Towards Intelligent Precision Oncology: A Hybrid AI-Driven Framework for Cancer Risk Stratification and Early Detection

Abstract

Early cancer detection remains one of the greatest challenges in precision oncology, often hindered by limited diagnostic specificity, rigid screening protocols, and poor adaptability across diverse patient populations. To address these gaps, we propose HyAICare, a hybrid, AI-driven framework designed to integrate multimodal biomedical data with a modular, adaptive learning pipeline. The system combines circulating cell-free DNA (cfDNA) fragmentomics, attention-enhanced cytological image analysis, and patient-specific clinical metadata to enable truly personalized, context-aware screening recommendations. HyAICare consists of three synergistic modules: (1) a Biomarker-Informed Deep Learning (BIDL) model for cfDNA-based cancer risk scoring, (2) a Tissue-Specific Risk Mapper (TSRM) for tumor origin localization using deep attention networks, and (3) an Adaptive Screening Recommender (ASR) employing LightGBM with SHAP-based explainability for individualized screening advice. Tested on a diverse, multi-ethnic dataset of 10,000 patients from three tertiary cancer centers, HyAICare achieved a 94.6% cancer detection accuracy and a 91.8% tissue localization rate, while reducing unnecessary screenings by 24.5%—demonstrating clear advantages over existing unimodal AI systems. Notably, the framework preserves interpretability through transparent feature attribution and visual attention maps, ensuring it remains clinically auditable and ethically scalable. These results position HyAICare as a practical, next-generation tool for early cancer detection in real-world, heterogeneous healthcare environments.