Go to OpenReview Public Article ORCID homepageEffects of Follow-Up Interval and Care Setting on Valve Severity Changes Between Successive Echocardiograms: Secondary Analysis of MIMIC-III EchoNotes Cohort (Preprint)
Abstract: Background: Echocardiography reports in routine care often encode clinical reasoning as ordinal valve-lesion severity. The EchoNotes database derived from MIMIC III harmonizes such report text into a reproducible ordinal schema across echocardiograms, enabling population-scale analysis of report-level change between successive exams. However, most prior work has focused on image-based automation (e.g., ejection fraction estimation) or label extraction from text, with far less attention to how these ordinal severities change over time. Critically, apparent change is confounded by follow-up interval (\mathrm{\Delta t}) and care setting (inpatient vs outpatient), yet the literature offers few \mathrm{\Delta t}-aware summaries or nonparametric, \mathrm{\Delta t}-standardized setting comparisons. Objective: The study conducted a secondary data analysis of the de‑identified EchoNotes dataset to provide an evidence‑based description of how ordinal valve-lesion severity levels change between successive examinations and to determine how apparent change depends on \mathrm{\Delta t} and care setting using observed data with uncertainty quantification. Methods: We analyzed the EchoNotes dataset that has 45,794 reports. For each patient and lesion of interest (aortic valve (AV) stenosis, AV regurgitation, mitral valve (MV) stenosis, and MV regurgitation), we chronologically ordered reports, formed successive within‑patient pairs, and retained evaluable ordinal states (0 = normal, 1 = mild, 2 = moderate, 3 = severe). We computed row‑normalized next‑visit transition matrices summarizing, for each baseline state, the distribution of the subsequent state. To make visit timing explicit, we stratified transitions by \mathrm{\Delta t} (< 7, 7-30, 30-90, and ≥ 90 days). Care‑setting differences were estimated by assigning pairs to inpatient or outpatient status. Because settings differ in follow‑up timing, we produced a nonparametric \mathrm{\Delta t}‑standardized contrast. Within each baseline state, inpatient and outpatient pairs were reweighted to a shared \mathrm{\Delta t} bin mix before computing difference matrices (inpatient minus outpatient). Uncertainty was quantified with a subject‑level bootstrap. Results: Left-sided lesion codes were frequently evaluable (AV regurgitation—72.7%, AV stenosis—65.2%, MV regurgitation—67.8%, MV stenosis—37.0%). The ordinal association analysis witnessed the strongest correlation between right-ventricle dysfunction and chamber enlargement (ρ ≈ 0.62). In \mathrm{\Delta t}-stratified next-visit transition matrices, diagonal persistence dominated short intervals, with off-diagonal movement rising at longer \mathrm{\Delta t} (most for MV regurgitation and least for AV stenosis). Care-setting comparisons showed more movement among inpatients in the non-\mathrm{\Delta t}-standardized matrices (e.g., 35% inpatient vs 53% outpatient for AV stenosis when concerning the moderate-to-moderate transitions), but \mathrm{\Delta t}-standardization attenuated most differences, with residual contrasts largely confined to higher-severity AV regurgitation where inpatients were more likely to improve from worse states. Conclusions: We propose an evidence based workflow that yields interpretable maps of how report level echo severities behave in routine care. Our \mathrm{\Delta t} aware results separate visit scheduling from biology, support fairer comparisons across services, and provide practical priors for simulation, cohort design, and evaluation of future multimodal or AI systems against an observational baseline.
External IDs:doi:10.2196/preprints.83320
Loading