Noninvasive assessment of left ventricular end-diastolic pressure using machine learning–derived phasic left atrial strain

While transthoracic echocardiography (TTE) assessment of left ventricular end-diastolic pressure (LVEDP) is critically important, the current paradigm is subject to error and indeterminate classification. Recently, peak left atrial strain (LAS) was found to be associated with LVEDP. We aimed to test the hypothesis that integration of the entire LAS time curve into a single parameter could improve the accuracy of peak LAS in the noninvasive assessment of LVEDP with TTE.We retrospectively identified 294 patients who underwent left heart catheterization and TTE within 24 h. LAS curves were trained using machine learning (100 patients) to detect LVEDP ≥ 15 mmHg, yielding the novel parameter LAS index (LASi). The accuracy of LASi was subsequently validated (194 patients), side by side with peak LAS and ASE/EACVI guidelines, against invasive filling pressures. Within the validation cohort, invasive LVEDP was elevated in 116 (59.8%) patients. The overall accuracy of LASi, peak LAS, and American Society of Echocardiography/European Association for Cardiovascular Imaging (ASE/EACVI) algorithm was 79, 75, and 76%, respectively (excluding 37 patients with indeterminate diastolic function by ASE/EACVI guidelines). When the number of LASi indeterminates (defined by near-zero LASi values) was matched to the ASE/EACVI guidelines (n = 37), the accuracy of LASi improved to 87%. Importantly, among the 37 patients with ASE/EACVI-indeterminate diastolic function, LASi had an accuracy of 81%, compared with 76% for peak LAS.LASi allows the detection of elevated LVEDP using invasive measurements as a reference, at least as accurately as peak LAS and current diastolic function guideline algorithm, with the advantage of no indeterminate classifications in patients with measurable LAS.