AI-derived left ventricular global longitudinal strain enables automated assessment of subclinical myocardial dysfunction and provides incremental prognostic information in type 2 diabetes

Subclinical myocardial dysfunction may develop in patients with type 2 diabetes mellitus (T2DM) before overt heart failure or reduction in left ventricular ejection fraction (LVEF). Left ventricular global longitudinal strain (LV GLS) can detect early myocardial deformation abnormalities, but conventional strain analysis remains limited by post-processing burden and operator dependence. We aimed to develop and validate an automated three-view AI-derived LV GLS framework and to evaluate its clinical and prognostic relevance in patients with T2DM. This multi-stage observational study included a broad-spectrum model-development cohort of 2,547 participants, a retrospective clinical cohort of 877 patients with T2DM, and a prospective external validation cohort of 200 patients with T2DM. Patients in both T2DM cohorts had preserved LVEF and no overt heart failure; mean age was 61.8 ± 9.7 and 60.9 ± 8.8 years, respectively, and 58.7% and 59.0% were men. Hypertension was present in 64.3% and 63.5% of the retrospective and prospective cohorts, respectively. An anatomy-guided, label-efficient AI framework derived three-view LV GLS from standard apical four-chamber, two-chamber, and long-axis cine loops and was validated against expert TomTec measurements. An imaging-defined subclinical myocardial dysfunction phenotype was defined by preserved LVEF, absence of overt heart failure or other major causes of LV dysfunction, and impaired TomTec-derived LV GLS ( > − 18%). The primary outcome was 5-year major adverse cardiovascular events (MACE), defined as cardiovascular death, nonfatal myocardial infarction, nonfatal stroke, heart-failure hospitalization, or coronary/peripheral revascularization, in the retrospective cohort. Prognostic models underwent internal validation using 1,000 bootstrap resamples with optimism correction. AI-derived LV GLS showed close agreement with TomTec-derived reference measurements, with intraclass correlation coefficients of 0.986, 0.984, and 0.981 in the development, retrospective, and prospective cohorts, respectively, and mean analysis times below 10 s per case. For the TomTec-defined imaging phenotype, AI-derived LV GLS alone achieved AUCs of 0.986 in both the retrospective and prospective cohorts; addition of clinical or metabolic-inflammatory variables provided minimal further discrimination, consistent with both methods measuring the same underlying deformation phenotype. During 5-year follow-up, 94 of 856 QC-valid patients (11.0%) experienced MACE. Abnormal AI-derived LV GLS ( > − 18%) was independently associated with MACE after multivariable adjustment (HR 2.00, 95% CI 1.27–3.16; P = 0.003). Addition of AI-derived LV GLS to the parsimonious clinical model increased the apparent 5-year AUC from 0.628 to 0.729 (ΔAUC 0.101, 95% CI 0.045–0.159); after bootstrap optimism correction, the corresponding AUCs were 0.586 and 0.699. Direct comparison showed similar prognostic discrimination for AI-derived and TomTec-derived LV GLS (0.729 vs. 0.732; ΔAUC − 0.002, 95% CI − 0.011 to 0.006). AI-derived three-view LV GLS enabled rapid and reproducible assessment of subclinical myocardial dysfunction in patients with T2DM and retained prognostic information comparable to conventional TomTec-derived GLS. Its principal advantage lies in automated and scalable strain quantification rather than superior prognostic discrimination. The prognostic findings require confirmation in an independent prospective outcome cohort.