SCH: Methods for Clinical Assessment Generation

Electronic health records (EHRs) are rich resources for computer-assisted clinical decision support systems (CDSS). There are over five decades of research in CDSS, resulting in methods in knowledge-based and machine-learning approaches. However, methods have been evaluated by different datasets or in different clinical applications, it is difficult to compare the performance of different methods.

This study will evaluate state-of-the-art methods for CDSS, focusing on the important application of clinical assessment generation, which predicts medical diagnoses and summaries of the key elements that lead to the medical diagnoses. The CDSS methods will be implemented and evaluated using the longitudinal EHR data from the US Veterans Health Administration (VHA) across the entire United States of America.

Clinical assessment generation will help medical education by generating clinically relevant case studies. Clinical assessment generation will improve patient care by assisting primary care physicians with clinical diagnoses, especially for patients with rare diseases, and by identifying medical specialties and subspecialties for their patients. This work will be transformative in CDSS.

Providers are trained to write notes with a problem-oriented subjective, objective, assessment, and plan (SOAP) structure, where assessments can be inferred from their subjective and objective sections of the SOAP notes. Clinical assessment generation can be considered as an application of text-to-text or graph-to-text generation. State of the art methods includes text-to-text transfer transformer, generative pre-trained transformer, generation with multihop reasoning flow, commonsense knowledge aware conversational model, abductive commonsense reasoning, and automated clinical assessment generation, all of which will be evaluated for clinical assessment generation using over 8 million VHA patients’ longitudinal EHRs from over 1,200 healthcare facilities in the US.

The evaluation metrics include bilingual evaluation understudy and Recall­Oriented Understudy for Gisting Evaluation. Clinicians will evaluate the clinical relevance and interpretability of each method. We will evaluate method performance by disease categories or subcategories, by different note types, and by different VHA hospital facilities. We will also evaluate method performance by race, gender, age, and minority groups.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.