A Novel Model for Generating Patient Laboratory Test Orders from Admission: Transformer Model Approach

There is a growing demand for medical pseudo-data that maintains statistical utility, enabling the analysis of a wide range of medical data without compromising patient privacy. Additionally, there is a growing need for effective sequence prediction in healthcare. The proposed model in this research, utilizes GPT-2, a Transformer model from natural language processing, to generate laboratory test order data for patients from admission to discharge while incorporating temporal processes. This research utilized patient data from Kochi Medical School Hospital, Japan, with specific inclusion criteria. The methodology involved converting patient profiles and order data into categorical formats and employing SentencePiece tokenization to create 8000 tokens for training the GPT-2 model. The model's performance was evaluated across three patterns, predicting laboratory test orders at various times during hospitalization. It was compared to a gradient boosting decision tree (GBDT) model using F-scores as the evaluation metric. Results indicated that while the GBDT model performed better based solely on admission information, the GPT-2 model showed superior performance when using time-series information post-admission. Our research proposed a model that successfully considered temporal processes to generate laboratory test order data for patients from admission to discharge using GPT-2, a Transformer model developed in the field of natural language processing.