Core Concepts
The authors propose a novel Large Language Multimodal Models (LLMMs) framework to predict chronic disease risk by integrating clinical notes and laboratory test results. By leveraging large language models, they achieve significant improvements in early-stage diabetes prediction accuracy.
Abstract
In the study, the authors address the limitations of previous research by collecting five years of electronic health records (EHRs) from a Taiwan hospital database. They focus on training large language models to predict chronic diseases, particularly diabetes. The proposed LLMMs framework combines text embedding encoders and multi-head attention layers to learn laboratory test values and merge blood features with chronic disease semantics. By utilizing models like clinicalBERT and PubMed-BERT with attention fusion, they achieve high accuracy in multiclass chronic disease and diabetes prediction. Transforming laboratory test values into textual descriptions using the Flan T-5 model further enhances prediction accuracy. The study demonstrates the effectiveness of leveraging numerical text data for training and inference in language models, significantly improving early-stage diabetes prediction.
Stats
Accuracy of 73% achieved in multiclass chronic diseases and diabetes prediction.
76% Area Under the ROC Curve (AUROC) achieved by transforming laboratory test values into textual descriptions.
Quotes
"Chronic diseases such as diabetes are the leading causes of morbidity and mortality worldwide."
"Our method combined a text embedding encoder and multi-head attention layer to learn laboratory test values."