Artificial Intelligence for Detection of Prognostically Significant Left Ventricular Dysfunction From Echocardiography

Background

Identification of left ventricular (LV) dysfunction following echocardiographic investigations remains problematic, particularly when the ejection fraction (EF) is preserved.

Objectives

The authors examined the operational characteristics of artificial intelligence LV dysfunction (AI-LVD) identification from routinely obtained echocardiographic measurements.

Methods

Following initial training in 126,136 (imputation cohort) and 254,735 (training cohort) cases from the National Echo Database of Australia, the AI-LVD was tested in 81,509 cases (last echo January 1, 2000-May 21, 2019) with no mitral valve intervention or pacemaker. This cohort comprised 41,796 men (51.3%) aged 62.3 ± 17.1 years and 39,713 women aged 63.2 ± 18.4 years, in whom 4,490 (5.5%), 3,734 (4.6%), and 59,297 (72.7%) had reduced, mildly reduced, and preserved EF, while 13,988 (17.2%) had no recorded EF and 39,940 (45.2%) had “indeterminate” filling pressures.

Results

Overall, the AI-LVD generated a (sex-specific) output in decile distributions consistent with increasingly higher levels of LV dysfunction and mortality—actual 5-year mortality rising from 5.7% to 66.3% and 2.3% to 64.2% in men and women, respectively. The prognostic capacity of the AI-LVD persisted in preserved EF, when adjusting for age, year of echo, and missing echo parameters—with adjusted hazard for all-cause mortality during 1,541 (812-2,682) days follow-up 4.93-fold (95% CI: 4.35-5.59) and 7.11-fold (95% CI: 5.85-8.64) higher in the highest vs lowest decile group in men and women, respectively.

Conclusions

A new AI-LVD algorithm using only echocardiographic measurements can reliably identify prognostically important LV dysfunction, including in preserved EF, even when key reporting parameters are missing. The AI-LVD can be used in real-time during routine echocardiography reporting.