LLMs integrate clinical knowledge for LNM prediction

Researchers from Zhejiang Lab and Peking University Cancer Hospital published a study in JMIR Medical Informatics introducing a knowledge-augmented framework that combines large language model reasoning with machine learning model outputs to predict lymph node metastasis (LNM) in lung cancer. The work addresses a well-known gap in clinical AI: conventional ML models capture statistical patterns from patient data but do not reason over the medical knowledge a specialist would apply, while LLMs encode broad medical knowledge but have historically underperformed data-driven models on structured clinical tasks.

Lymph node metastasis is a decisive factor in lung cancer staging. Its presence or absence determines surgical eligibility and the need for neoadjuvant therapy. Accurate preoperative LNM prediction is challenging: standard imaging has limits, and misclassification leads to suboptimal treatment decisions. Both radiomics features and deep learning have been applied to this problem, with ML models trained on patient clinical features currently representing the strongest data-driven approaches.

The framework works as an ensemble. First, a traditional ML model predicts LNM probability from structured clinical features. An LLM then receives that probability alongside the patient's full clinical data, including demographics, lab values, CT report text, and disease history. The LLM is prompted to estimate LNM risk independently from the clinical data, then revise its estimate by incorporating the ML model output and its calibrated performance context. Multiple LLM responses are collected for the same patient and aggregated as the final prediction. This chain-of-thought design forces the model to engage its medical knowledge before seeing the ML output, reducing the tendency to blindly defer to the machine learning result.

The study offers a practical template for hybrid clinical prediction: rather than replacing ML with LLMs or using LLMs standalone, it shows LLMs can function as knowledge-grounded interpreters of ML outputs. The framework handles both structured data and unstructured free-text clinical notes. The research group has published prior work in JMIR and IEEE on NLP and ML for LNM prediction in non-small cell lung cancer, making this a methodological extension of an established research program. The ensemble pattern - LLM as a reasoning layer over ML predictions - is generalisable to other clinical risk tasks where medical knowledge should complement pattern-fitting.