1876P A combination model of electronic patient-reported outcomes (ePROs) and lab measurements in prediction of immune related adverse events (irAEs) and treatment response of immune checkpoint inhibitor (ICI) therapies

Abstract

ICIs have introduced irAEs, arising from various organs without temporal connection to the therapy. Furthermore, treatment benefit evaluation with ICIs is often difficult. Data collected from various sources could be used to build machine learning (ML) based tools to improve the early detection of irAEs and treatment benefit which could improve patient care, quality-of-life, and cost-effectiveness of ICIs. This study combined datasets from four separate sources. The first dataset consisted of 16 540 reported symptoms collected with Kaiku Health digital platform from 33 ICI treated cancer patients. Other dataset included laboratory measurements, prospectively collected irAEs (n=26), including dates and CTCAE based severity, and investigator assessed treatment responses (n=33). Datasets were combined from 4 to 12 weeks of the treatment initiation to analyze the relationships between the variables. The symptoms reported less than two weeks before lab measurements were linked together. When analyzing the days before the onset of irAEs, all data from two weeks before the diagnosis were considered. The final dataset used in the analysis consisted of 227 samples from 33 patients including 16 monitored symptoms and 11 lab values. Increases in ALT, ALP, Creatinine and especially thyrotropin correlated strongly with clinical benefit (CB). From the monitored symptoms itching, nausea and dizziness had positive correlation with CB. ALP, leukocytes and neutrophils had positive and nausea negative correlation with irAEs. Rash and increases in thyrotropin correlated notably with the upcoming onset of irAEs. The results of the study show that irAEs and treatment benefit both correlate with ePRO collected symptom data and laboratory values. It is feasible that both sources of data could be used in building ML based prediction models for irAEs and treatment benefit.