Development and internal evaluation of a model combining ECG waveforms with machine reports to identify recorded right ventricular systolic dysfunction in adults

Abstract Background ECG waveform models can identify structural and functional cardiac abnormalities. Their incremental value beyond jointly modeled machine measurements and interpretation text requires evaluation using the same patients and reference labels. Methods We linked MIMIC-IV-ECG v1.0 with MIMIC-IV-Echo v1.0.1 from a single hospital and selected one index ECG–echocardiography pair per eligible adult. The primary training cohort comprised 25,315 patients, including 3,165 endpoint-positive patients. The primary endpoint was recorded global or unspecified reduction in right ventricular systolic function. A selected machine-feature model (B; XGBoost) was compared with the same configuration refitted with an additional convolutional-network waveform score (BW). Development, calibration and threshold selection were separated from internal testing. The primary comparison was test specificity at model-specific thresholds selected for a 90% validation sensitivity target. Actual test sensitivities were reported separately. Nominal 95% confidence intervals were obtained from 2,000 paired patient-bootstrap resamples of the frozen systems. Results The primary test cohort included 5,505 patients, of whom 703 were endpoint-positive. The area under the receiver operating characteristic curve (AUROC) increased from 0.782 (95% CI 0.764–0.798) to 0.850 (95% CI 0.835–0.865), a difference of 0.069 (95% CI 0.057–0.080). Specificity increased from 40.19% to 49.77%, a difference of 9.58 percentage points (95% CI 8.40–10.79), corresponding to 460 fewer false positives (95% CI 403–518). Actual sensitivity was 92.18% versus 91.89%, a difference of − 0.28 percentage points (95% CI − 2.29–1.63). Sensitivity was lower in the operationally complete-measurement subset, and a secondary severity-label scenario showed a sensitivity disadvantage. Conclusions Adding a waveform score improved discrimination and reduced false positives within this frozen internal evaluation. The findings do not establish sensitivity noninferiority, clinical utility or external applicability. Registration: OSF grs6q; stage-specific registration.