17, ICR=0.18). Besides, a smaller difference between training and testing AUC was observed in treatment-specific models, suggesting a mitigated issue of overfitting.
Conclusion: Three treatment-specific models were developed and demonstrated satisfactory and superior model discriminability and calibration performance than the CC model in external testing samples, meanwhile showing less degree of model overfitting.
Funding Support, Disclosures, and Conflict of Interest: Innovation and Technology Fund (ITS/080/19), the Innovation and Technology Commission Project of Strategic Importance (P0035421) The Hong Kong Polytechnic University, and Shenzhen-Hong Kong-Macau S&T Program(Category C)(SGDX20201103095002019) Shenzhen Basic Research Program (R2021A067)