Establishment and validation of a dynamic prognostic model using serial circulating tumor DNA (ctDNA) for endemic EBV-related nasopharyngeal carcinoma (NPC): A secondary analysis of EP-SEASON.

6049 Background: Published risk prediction tools have focused on pretreatment factors, whereas the accuracy remains challenging in cancer care. Emerging evidences emphasize the dynamic rather than static recurrence risks during treatment course, and non-invasive diagnostics tools have advanced opportunities for serial tumor assessments. Here, we present an effective dynamic risk individualized prediction model (NPC-DRIM) incorporating serial ctDNA data, using the endemic EBV-related NPC as a model. Methods: This study included 1000 patients (pts) enrolled from a prospective biomarker study EP-SEASON, with complete longitudinal ctDNA data at 11 timepoints across treatment: after each neoadjuvant chemotherapy (NAC) circle (T1-3), every week during radiotherapy (T4-T9), within 1 week after radiotherapy (T10), and 1-3 months after radiotherapy (T11). Pts were divided into subcohort NAC (n=752) and subcohort no-NAC (n=248) according to receiving NAC or not, and randomly 70/30% split into training and validation cohort. Time-series and statistical features characterizing the dynamic change of ctDNA at each timepoint were extracted. The NPC-DRIM at T3-T11 were developed using the features selected via Cox univariate analysis in training cohort and then validated. The performance of NPC-DRIM was determined by C-index, time-dependent AUC, calibration curves, and decision curves, and compared with existing models. Results: The NPC-DRIM incorporated 4 clinical variables, 8 time-series features and 10 statistical features of ctDNA data. The C-index for predicting recurrence increased with time: 0.64 at T2, 0.69 at T3-T4, 0.70 at T5, 0.71 at T6, 0.73 at T7-T9, 0.77 at T10, and 0.76 at T11 in subcohort NAC ; 0.70 at T5, 0.68 at T6, 0.82 at T7, 0.78 at T8, 0.73 at T9, 0.74 at T10, and 0.83 at T11 in subcohort no-NAC . The NPC-DRIM at T11 had statistically improved outcome prediction compared to other dynamic models (Landmark Cox and Joint Model), and static models (AHR_Chen, RPA_Guo, RPA_Lee, and AJCC_8th staging system) (Table). For individualized dynamic risk prediction, we developed a web-based calculator to visualized the estimated changing recurrence risks. In addition, we showed that the high-risk pts identified by NPC-DRIM benefit from immune checkpoint inhibitors (ICI), while the low-risk pts did not. Conclusions: We introduce for the first time that the dynamic risk prediction model NPC-DRIM outperformed the conventional models, facilitating personalized therapeutic paradigms. Clinical trial information: NCT03855020 . Subcohort NAC Subcohort no-NAC C-index p value C-index p value NPC-DRIM 0.76 0.83 Landmark Cox 0.65 0.01 0.63 <0.01 Joint Model 0.63 <0.01 0.61 0.01 AHR Model (Chen et al. 2021) 0.61 <0.01 0.57 <0.01 RPA Model (Guo et al. 2019) 0.59 <0.01 0.59 <0.01 RPA Model (Lee et al. 2019) 0.59 <0.01 0.66 0.02 AJCC_8th 0.56 <0.01 0.57 <0.01