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

Figures

• 

  Figure 1. Study Flowchart This graph shows the distribution and flow of NEDA studies/reports and individual cases used to develop and test the AI-LVD (with key steps described in more detail in the Supplemental Material). Initially a random 30% split of the data set enabled training of an imputation model that accounted for missing measurement data. The remaining 70% of the data set underwent deployment of a natural language processing (NLP) AI system, followed by removal of echo studies with no valid comments on the echo report relating to the presence or absence of mitral regurgitation, and removal of studies where an individual with multiple studies was represented more than once in different data sets. The AI-LVD development process involved a random 75:25 split of the remaining database, with 75% to train the AI-LVD (after removal of repeat echo studies and prior mitral valve intervention). The final 25% was used to test the performance of the AI-LVD model after selecting the last echo for each patient, the absence of prior mitral valve intervention, age >18 years, and echo between January 1, 2000, and the census date. AI-LVD = artificial intelligence-based detection of left…

• 

  Central Illustration. Measurement-Based Echocardiography AI for Identifying Prognostically Significant Left Ventricular Dysfunction Echocardiographic measurement data (no image analysis) indicated in the left panel is taken into the AI model (middle panel), producing a probability output. The right panel demonstrates the unadjusted pattern of actual 1- and 5-year mortality for men (top graph) and women (bottom graph) according to the sex-specific, decile distribution of AI output. LVEF = left ventricular ejection fraction; MVR = mitral valve replacement; other abbreviation as in Figure 1.

• 

  Figure 2. Adjusted Risk of Long-Term, All-Cause Mortality According to the AI-LVD Probability Output (Men) This graph plots the Cox proportional hazards for long-term all-cause mortality adjusted for age, year of (last) echo (reference group for comparison – 2000-04), the number of key echo variables reported/available to the AI-LVD (0-8), and the decile distribution of AI-LVD output in men (reference group for comparison – lowest decile group being an output probability of <0.421). Abbreviations as in Figure 1.

• 

  Figure 3. Adjusted Risk of Long-Term, All-Cause Mortality According to the AI-LVD Probability Output (Women) This graph plots the Cox proportional hazards for long-term all-cause mortality adjusted for age, year of (last) echo (reference group for comparison – 2000-04), the number of key echo variables reported/available to the AI-LVD (0-8), and the decile distribution of AI-LVD output in women (reference group for comparison – lowest decile group being an output probability output of <0.422). Abbreviations as in Figure 1.