[2503.05893] Zero-shot Medical Event Prediction Using a Generative Pre-trained Transformer on Electronic Health Records

0 1 minute read

250305893 Zero shot Medical Event Prediction Using a Generative Pre trained Transformer.png

[Submitted on 7 Mar 2025 (v1), last revised 7 Jul 2025 (this version, v2)]

View the PDF file from the paper entitled Zero-Shot using an adapter trained before training on electronic health records, by Ekaterina Rickop and 9 other authors

PDF HTML (experimental) view

a summary:The longitudinal data in electronic health records (EHRS) represents the clinical history of the individual through a series of blog concepts, including diagnosis, procedures, medicines and laboratory tests. The trained before training (GPT) transformers can take advantage of these data to predict future events. While adjusting these models can enhance the performance of the task, it becomes expensive when applied to many clinical prediction tasks. On the contrary, the pre -foundation model can be used to prepare a zero prediction, providing a developmental alternative to separate models to control each result.

This study provides the first comprehensive analysis of the zero prediction with the EHRS foundation, which leads to a new pipeline that formulates the medical concept as the task of obstetric modeling. Unlike the methods of supervision that require intense designer data, our model enables the model to predict the next medical event only of prior knowledge. We evaluate performance through multiple time prospects and clinical categories, which indicates the ability of the model to capture the inherent time and complex patient paths without overseeing the task.

The typical performance of the following medical concept was evaluated using accuracy and recall standards, achieving a better accuracy of 0.614 and calling 0.524. For 12 main diagnostic cases, the model showed a strong zero performance, achieving high real positive rates while maintaining low false positives.

We explain the strength of the EHR GPT foundation in capturing various apparent patterns and enabling the uncommon prediction of clinical results. This ability enhances the diversity of predictive health care models and reduces the need for task training, providing more developmental applications in clinical settings.