A language model trained on clinical notes can accurately estimate clinical and operational tasks within a health system, according to a study published online June 7 in Nature.
Lavender Yao Jiang, from NYU Langone Health in New York City, and colleagues trained a large language model for medical language in the NYU Langone Health System (NYUTron). The language model-based approach included four steps: data collection using unlabeled and labeled clinical notes; pretraining; fine-tuning across a range of clinical and operational predictive tasks; and deployment. The approach was examined for prediction of five tasks within a health system: 30-day all-cause readmission, in-hospital mortality, comorbidity index, length of stay, and insurance denial.
The researchers found that on prediction tasks, NYUTron had an area under the curve of 78.7 to 94.9 percent. Compared with traditional models, this represented an improvement of 5.36 to 14.7 percent.
【MORE】