Association of left ventricular end-systolic volume with peri-echocardiographic acute kidney injury proxy beyond ejection fraction: a retrospective linked echocardiographic cohort study

Abstract

Background

Left ventricular ejection fraction (LVEF) remains the default echocardiographic summary of systolic performance, yet its ratio structure makes it sensitive to geometry and loading conditions. Left ventricular end-systolic volume (LVESV) may capture adverse remodeling more directly. The relation of LVESV to peri-echocardiographic renal dysfunction in hospitalized patients remains uncertain.

Methods

This retrospective linked echocardiographic cohort study analyzed 1,022 complete biplane studies from 784 patients in the credentialed PhysioNet MIMIC-IV-ECHO-Ext-LVVOLUMES-A4C-ROI resource linked to MIMIC-IV v3.1 clinical tables. The primary endpoint was an acute kidney injury (AKI) proxy defined as maximum serum creatinine ≥ 2.0 mg/dL within a ± 24-hour peri-echocardiographic window. Hierarchical logistic regression with patient-level cluster-robust standard errors included age, sex, first-day SOFA score, and vasopressor exposure; 5 chained-equation imputations addressed missing SOFA. Temporal analysis compared echocardiography time with the first creatinine value meeting the endpoint threshold.

Results

AKI proxy was present in 80 of 1,022 studies (7.8%). Among these 80 studies, 73 (91.3%) reached the creatinine threshold before echocardiography and 7 (8.8%) after echocardiography. For AKI proxy, LVESV showed modestly higher discrimination than LVEF (AUC 0.572 vs. 0.548; ΔAUC 0.024; p = 0.253). In the fully adjusted hierarchical model, LVESV remained associated with AKI proxy (OR 1.40 per SD increase, 95% CI 1.10–1.78; p = 0.006), whereas LVEF did not (OR 0.83 per SD decrease, 95% CI 0.62–1.10; p = 0.188). Model 3 yielded the highest C-statistic (0.727) and the lowest AIC (517.0), but BIC was not improved and reclassification gains were modest.

Conclusions

LVESV was associated with peri-echocardiographic AKI proxy after severity adjustment, but the temporal analysis indicates that the signal predominantly reflected concurrent renal illness rather than prospective prediction. Any incremental value beyond LVEF was modest and should be interpreted within the limits of a non-KDIGO creatinine-based proxy endpoint.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12872-026-06133-9.

Introduction

LVEF remains the most familiar summary of left ventricular systolic performance and still anchors heart failure phenotyping, therapeutic eligibility, and many downstream management decisions [1, 2]. Contemporary guideline documents nevertheless acknowledge that the same nominal LVEF can arise from very different geometric and loading states [3].

LVEF is not a direct measure of myocardial contractility. Its ratio structure makes it vulnerable to chamber geometry, afterload, and compensatory fiber recruitment, which can preserve a numerically acceptable ejection fraction despite adverse remodeling [4, 5]. Left ventricular geometry therefore matters clinically, and end-systolic enlargement may convey information that a percentage-based metric leaves behind [6].

Contemporary echocardiographic volume workflows have improved substantially, and agreement with reference standards is now clinically usable in many routine settings [7, 8]. Outcome studies have also shown that ventricular volumes can track prognosis independently of, or more coherently than, ejection fraction alone [9, 10]. Related observations in preserved and reduced ejection-fraction syndromes have kept attention on LVESV as a structural expression of residual cavity burden after contraction [11, 12].

The present linked-database study evaluated whether LVESV was associated with peri-echocardiographic renal dysfunction beyond LVEF in hospitalized patients undergoing echocardiography. Because creatinine-defined renal dysfunction may precede, follow, or coincide with imaging, the analysis was structured to distinguish a concurrent cardiorenal association from a prospective prediction signal.

We hypothesized that greater LVESV would remain associated with a peri-echocardiographic creatinine-threshold AKI proxy after severity adjustment and inclusion of LVEF, while prespecifying temporal and performance analyses to determine whether any observed signal supported prediction or instead reflected concurrent illness burden.

Methods

Study design, reporting framework, data source, and cohort construction

This retrospective linked echocardiographic cohort study used the credentialed PhysioNet resource and the MIMIC-IV v3.1 hospital and ICU databases [13, 14]. The echocardiographic image subset was defined from MIMIC-IV-ECHO-Ext-LVVOLUMES-A4C-ROI [15]. Reporting followed STROBE and RECORD guidance for observational studies using routinely collected health data [16, 17]. Structured echocardiographic metadata were processed in Google BigQuery and linked to de-identified hospital admissions and ICU stays through subject, admission, and stay identifiers with prespecified time rules.

The master extraction contained 1,064 linked echocardiographic studies. Eligibility required a linked hospital admission, an echocardiographic timestamp, and complete biplane LVEF and LVESV measurements. Forty-two studies lacked complete biplane LVEF or LVESV measurements and were excluded before inferential modeling, leaving an analytical cohort of 1,022 studies from 784 patients. MIMIC dates are de-identified and shifted; therefore, absolute calendar dates were not interpreted, and timing rules used relative intervals between admission, ICU stay, laboratory measurements, and echocardiography. This locked analytical cohort was used consistently for all main tables, figures, and regression models.

Outcome definitions and temporal analysis

The primary endpoint was an AKI proxy defined as maximum serum creatinine ≥ 2.0 mg/dL within a ± 24-hour peri-echocardiographic window. This operational definition used a fixed creatinine threshold rather than KDIGO change-based criteria and therefore does not represent adjudicated AKI [18]. Secondary endpoints were vasopressor exposure and in-hospital mortality. LVEF and LVESV were treated as the principal echocardiographic predictors. An exploratory pseudonormal subgroup was prespecified as LVESV at or above the analytical-cohort 75th percentile with LVEF ≥ 40%.

Temporal analysis was restricted to AKI-positive studies in the analytical cohort. The time of the first creatinine value meeting the endpoint threshold was compared with the echocardiographic study timestamp. Admission baseline creatinine was defined as the first creatinine measured during the indexed admission and was used only for contextual temporal analyses, not for primary endpoint adjudication.

Supplementary covariates

Severity adjustment included age, sex, first-day SOFA score, and vasopressor exposure. Comorbidity flags for chronic kidney disease, diabetes mellitus, heart failure, hypertension, coronary artery disease, atrial fibrillation, and chronic obstructive pulmonary disease or asthma were extracted from diagnoses_icd. Nephrotoxic drug exposure within the peri-echocardiographic window was derived from prescriptions and included aminoglycosides, vancomycin, nonsteroidal anti-inflammatory drugs, angiotensin-converting enzyme inhibitor or angiotensin receptor blocker exposure, loop diuretics, intravenous contrast, and amphotericin B. A hospital-wide renal replacement therapy proxy was derived from ICU procedure events and hospital procedure codes. Body mass index (BMI) was extracted from structured anthropometric records where available and evaluated in a complete-case sensitivity analysis; one implausible BMI value outside the prespecified plausible range of 10–80 kg/m² was excluded from the BMI sensitivity dataset.

Statistical analysis

Receiver-operating-characteristic analysis compared LVEF and LVESV for the three study outcomes, with paired bootstrap testing of ΔAUC [19]. Hierarchical logistic regression used patient-level cluster-robust standard errors to account for repeated studies within individuals [20, 21]. Model 1 included age, sex, first-day SOFA score, and vasopressor exposure. Model 2 added LVEF. Model 3 added LVESV to test incremental association beyond LVEF. Variance inflation factors were computed for Model 3.

Continuous variables were summarized as mean ± SD or median with range when distributional features required nonparametric description. Categorical variables were summarized as n (%). Between-group comparisons used Student’s t-test or the Mann–Whitney U test for continuous variables according to distribution and χ² or Fisher’s exact tests for categorical variables. All P values were two-sided, with P < 0.05 considered statistically significant.

Five chained-equation imputations addressed missing first-day SOFA values in the main analysis [22]. Prespecified sensitivity analyses added nephrotoxic exposure and loop diuretic use, added chronic kidney disease, added broader comorbidity adjustment, and added the renal replacement therapy proxy. ICU subgroup analysis used observed SOFA values without imputation. Non-ICU, surgical, and elective subgroup models were treated as exploratory and were not emphasized when event counts were too low for stable multivariable estimation. Data extraction and linkage were performed in Google BigQuery. Additional tabulation, model checking, BMI-adjusted sensitivity modeling, ROC summaries, and document quality checks were performed in Python 3.12.13 using pandas 3.0.3, statsmodels 0.14.6, and scikit-learn 1.8.0.

Results

AKI proxy was present in 80 of 1,022 studies (7.8%). Patients with AKI proxy had larger LVESV, larger LVEDV, more vasopressor exposure, and lower female representation than those without AKI proxy, whereas the mean LVEF difference was modest and did not reach conventional significance (Table 1). Chronic kidney disease was present in 83.8% of AKI-positive studies compared with 22.3% of AKI-negative studies. Diabetes, heart failure, hypertension, and coronary artery disease were also more frequent in the AKI group (Supplementary file 1: Table S1). Any nephrotoxic exposure occurred in 25.0% of AKI-positive studies and 6.4% of AKI-negative studies, driven mainly by vancomycin and loop diuretics (Supplementary file 1: Table S2).

Table 1.

Baseline characteristics of the analytical cohort by AKI proxy status (No AKI proxy n = 942; AKI proxy n = 80)

Variable	No AKI proxy (n = 942)	AKI proxy (n = 80)	P value
Age (years), mean ± SD	65.4 ± 13.4	67.1 ± 12.1	0.252
Female sex, n (%)	515 (54.7%)	32 (40.0%)	0.014
LVEF biplane (%), mean ± SD	59.0 ± 13.0	55.6 ± 16.1	0.065
LVESV biplane (mL), mean ± SD	41.7 ± 31.0	57.1 ± 48.7	0.007
LVEDV A4C (mL), mean ± SD	99.0 ± 45.5	114.8 ± 62.3	0.029
Vasopressor use, n (%)	14 (1.5%)	10 (12.5%)	< 0.001

Temporal analysis showed that the creatinine threshold was reached before echocardiography in 73 of 80 AKI-positive studies (91.3%) and after echocardiography in 7 studies (8.8%). The median time from echocardiography to threshold crossing was − 30 h (range − 445 to -1) in the pre-echo group and + 13 h (range 5 to 20) in the post-echo group. Admission baseline creatinine was higher in studies with pre-echo renal deterioration than in the small post-echo subset (median 3.0 vs. 1.8 mg/dL). These distributions indicate that echocardiography was usually obtained during or after renal dysfunction rather than before it (Supplementary file 1: Table S3 and Figure S1).

For AKI proxy, discrimination remained modest for both echocardiographic indices. LVESV yielded an AUC of 0.572 and LVEF an AUC of 0.548, with no significant difference between them (ΔAUC + 0.024; p = 0.253). For vasopressor exposure, LVEF outperformed LVESV (AUC 0.734 vs. 0.639; ΔAUC − 0.095; p = 0.014). For in-hospital mortality, discrimination was similar and low-to-moderate for both metrics (Table 2). Across the hierarchical clinical models, the C-statistic was 0.720 for Model 1, 0.717 for Model 2, and 0.727 for Model 3. AIC improved slightly with Model 3, whereas BIC did not. Category-free reclassification from Model 2 to Model 3 was modest (NRI + 0.155; IDI + 0.008), and decision-curve analysis showed only small net-benefit separation at lower threshold probabilities (Supplementary file 1: Table S4 and Figure S2).

Table 2.

Discrimination of LVEF and LVESV across study outcomes

Outcome	LVEF AUC	LVESV AUC	ΔAUC	Bootstrap P value
AKI proxy	0.548	0.572	+ 0.024	0.253
Vasopressor exposure	0.734	0.639	-0.095	0.014
In-hospital mortality	0.685	0.669	-0.015	0.687

In the hierarchical multivariable analysis, first-day SOFA score and vasopressor exposure remained the dominant clinical covariates. LVEF was not associated with AKI proxy in Model 2 (OR 1.07 per SD decrease, 95% CI 0.86–1.34; p = 0.532). After LVESV was added, the LVESV term remained associated with AKI proxy (OR 1.40 per SD increase, 95% CI 1.10–1.78; p = 0.006), whereas the LVEF term shifted toward the null (OR 0.83, 95% CI 0.62–1.10; p = 0.188; Table 3). The joint distribution of LVEF and LVESV and the overlapping ROC curves are shown in Fig. 1. Variance inflation factors remained below 3.0 for all covariates, including LVEF (2.27) and LVESV (2.57), which supported collinearity-aware interpretation rather than model instability (Supplementary file 1: Table S5).

Table 3.

Hierarchical logistic regression models for AKI proxy

Variable	Odds ratio	95% CI	P value
Model 1: Clinical baseline
Age (per year)	1.01	0.99–1.03	0.151
Male sex	1.51	0.85–2.69	0.158
First-day SOFA score	1.17	1.09–1.26	< 0.001
Vasopressor exposure	4.93	1.76–13.84	0.002
Model 2: Clinical baseline + LVEF
Age (per year)	1.01	1.00-1.04	0.142
Male sex	1.48	0.82–2.65	0.193
First-day SOFA score	1.17	1.09–1.26	< 0.001
Vasopressor exposure	4.59	1.61–13.03	0.004
LVEF (per SD decrease)	1.07	0.86–1.34	0.532
Model 3: Clinical baseline + LVEF + LVESV
Age (per year)	1.02	1.00-1.04	0.077
Male sex	1.24	0.67–2.32	0.496
First-day SOFA score	1.16	1.08–1.25	< 0.001
Vasopressor exposure	4.72	1.58–14.11	0.005
LVEF (per SD decrease)	0.83	0.62–1.10	0.188
LVESV (per SD increase)	1.40	1.10–1.78	0.006

Fig. 1.

LVESV and LVEF in relation to AKI proxy. Panel A shows AKI proxy rate across LVEF quintiles with mean LVESV overlaid. Panel B shows ROC curves for AKI proxy. Panel C shows the joint distribution of LVEF and LVESV by AKI proxy status. Dashed lines denote the exploratory LVEF threshold of 40% and the analytical-cohort LVESV 75th percentile

The cohort included 512 ICU studies (50.1%) and 510 non-ICU studies (49.9%). Medical services accounted for 81.8% of studies, surgical services for 13.3%, and other services for 4.9%. Observation admissions were common, but emergency admissions carried the largest absolute number of AKI events (Supplementary file 1: Table S6). AKI proxy occurred in 69 of 512 ICU studies (13.5%) and in 11 of 510 non-ICU studies (2.2%). In the ICU subgroup, LVESV remained associated with AKI proxy after adjustment with observed SOFA values (OR 1.47, 95% CI 1.09–1.97; p = 0.011; C-statistic 0.735). Finer multivariable subgroup models were not emphasized outside the ICU because event counts were too low for stable estimation (Supplementary file 1: Table S7).

Sensitivity analyses changed the strength of the LVESV signal more than its direction. Adjustment for nephrotoxic exposure and loop diuretic use preserved the association (OR 1.32, 95% CI 1.05–1.66; p = 0.017). Adjustment for chronic kidney disease attenuated the association (OR 1.29, 95% CI 0.94–1.78; p = 0.112). Broader adjustment for chronic kidney disease, diabetes, nephrotoxic exposure, and loop diuretics further weakened the LVESV term (OR 1.19, 95% CI 0.88–1.62; p = 0.260). Addition of the renal replacement therapy proxy produced a similar attenuation (OR 1.20, 95% CI 0.87–1.66; p = 0.266; Supplementary file 1: Table S8).

BMI was available for 963 of 1,022 studies (94.2%), including 76 of 80 AKI-positive studies. After exclusion of one implausible BMI value and complete-case restriction for observed SOFA and BMI, the BMI-adjusted sensitivity model included 484 studies and 65 AKI-proxy events. LVESV remained associated with AKI proxy in this sensitivity model (OR 1.40 per SD increase, 95% CI 1.04–1.90; p = 0.029), whereas BMI was not associated with the endpoint (OR 1.04 per 5 kg/m² increase, 95% CI 0.80–1.34; p = 0.791; Supplementary file 1: Table S9).

Discussion

This linked cohort analysis clarified the interpretation more than it changed the direction of the signal. LVESV remained associated with the peri-echocardiographic AKI proxy after adjustment for age, sex, SOFA score, vasopressor exposure, and LVEF, but the effect size was modest and discrimination stayed in a low-to-moderate range. The central message is therefore not prospective prediction. The more defensible conclusion is that enlarged end-systolic volume tracked a concurrent cardiorenal illness state that LVEF alone did not fully summarize.

The temporal analysis was decisive. In 73 of 80 AKI-positive studies, the creatinine threshold had already been crossed before the echocardiographic examination. Bedside echocardiography in critical care is often obtained after hemodynamic or renal deterioration becomes clinically evident, and our temporal distribution fit that practice pattern [23]. The small post-echo subset prevented a stable prospective-style adjusted model, so any claim that LVESV predicted later AKI would have exceeded what the data can support.

The physiologic rationale for studying LVESV remains sound within this interpretation. LVESV reflects the residual ventricular cavity after contraction and is more directly linked to remodeling than a ratio-based ejection fraction [6]. Contemporary echocardiography has improved the reproducibility of volume assessment, although agreement with reference standards still depends on acquisition quality and method [7, 24]. Studies in ambulatory and disease-specific cohorts have likewise shown that ventricular volumes can carry information that a preserved ejection fraction partially obscures [9, 25]. Our data align with that structural concept, but only within the restricted context of a peri-echocardiographic creatinine-based endpoint.

The multivariable behavior of LVEF and LVESV also deserves careful interpretation. The LVEF coefficient moved toward the opposite direction after LVESV entered the model, but the variance-inflation profile remained below the usual concern threshold. This pattern is consistent with suppression between two geometrically related variables rather than catastrophic multicollinearity. At the same time, model-performance metrics prevented overstatement. C-statistic gains were small, reclassification was driven mainly by non-events and BIC did not favor the more complex model. Calibration of Model 3 was imperfect (Hosmer-Lemeshow P = 0.030); the small discrimination gain should therefore not be interpreted as a robust improvement in absolute risk estimation. Any complementary value was therefore incremental rather than transformative.

Confounding by chronic renal burden and treatment intensity likely explained part of the observed association. Chronic kidney disease was present in more than four fifths of AKI-positive studies, and the LVESV term weakened after CKD adjustment. Nephrotoxic exposure, especially vancomycin and loop diuretics, also clustered strongly in AKI-positive studies. The renal replacement therapy proxy did not materially rescue the signal. Together, these findings suggest that the association linked LV remodeling to a broader burden of illness, chronic kidney disease, and treatment intensity rather than to a clean isolated renal outcome.

Heterogeneity across care settings further limited generalizability. Half of the cohort came from ICU-linked studies, and nearly all stable multivariable information about AKI was concentrated there. The ICU subgroup still showed an LVESV signal after adjustment with observed SOFA, but non-ICU, surgical, and elective strata contained too few events for dependable effect estimation. Those sparse-data conditions argue against broad extrapolation and support treating finer subgroup findings as exploratory.

This study has several limitations. The endpoint was an AKI proxy defined by an absolute creatinine threshold rather than KDIGO change criteria, urine output could not be adjudicated reliably for the full cohort, and admission baseline creatinine was only a practical proxy [18]. Dialysis was not part of the primary endpoint, although an auxiliary renal replacement therapy proxy was examined. Echocardiography indication, full fluid balance, and nephrotoxic burden were incompletely captured. BMI could be assessed only as a complete-case sensitivity analysis and should not be interpreted as eliminating residual confounding by body size, adiposity, or treatment intensity. The source data came from a single US tertiary-care health system, and the echocardiographic subset may not represent all hospitalized patients undergoing routine echocardiography. The study remained retrospective and observational, so association should not be interpreted causally. The next useful step is a prospectively time-anchored study that pairs echocardiography with adjudicated KDIGO trajectories before and after imaging.

Conclusions

In this linked echocardiographic cohort, LVESV was associated with a peri-echocardiographic creatinine-threshold AKI proxy after adjustment for clinical severity and LVEF. However, most AKI-proxy events had already occurred before echocardiography, discrimination was modest, and incremental model-performance gains were small. LVESV should therefore be interpreted as a structural marker tracking concurrent cardiorenal illness burden rather than as a stand-alone prospective AKI prediction tool.

Abbreviations

AKI

Acute kidney injury

AUC

Area under the curve

CAD

Coronary artery disease

CKD

Chronic kidney disease

DCA

Decision-curve analysis

ICU

Intensive care unit

IDI

Integrated discrimination improvement

LVEDV

Left ventricular end-diastolic volume

LVEF

Left ventricular ejection fraction

LVESV

Left ventricular end-systolic volume

NRI

Net reclassification improvement

OR

Odds ratio

RECORD

REporting of studies Conducted using Observational Routinely-collected health Data

ROC

Receiver-operating-characteristic

RRT

Renal replacement therapy

SD

Standard deviation

SOFA

Sequential Organ Failure Assessment

STROBE

Strengthening the Reporting of Observational Studies in Epidemiology

Heading Hasan Burak Isleyen, MD

Hasan Burak Isleyen is an Assistant Professor of Cardiology at Nişantaşı University Faculty of Medicine and serves as Education and Research Responsible Physician at Bezmialem Vakıf University Department of Cardiology. His research interests include intensive care cardiology, echocardiography, cardio-oncology, cardiovascular outcomes, clinical prediction models, and secondary analyses of large-scale clinical databases.

Authors’ contributions

HBI and MK conceptualized the study. HBI, FK and CA designed the methodology. HBI performed the formal analysis. HBI and FK curated the data. HBI and AAS developed the visualizations. STY, SB, CA, NZE, AAS and MK contributed to investigation and interpretation. HBI drafted the manuscript. All authors critically revised the manuscript, approved the final version, and agree to be accountable for all aspects of the work.

Funding

The authors received no external funding for this study.

Data availability

The data that support the findings of this study are available from the PhysioNet MIMIC-IV and MIMIC-IV-ECHO resources under credentialed access and are therefore not redistributed by the authors. Researchers can obtain access directly from PhysioNet after completion of the required training and data use process. The analysis code and reproducibility materials are archived in Zenodo [26].

Declarations

Ethics approval and consent to participate

The study used de-identified credentialed data obtained through PhysioNet. Individual informed consent was waived under local policy. The project was approved by the Department of Cardiology Academic Board, Bezmialem Foundation University Faculty of Medicine (Meeting No. 3, 12 March 2026; Decision No. 5).

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.