Evaluation of Diagnostic Apps and Prediction Models for Myocardial Infarction and Other Causes of Chest Pain: Informing Patient Use

Abstract

Abstract Background: Symptom checker (SC) applications output possible diagnoses based on user’s symptoms. They may influence patients’ care seeking behavior but remain understudied, especially for high-risk diseases including acute myocardial infarction (AMI). Objective: This study used risk factor and symptom data reported by patients presenting with chest pain to an ED to evaluate the accuracy of Ada, WebMD, and Isabel SCs in diagnosing high-risk, cardiac, and low risk, noncardiac causes of chest pain. We hypothesized (1) SCs would miss cases of AMI, (2) SCs would over-diagnose AMI in noncardiac, low risk cases. Methods: From a dataset of 1872 cases of patients with chest pain, fifty high-risk cases (S1) were randomly sampled. 29 cases (S2) were selected as low risk, noncardiac, and included additional noncardiac symptoms and diagnoses. Samples were entered into the SCs, and matches were identified with top 5 app suggestions (M1-M5). SC performance was compared with a logistic regression (LR) model previously trained on the original dataset to predict AMI. Results: WebMD: (S1) Acute coronary syndrome (UA and AMI)- 100% sensitive, 13.3% specific, PPV-43.5%, NPV-100%. Identified 100% of AMIs, 100% of UAs. (S2) Identified 24.1% of S2 low risk, noncardiac diagnoses. Suggested AMI first for 34.5% of cases and only nonurgent diagnoses (true negatives) for 3.4% of cases. Isabel: (S1) ACS - 75% sensitive, 83.3% specific, PPV-75%, NPV-83.3%. Identified 100% AMIs, 44.4% UAs. (S2) Identified 24.1% of S2 noncardiac diagnoses, suggested AMI first for 17.2%, true negatives 0%. Ada: (S1) ACS - 95% sensitive, 56.7% specific PPV-59.4%, NPV-94.4%. Identified 100% of AMIs, 88.9% of UAs. (S2) Identified 48.3% of S2 noncardiac diagnoses, suggested AMI first for 34.5%, true negatives 17.2%. LR model: (S1) ACS – 100% sensitive. Suggested ACS for 59% S2 cases. True negative rate (41%) was significantly higher than WMD (3.4%) or Isabel (0%), (P =.001). Conclusions: All 3 SC apps identified 100% of AMIs in their top 5 suggestions and were highly sensitive to ACS. However, SCs were risk averse and limited in their identification of noncardiac diagnoses in low-risk patients. The LR model had significantly better discrimination with low-risk patients and potential to decrease excess care.