A Novel Bayesian General Medical Diagnostic Assistant Achieves Superior Accuracy With Sparse History: A Performance Comparison of 7 Online Diagnostic Aids and Physicians

Alicia M Jones; Daniel R Jones

doi:10.3389/frai.2022.727486

A Novel Bayesian General Medical Diagnostic Assistant Achieves Superior Accuracy With Sparse History: A Performance Comparison of 7 Online Diagnostic Aids and Physicians

Front Artif Intell. 2022 Jul 22:5:727486. doi: 10.3389/frai.2022.727486. eCollection 2022.

Authors

Alicia M Jones¹, Daniel R Jones¹

Affiliation

¹ Eureka Clinical Computing, Eureka Springs, AR, United States.

Abstract

Online AI symptom checkers and diagnostic assistants (DAs) have tremendous potential to reduce misdiagnosis and cost, while increasing the quality, convenience, and availability of healthcare, but only if they can perform with high accuracy. We introduce a novel Bayesian DA designed to improve diagnostic accuracy by addressing key weaknesses of Bayesian Network implementations for clinical diagnosis. We compare the performance of our prototype DA (MidasMed) to that of physicians and six other publicly accessible DAs (Ada, Babylon, Buoy, Isabel, Symptomate, and WebMD) using a set of 30 publicly available case vignettes, and using only sparse history (no exam findings or tests). Our results demonstrate superior performance of the MidasMed DA, with the correct diagnosis being the top ranked disorder in 93% of cases, and in the top 3 in 96% of cases.

Keywords: AI medical diagnosis; Bayesian medical diagnosis; Bayesian network; comparison of physicians with AI decision support; diagnostic decision support system; diagnostic performance; general medical diagnostic assistant; symptom checkers.