Baseline Signals in Advanced Heart Failure
Across two distinct advanced heart-failure settings, baseline measurements captured outcome-related differences more consistently than later change measures, but retrospective designs and limited validation prevent clinical conclusions.
Advanced heart failure can involve electrical delay, weak contraction, altered chamber geometry, and blood stasis. Two recent studies examined whether measurements already available near the start of care could distinguish later outcomes in different high-risk groups. One analyzed electrocardiograms around cardiac resynchronization therapy (CRT), a device-based treatment for selected people with systolic heart failure and delayed ventricular activation. The other combined conventional statistics and machine learning in people with ischemic heart failure and a confirmed left-ventricular thrombus. Both found signals in baseline measurements, but neither study established a prospectively validated prediction tool.
Evidence
The CRT study retrospectively analyzed 415 recipients at a Swedish tertiary center. Median age was 72.8 years, 77.3% were men, median baseline left-ventricular ejection fraction was 27.5%, and median follow-up was 2.8 years. Investigators measured left-ventricular activation time (LVAT) on a standard 12-lead electrocardiogram as the interval from QRS onset to the largest deflection in lead V6. The primary endpoint combined heart-failure hospitalization and all-cause mortality; 171 participants reached it.
Median LVAT was 78 milliseconds before implantation and 88 milliseconds afterward. Among the 389 participants with left bundle-branch block or nonspecific intraventricular conduction delay, each 10-millisecond increase in baseline LVAT was associated with a lower unadjusted rate of the composite endpoint (hazard ratio 0.92, 95% confidence interval 0.86–0.99). After adjustment for clinical factors, the estimate was weaker and crossed the null (hazard ratio 0.93, 95% confidence interval 0.85–1.01; p=0.086). The association appeared in a Kaplan–Meier analysis of the left bundle-branch block subgroup, but not in the smaller non-left-bundle groups. Post-implant LVAT was not associated with the endpoint. Importantly, the authors reported that LVAT did not outperform the established QRS-duration measure.
The thrombus study retrospectively included 190 patients treated at one center from 2015 through 2024. All had ischemic heart failure, imaging-confirmed left-ventricular thrombus, anticoagulation documented in the study, and serial echocardiography. By six months, 86 patients had documented thrombus regression and 104 did not. At baseline, the non-regression group had a larger median thrombus area (2.97 versus 2.16 square centimeters), a higher mean CHA₂DS₂-VA vascular-risk score (3.49 versus 2.80), and a lower mean ejection fraction (32.32% versus 35.03%).
In a two-variable model, each one-point increase in CHA₂DS₂-VA was associated with higher odds of six-month non-regression (adjusted odds ratio 1.35, 95% confidence interval 1.09–1.69); thrombus size had a positive but statistically uncertain association. The model's apparent area under the receiver-operating-characteristic curve was 0.644. An internally evaluated CatBoost model produced a cross-validation AUC of 0.762 plus or minus 0.055. Its leading features included left-atrial diameter, pulmonary-artery pressure, platelet count, and left-ventricular end-diastolic diameter. These feature rankings indicate predictive contribution inside this dataset, not causation.
Thrombus regression status alone was not significantly associated with one-year major adverse cardiovascular events. In a separate model, however, baseline thrombus size and CHA₂DS₂-VA were associated with those events, with adjusted odds ratios of 1.14 and 1.57 per unit increase, respectively; model AUC was 0.713.
Analysis — Baseline phenotype versus measured change
The cross-study connection is an analysis, not an established clinical rule. In both cohorts, the more informative signals were present at baseline: electrical activation before CRT in one study, and thrombus burden plus vascular, chamber-remodeling, and hematologic features in the other. Later measurements were less decisive—post-implant LVAT did not track the CRT composite endpoint, while six-month thrombus regression alone did not separate one-year event rates. This convergence suggests a research hypothesis: an initial cardiopulmonary phenotype may integrate accumulated disease substrate that a single follow-up change measure only partly captures. The studies also show why that hypothesis needs restraint. Baseline LVAT lost conventional statistical significance after adjustment, and the thrombus models were trained and evaluated within one small cohort. The two measurements describe different mechanisms and cannot be merged into one score from these data. Their common value is methodological: future studies could prespecify baseline and longitudinal measures, then test whether either adds reproducible information beyond established predictors in external populations.
Limitations
Both studies were retrospective and observational, so their associations cannot show that a measurement caused an outcome or that acting on it would improve outcomes. The CRT cohort came from one experienced tertiary center, included relatively few women, and had very small right-bundle and other morphology subgroups. Selection, referral, attrition, unmeasured confounding, pacing-related measurement variability, and unavailable lead-position details could affect its estimates. Its adjusted LVAT result was statistically uncertain, and LVAT did not outperform QRS duration.
The thrombus cohort was also single-center and included only 190 patients with ischemic heart failure. Treatment and imaging schedules varied; echocardiography could miss small thrombi, and thrombus area did not capture three-dimensional structure. The machine-learning split was small, lacked external validation and calibration, and could overfit. Feature importance is not a biological mechanism. The article also reports different p values for the same regression-versus-event comparison in its table and narrative, although both are nonsignificant. These population-level findings do not predict an individual patient's course. Prospective, multicenter validation in more diverse cohorts would be needed to determine whether either baseline signal adds dependable information.