Abstract 1677P
Background
Although ICIs revolutionized modern oncology, a significant challenge remains to minimize immune related adverse events (irAEs), and subsequent financial burden due to treatment delay and hospitalization. There is a paucity of data about predicting biomarkers for ICI-related toxicity. In this study we aimed to use machine learning (ML) algorithms to study the profile of ICI toxicity.
Methods
A computer algorithm was developed using SAP Business Objects BI Platform 4.2 to extract electronic medical records (EMR) of patients receiving ICIs and identify irAEs. To avoid potential confounding effects, the study included only patient receiving ICIs without concurrent chemotherapy, biological or hormonal therapy or radiotherapy during the study period. The algorithm used all available data on each patient; Namely, structured data (e.g., numeric lab results) and unstructured data (e.g., patient notes written by the treating physician). The code mapped irAEs drawn from the structured database, as well as irAEs drawn from medical notes using ML models in Natural Language Processing (NLP) with MDCLONE Studio V1.1.2 tool. Association between variables and irAEs were assessed using chi-squared and t tests.
Results
The study cohort included 1617 patients with solid tumors who received ICIs between the years 2010-2021 in the Division of Oncology in Rambam Health Care Campus. 1104 (68%) were men, mean age was 69 years. 910 (56%) patients developed grade 3&4 irAEs during and after the ICIs treatment. Gender, BMI and pretreatment derived neutrophil-to-lymphocytes ratio (dNLR) were not associated with higher irAEs in the general cohort. Younger age and PDL-1/CTLA-4 combination were found to be associated with higher irARs rates (P=0.001 and P<.001, respectively). In subgroup analysis, young age was found to be associated with hepatotoxicity and hematologic irAEs (P<.001 and P=0.01, respectively), female gender was associated with endocrine toxicity (P=0.024) and high BMI and low dNLR were associated with renal irAEs (P=0.001 and P=0.029, respectively).
Conclusions
Using ML tools in real-world data setting, several key characteristics were identified to be correlated with high tendency of irAEs.
Clinical trial identification
Editorial acknowledgement
Legal entity responsible for the study
The authors.
Funding
Has not received any funding.
Disclosure
All authors have declared no conflicts of interest.