Narrative Feature or Structured Feature? A Study of Large Language Models to Identify Cancer Patients at Risk of Heart Failure

A large language model outperformed traditional machine learning methods in identifying cancer patients at risk of developing heart failure, according to a study using electronic health records from the University of Florida Health system [1]. The research, led by Ziyi Chen, examined a cohort of 12,806 patients diagnosed with lung, breast, and colorectal cancers. Within that group, 1,602 individuals later developed heart failure, a known cardiotoxicity risk of cancer treatments that can worsen outcomes and survivorship [1][2]. The study tested several machine learning approaches, including traditional support vector machines, a deep learning model called Time-Aware long short-term memory (T-LSTM), and the widely used transformer model BERT [1][2]. The large language model GatorTron-3.9B achieved the highest F1 scores, outperforming support vector machines by 39%, T-LSTM by 7%, and BERT by 5.6% [1][2]. The researchers attributed the improvement to novel narrative features derived from structured medical codes, which increased feature density [1][2]. The study used electronic health records to build its predictive models, a practice that relies on high-quality training datasets, which are often difficult and expensive to produce because of the time required to label data [8]. The University of Florida, where the patient data originated, is a public land-grant research university in Gainesville and a senior member of the State University System of Florida, classified as an R1 doctoral university with very high research activity [10]. The findings were posted on the preprint server arXiv, with the initial submission on March 18, 2024, and subsequent revisions through June 2026 [1].