802P A machine learning-based framework for efficacy and safety analysis of PD-1PD-L1 inhibitor application in gynaecological cancer on incomplete post-marketing surveillance dataset

Abstract

Post-marketing surveillance of antineoplastic agents is performed to evaluate the efficacy and safety in patients aiming at expanding drug indications and discovering potential adverse reactions. The real-world data is fraught with multiple missing values. Literature addressing different strategies for dealing with missing data in such a situation is scarce. Using machine learning algorithms for predicting therapeutic response to PD-1/PD-L1 Inhibitor has attracted attention in recent years. However, training a predictive model usually requires imaging or biomarker information, which is rarely available in the post-marketing surveillance data. To address these challenges, we propose an ensemble learning-based framework that utilizes a multi-model fusion strategy at the stages of both data imputation and outcome prediction. The proposed method is evaluated on a gynecological cancer post-marketing surveillance dataset with 118 patient samples, collected from Sep 2019 to Feb 2021. Compared to several existing imputation algorithms, our method demonstrates a superior capability to obtain high-quality imputed values, effectively boosting the overall performance of outcome prediction. The study also compares three PD-1 inhibitors, including Camrelizumab (with 51 patient samples), Sintilimab (44 samples), and Toripalimab (23 samples). Statistical results show that Toripalimab presents a significant efficacy improvement in the RECIST metric, with 48% CR, 35% PR, 4% SD, and 13% PD, compared to Camrelizumab (29% CR, 22% PR, 14% SD, 24% PD) and Sintilimab (18% CR, 48% PR, 20% SD, 14% PD). The death rates for Camrelizumab, Sintilimab, and Toripalimab, are 17.6%, 13.6%, and 0%. With respect to adverse reactions (AE), the rates of organ dysfunction are 62.7%, 50.0%, and 43.5%, and the rates of general discomfort are 17.6%, 47.7%, and 17.4%, for Camrelizumab, Sintilimab, and Toripalimab, respectively. Our method can be used as an effective analytical tool for efficacy and safety evaluation on an incomplete post-marketing surveillance dataset.