Abstract 1096: Estimating risk factors and predicting various cancer types using integrated real-world data: An AI-based approach

Abstract Cancer remains a critical global health issue with numerous risk factors. However, comprehensive large-scale studies in the United States, particularly those examining cancer-related medical conditions, are limited. This study aims to develop AI-driven precision medicine algorithms by integrating data from electronic health records (EHR) to identify patients at elevated risk for various cancers, using age-, gender-, race-, and area deprivation index (ADI)-stratified data. By employing disease, demographic, and ADI data as covariates, we apply interpretable machine learning models to uncover insights into cancer risk factors. We explored eight machine and deep learning algorithms including logistic regression (LR) and random forest (RF) using 1, 229 truncated unique ICD-10-CM codes (selected through related risk analysis) as binary features across a total pool of 1.6 million patients with 5-fold cross-validation. First, we applied both LR and RF methods for all patients. Using LR, we observed that conditions such as nicotine dependence, irregular menstruation, and urinary disorders were highly associated with subsequent cancer diagnoses, while acute pharyngitis and cardiac murmurs were associated with a decreased risk, achieving an overall accuracy of 86%. RF further identified conditions, including acute pharyngitis, nicotine dependence, and other soft tissue disorders with high predictive value, achieving 96% training accuracy and 86% test accuracy. Finally, eight machine and deep learning algorithms were evaluated on various sample sizes, ranging from 200 to 20, 000, assessing accuracy, precision, recall, F1 score, ROC AUC, and MCC. Notably, RF accuracy increased from 69% to 78%, and ROC AUC from 79% to 86% as sample size grew, underscoring the model's robustness with larger datasets. Our findings contribute to early cancer detection and provide insights into disease progression. This comprehensive analysis has the potential to improve early detection, diagnostic pathways, and understanding of the pathological processes underlying cancer worldwide, as well as highlight variable performance across methodologies on this kind of data. Citation Format: Md Ashad Alam, Nick Duesbery, Daniel Fort. Estimating risk factors and predicting various cancer types using integrated real-world data: An AI-based approach [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 1096.