Reasoning Language Model for Personalized Lung Cancer Screening

2025-09-07

AI Summary

In lung cancer screening, Lung-RADS relies solely on nodule imaging features, limiting its ability to simultaneously achieve high sensitivity and specificity. To address this, we propose the first reasoning-oriented language model that jointly integrates chest CT scans and longitudinal electronic health records (EHRs). Our method employs structural equation modeling and chain-of-thought prompting to enable traceable, interpretable, individualized risk assessment. It explicitly decomposes and quantifies contributions from heterogeneous risk factors—including imaging characteristics, clinical variables, and temporal dynamics—thereby overcoming the limitations of conventional single-modality stratification. Leveraging a synergistic pipeline of dataset curation, knowledge distillation, supervised fine-tuning, and reinforcement learning, our model achieves significant improvements on the NLST dataset: +8.2% in AUC, +12.5% in sensitivity, and +9.3% in specificity. This advances precision lung cancer screening with clinically trustworthy, transparent decision support.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Language Grounding & Multi-modal NLP
• Computer Vision: Multi-modal Vision

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Economics, Online Markets and Human Computation: LLM based quality controls for crowd work

Abstract

Accurate risk assessment in lung cancer screening is critical for enabling early cancer detection and minimizing unnecessary invasive procedures. The Lung CT Screening Reporting and Data System (Lung-RADS) has been widely used as the standard framework for patient management and follow-up. Nevertheless, Lung-RADS faces trade-offs between sensitivity and specificity, as it stratifies risk solely based on lung nodule characteristics without incorporating various risk factors. Here we propose a reasoning language model (RLM) to integrate radiology findings with longitudinal medical records for individualized lung cancer risk assessment. Through a systematic study including dataset construction and distillation, supervised fine-tuning, reinforcement learning, and comprehensive evaluation, our model makes significant improvements in risk prediction performance on datasets in the national lung screening trial. Notably, RLM can decompose the risk evaluation task into sub-components, analyze the contributions of diverse risk factors, and synthesize them into a final risk score computed using our data-driven system equation. Our approach improves both predictive accuracy and monitorability through the chain of thought reasoning process, thereby facilitating clinical translation into lung cancer screening.

Problem

• Personalized lung cancer risk assessment using radiology and medical records
• Improving predictive accuracy and monitorability in screening evaluations
• Integrating diverse risk factors beyond nodule characteristics alone

Innovation

• RLM integrates radiology findings with longitudinal medical records
• Uses supervised fine-tuning and reinforcement learning for prediction
• Decomposes risk evaluation into sub-components with data-driven equation