The article discusses a method to reduce diagnostic errors by using Latent Dirichlet Allocation (LLMs) to identify pieces of evidence in Electronic Health Records (EHR) that show increased or decreased risk of specific diagnoses. It introduces a Neural Additive Model for making predictions backed by evidence at the uncertain times for clinicians, aiming to mitigate delays in diagnosis and errors from incomplete differentials. The authors conduct an evaluation by simulating how it might be used by a clinician to decide between a pre-defined list of differential diagnoses.
Publication date: 15 Feb 2024
Project Page: https://github.com/dmcinerney/ehr-diagnosis-env, https://github.com/dmcinerney/ehr-diagnosis-agent, https://github.com/dmcinerney/ehr-diagnosis-env-interface
Paper: https://arxiv.org/pdf/2402.10109
