A deep learning model trained on 325,377 echocardiography videos from 62,818 patients at Cedars-Sinai can detect chronic kidney disease directly from routine cardiac ultrasound — without a blood or urine test. Trained on parasternal long-axis clips, the model achieved AUC 0.756 on internal validation and held consistent performance (AUC ~0.718–0.719) across two independent external cohorts totaling over 43,000 patients at Stanford and Kaiser Permanente Northern California.
The clinical logic is elegant: CKD and cardiovascular disease share overlapping structural signatures — left ventricular hypertrophy, diastolic dysfunction, pericardial changes — that a neural network can detect even when cardiologists aren't specifically looking for renal disease. With 60% of CKD cases globally undiagnosed, opportunistic screening during routine echocardiography could intercept disease years before kidney function becomes critically compromised.
The AUC values (~0.72–0.76) represent modest but clinically meaningful discrimination — not a replacement for serum creatinine or eGFR, but a potentially impactful triage layer for populations with limited nephrology access. The large, multi-site validation is a genuine strength. Key limitations include reliance on retrospective data, unknown generalizability to non-US healthcare settings, and the absence of a prospective clinical utility study showing improved patient outcomes. As a preprint not yet peer-reviewed, these findings require independent scrutiny before clinical translation. Still, the paradigm of using cardiac imaging as a metabolic disease detector is increasingly credible and practically compelling.
