A Multi-Centre Multi-Omics Study of Critical Weight Loss Prediction in Nasopharyngeal Carcinoma Patients Undergoing Chemo-Radiotherapy
J SUN1*, S Lam2, X Teng3, J Zhang4, Z Ma5, C Liu6, W Li7, H Xiao8, Y Huang9, X Han10, F Lee11, W Yip12, A Cheung13, H Lee14, K Au15, J Cai16, (1) HKU/POLYU, Tko, ,HK, (2) Duke Kunshan University, Kunshan, ,CN, (3) The Hong Kong Polytechnic University, Hong Kong, Hong Kong, HK, (4) Duke Kunshan University, Nantong, 32, CN, (5) The Hong Kong Polytechnic University, ,,(6) The Hong Kong Polytechnic University, Hong Kong, Hong Kong, (7) The Hong Kong Polytechnic University, ,,(8) The Hong Kong Polytechnic University, Hong Kong, 91, CN, (9) The Hong Kong Polytechnic University, Hung Hom, Kowloon, (10) ,,,(11) Queen Elizabeth Hospital, Hong Kong, ,HK, (12) Queen Elizabeth Hospital, Hong Kong, ,CN, (13) ,,,(14) The University Of Hong Kong, ,,(15) Hong Kong Queen Elizabeth Hospital, ,,(16) Hong Kong Polytechnic University, Hong Kong, ,CN
Presentations
PO-GePV-T-108 (Sunday, 7/10/2022) [Eastern Time (GMT-4)]
ePoster Forums
Purpose: To develop a clinically generalizable multi-omics model for predicting critical weight loss (CWL) in nasopharyngeal carcinoma (NPC) patients during concurrent chemoradiotherapy (CRT) via a multi-centre study.
Methods: NPC patients from two hospitals were retrospectively collected and assigned as primary (N=64) and validation (N=59) cohort, respectively. The threshold of CWL was set to be percentage weight loss > 10% during CRT. Shape, energy, and texture feature of radiomics (Rx) and dosiomics (Dx) were extracted from computed tomographic images and dose maps, respectively, from four organ structures including primary gross-tumor-volume (GTVnp), nodal gross-tumor-volume (GTVn), bilateral parotid glands and larynx. A univariate filter (p-value < 0.1) was used, followed by forward ridge recursive feature elimination classifier to select predictive features in the primary cohort for Support Vector Machine modelling. Pre-treatment body mass index (BMI) was set as the baseline model (BMI-only), three omics-models (BMI+Rx, BMI+Rx+Dx, Rx+Dx) were then developed. Model performance was evaluated in validation cohort with Receiver-Operation-Characteristics (ROC) curve and Decision Curve Analysis (DCA).
Results: Four Rx features and eight Dx features were identified among the studied models. Model with BMI+Rx+Dx obtained the highest training (primary) score, achieving an area under ROC curve (AUC) = 0.91, outperforming other comparing models. This model also yielded the best validation score in validation cohort with AUC = 0.79. Compared to the BMI-only model, addition of either Rx or Dx features improved the model’s predictive power (validation AUC: BMI-only= 0.42; BMI+Rx=0.61; BMI+Rx+Dx=0.79). Furthermore, the BMI+Rx+Dx model also demonstrated the highest net benefit compared to “treat-all/treat-none” patients in DCA when the threshold probability was larger than 0.35.
Conclusion: The multi-omics model (BMI+Rx+Dx) outperformed other comparing models, demonstrating its strong capability and net benefit in differentiating the clinical heterogeneity of patients’ CWL during RT. Further, external validation of the model highlighted its generalizability toward potential clinical applications.
Funding Support, Disclosures, and Conflict of Interest: Innovation and Technology Fund (ITS/080/19), the Innovation and Technology Commission Project of Strategic Importance (P0035421) Shenzhen-Hong Kong-Macau S&T Program (Category C)(SGDX20201103095002019) Shenzhen Basic Research Program (R2021A067)
Keywords
Texture Analysis, Radiation Dosimetry, Radiation Therapy
Taxonomy
Contact Email