Differentiating HFmr/rEF from HFpEF using standard 12-lead ECG measurements: an interpretable machine learning study

Q: What are the key findings of this study?

A parsimonious random forest model using 12 routine ECG variables differentiated HFmr/rEF from HFpEF with an AUC of 0.832, performing comparably to a comprehensive model using clinical, laboratory, and ECG predictors.

Q: What is the clinical evidence from this study?

Study design: Cohort. Population: Heart failure (HFmr/rEF vs HFpEF) (n=495). Intervention: 12-feature random forest machine learning model vs. Comprehensive random forest model. Primary outcome: Area under the curve (AUC) for differentiating HFmr/rEF from HFpEF (AUC 0.832, 95% CI 0.766 to 0.899, p=0.288).

Q: What does this research mean for the field?

A parsimonious machine learning model using 12 standard ECG measurements differentiates HFmr/rEF from HFpEF with good discrimination (AUC 0.832), performing similarly to models using comprehensive clinical and laboratory data. Novelty: ClaimNovelty.METHODOLOGICAL. Consensus alignment: ConsensusAlignment.NEUTRAL.

Chenxu Zhao; Po Jie Chan; Scott Dougherty; Yiting Fan; Alex Pui‐Wai Lee

doi:10.1136/openhrt-2026-004127

July 1, 2026Open HeartOpen Access

Differentiating HFmr/rEF from HFpEF using standard 12-lead ECG measurements: an interpretable machine learning study

View Full Paper

Ask AI

Bookmark

Share

Key result

Parsimonious 12-lead ECG model discriminates HFmr/rEF from HFpEF with an AUC of ~0.83.

AUC 0.832
95% CI 0.766 to 0.899
P=0.288
n=495

Figures

Why the study?

Differentiating HFmr/rEF from HFpEF guides therapy, but echocardiography may be delayed or unavailable.

Does a machine learning model using standard 12-lead ECG measurements accurately differentiate HFmr/rEF from HFpEF in hospitalized patients with heart failure?

Key Points

To develop and validate machine learning models using routine 12-lead electrocardiogram (ECG) measurements to differentiate heart failure with mildly reduced or reduced ejection fraction (HFmr/rEF) from preserved ejection fraction (HFpEF).
Analyzed a retrospective cohort of 495 hospitalized patients with heart failure (254 HFmr/rEF and 241 HFpEF) split 70/30 into a training set (n=344) and a held-out test set (n=151).
Trained and tuned Random Forest (RF), Extreme Gradient Boosting, and Support Vector Machine models using fivefold cross-validation, using the Boruta algorithm to identify 12 key ECG features.
A parsimonious RF model using 12 Boruta-selected ECG variables differentiated HFmr/rEF from HFpEF with an AUC of 0.832 (95% CI 0.766 to 0.899) and 76.4% accuracy in the test set.
The 12-feature ECG model showed no significant AUC difference compared to an RF model using all clinical, laboratory, and ECG predictors (AUC 0.804, 95% CI 0.734 to 0.875; p=0.288) or all ECG features alone (AUC 0.821, 95% CI 0.752 to 0.890; p=0.092).

Study Design

Type

Cohort (n=495)

Multicenter

No

Structured PICO

Does a machine learning model using standard 12-lead ECG measurements accurately differentiate HFmr/rEF from HFpEF in hospitalized patients with heart failure?

P

Population

495 hospitalized adults with symptomatic heart failure, evaluated to differentiate mildly reduced/reduced ejection fraction from preserved ejection fraction using 12-lead ECG data.

E

Exposure

Machine learning models (Random Forest, XGBoost, Support Vector Machine) using routine 12-lead ECG data, specifically a parsimonious Random Forest model using 12 Boruta-selected ECG variables.

C

Comparator

Echocardiography-derived left ventricular ejection fraction (LVEF) as the reference standard, and comparison against comprehensive models using clinical, laboratory, and ECG predictors.

O

Outcome

Classification of heart failure phenotype (HFmr/rEF vs HFpEF) evaluated by area under the curve (AUC) and accuracy on a held-out test set.surrogate

Main Result

Effect estimate: AUC 0.832 (95% CI 0.766 to 0.899)

Absolute Event Rate: 0.832% vs 0.804%

p-value: p=0.288

Limitations

Single-centre retrospective design and reliance on vendor-specific structured ECG data may limit generalisability.
Limited sample size required combining HFmrEF with HFrEF, which may obscure an intermediate phenotype.
ECG and echocardiography were not always contemporaneous (median 5 days).
Unmeasured factors such as interim treatment, device therapy, HF aetiology and medication changes could affect both ECG features and EF category.
Evaluated diagnostic performance on a held-out test set but not clinical impact in a prospective workflow.
Single-center cohort
Retrospective design
Requires external validation and prospective evaluation

Ask about this paper

Answers are grounded in this paper and cite their sources.

Implication

A machine learning model using 12 standard ECG features can accurately differentiate HFmr/rEF from HFpEF, potentially aiding in rapid phenotyping when echocardiography is unavailable.

Abstract

BACKGROUND: Differentiating heart failure (HF) with mildly reduced/reduced ejection fraction (HFmr/rEF) from HF with preserved ejection fraction (HFpEF) guides therapy but echocardiography may be delayed or unavailable. We developed and validated machine learning models using routine 12-lead ECG data to classify HF phenotypes. METHODS: In this retrospective cohort of hospitalised patients with HF, predictors available at or before the index ECG were used. HFmrEF was pooled with HFrEF (left ventricular ejection fraction <50%) for model development. Data were split 70/30 into training and held-out test sets. Random forest (RF), Extreme Gradient Boosting and support vector machine models were trained and tuned using fivefold cross-validation in the training set. Boruta was used to select key ECG features. Test-set performance was evaluated by area under the curve (AUC) and accuracy; AUCs were compared using DeLong's test. RESULTS: Overall, 495 patients were included (254 HFmr/rEF; 241 HFpEF). RF consistently performed best. Using ECG features alone, RF achieved an AUC of 0.821 (95% CI 0.752 to 0.890) and accuracy of 75.7%. A parsimonious RF model using 12 Boruta-selected ECG variables achieved an AUC of 0.832 (95% CI 0.766 to 0.899) and accuracy of 76.4%, with no significant AUC difference versus a comprehensive RF model using clinical/laboratory/ECG predictors (AUC 0.804, 95% CI 0.734 to 0.875; p=0.288) or the model using all ECG features (p=0.092). Adding X-ray cardiomegaly did not improve performance. CONCLUSION: A parsimonious RF model based on a small set of standard ECG measurements differentiates HFmr/rEF from HFpEF with good discrimination, supporting ECG as an adjunct for phenotyping when echocardiography is not immediately available.

Discussion

No takes yet. Share an insight, caveat, or question.

Cite This Study

Zhao et al. (2026) conducted a cohort in Heart failure (HFmr/rEF vs HFpEF) (n=495). 12-feature random forest machine learning model vs. Comprehensive random forest model was evaluated on Area under the curve (AUC) for differentiating HFmr/rEF from HFpEF (AUC 0.832, 95% CI 0.766 to 0.899, p=0.288). A parsimonious random forest model using 12 routine ECG variables differentiated HFmr/rEF from HFpEF with an AUC of 0.832, performing comparably to a comprehensive model using clinical, laboratory, and ECG predictors.

synapsesocial.com/papers/6a5b38f68167787360d24e3b https://doi.org/10.1136/openhrt-2026-004127

View Full Paper

Ask AI

Bookmark

Share

Differentiating HFmr/rEF from HFpEF using standard 12-lead ECG measurements: an interpretable machine learning study

Figures

Key Points

Study Design

Structured PICO

Main Result

Limitations

Ask about this paper

Implication

Abstract

Discussion

Cite This Study

Differentiating HFmr/rEF from HFpEF using standard 12-lead ECG measurements: an interpretable machine learning study

Figures

Key Points

Study Design

Structured PICO

Main Result

Limitations

Ask about this paper

Implication

Abstract

Discussion

Cite This Study

Keep reading

Keep reading