A Large Language Model Screening Tool to Target Patients for Best Practice Alerts: Development and Validation

Best Practice Alerts (BPAs) are alert messages to physicians in the electronic health record that are used to encourage appropriate use of health care resources. While these alerts are helpful in both improving care and reducing costs, BPAs are often broadly applied nonselectively across entire patient populations. The development of large language models (LLMs) provides an opportunity to selectively identify patients for BPAs.

In this paper, we present an example case where an LLM screening tool is used to select patients appropriate for a BPA encouraging the prescription of deep vein thrombosis (DVT) anticoagulation prophylaxis. The artificial intelligence (AI) screening tool was developed to identify patients experiencing acute bleeding and exclude them from receiving a DVT prophylaxis BPA.

Our AI screening tool used a BioMed-RoBERTa (Robustly Optimized Bidirectional Encoder Representations from Transformers Pretraining Approach; AllenAI) model to perform classification of physician notes, identifying patients without active bleeding and thus appropriate for a thromboembolism prophylaxis BPA. The BioMed-RoBERTa model was fine-tuned using 500 history and physical notes of patients from the MIMIC-III (Medical Information Mart for Intensive Care) database who were not prescribed anticoagulation. A development set of 300 MIMIC patient notes was used to determine the model's hyperparameters, and a separate test set of 300 patient notes was used to evaluate the screening tool.

Our MIMIC-III test set population of 300 patients included 72 patients with bleeding (ie, were not appropriate for a DVT prophylaxis BPA) and 228 without bleeding who were appropriate for a DVT prophylaxis BPA. The AI screening tool achieved impressive accuracy with a precision-recall area under the curve of 0.82 (95% CI 0.75-0.89) and a receiver operator curve area under the curve of 0.89 (95% CI 0.84-0.94). The screening tool reduced the number of patients who would trigger an alert by 20% (240 instead of 300 alerts) and increased alert applicability by 14.8% (218 [90.8%] positive alerts from 240 total alerts instead of 228 [76%] positive alerts from 300 total alerts), compared to nonselectively sending alerts for all patients.

These results show a proof of concept on how language models can be used as a screening tool for BPAs. We provide an example AI screening tool that uses a HIPAA (Health Insurance Portability and Accountability Act)-compliant BioMed-RoBERTa model deployed with minimal computing power. Larger models (eg, Generative Pre-trained Transformers-3, Generative Pre-trained Transformers-4, and Pathways Language Model) will exhibit superior performance but require data use agreements to be HIPAA compliant. We anticipate LLMs to revolutionize quality improvement in hospital medicine.