Researchers integrate radiology text and EHR for renal malignancy prediction

A JMIR preprint by Fan, Liang, Sun, Pan, Terry, and Xu presents a retrospective cohort study that builds a multimodal prediction pipeline to estimate renal tumor malignancy from combined radiology report text and structured EHR data. The authors report using large language models to extract abnormality characteristics and a pretrained biomedical transformer, RadBERT, to generate contextual text embeddings. The preprint evaluates three fusion strategies, and reports that early fusion achieved an area under the ROC curve of 0.818 (± 0.010), with textual features from RadBERT driving the largest improvement and LLM-extracted structured characteristics yielding modest additional gains, per the preprint.

Per the preprint, the study fuses features from structured EHR variables with text-derived features using early, middle, and late fusion approaches. Performance was measured using standard classification metrics including:

• •accuracy
• •precision
• •recall
• •specificity
• •AUC
• •F1-score

The manuscript reports that RadBERT contextual embeddings outperformed simpler extracted features, while the LLM-based abnormality extraction contributed incremental benefit when combined with embeddings, according to the authors.

Editorial analysis: Multimodal approaches that combine contextual transformer embeddings with tabular EHR features are increasingly common in clinical ML research because they can capture complementary information from narrative reports and structured records. Observers in the field note that text embeddings from domain-pretrained transformers often yield larger gains than handcrafted or rule-extracted features, while extraction pipelines based on LLMs can add value but introduce additional computational and validation burdens.

Editorial analysis: Key practical considerations for adoption include the cost of running domain LLMs at scale, the need to validate text-extracted labels against chart review, and model interpretability requirements in clinical decision support. The preprint focuses on model performance metrics; it does not, in the available manuscript, provide deployment or prospective validation details.

Editorial analysis: Readers should watch for the peer-reviewed JMIR Med Inform version for expanded methodological details, dataset size and characteristics, external validation results, and any code or model release that would enable reproduction and comparative benchmarking.