Multimodal AI estimates hemodynamic echocardiographic indicators of structural heart disease from heart sounds and pulse waves

Abstract Echocardiography is the reference method for cardiac hemodynamics but requires dedicated hardware and an expert operator, whereas the stethoscope reports only whether disease is likely. This study asked whether a stethoscope examination can instead yield triage-level estimates of the quantities that echocardiography measures. A multimodal (heart sound, pulse wave and demographics) model estimated continuous echocardiographic indicators from one 15-second recording of chest phonocardiography and finger photoplethysmography with a single device. In 1,502 recordings from 1,339 consecutive patients with same-day echocardiography, out-of-fold Pearson correlation averaged 0.402 across eight pre-registered indicators and 0.409 across a post hoc protocol-acquired panel; in 118 held-out recordings, 0.516 and 0.453. Predictions were compressed toward the cohort mean, so the output is a triage-level estimate: a pre-specified rule-out rule reached a sensitivity of 0.95 at a specificity of 0.27. Flow-generated quantities were recovered whereas wall thickness was not. All data come from one center; transportability is untested.