Does Electrocardiogram-Based Artificial Intelligence Detect Aortic Stenosis Solely Through Conventional Left Ventricular Hypertrophy Voltage Criteria? ― An Analysis of the Clinical Deep Data Accumulation System (CLIDAS) Database ―

Q: What is the clinical evidence from this study?

Study design: Observational. Population: Aortic stenosis (n=51713). Intervention: Deep learning AI-ECG model vs. Conventional left ventricular hypertrophy (LVH) voltage criteria (Sokolow-Lyon and Cornell). Primary outcome: Discriminative performance (AUC) for detecting aortic stenosis (ICD-10 code I35.0) in the independent test cohort (AUC 0.849, 95% CI 0.832-0.865).

Q: What does this research mean for the field?

Artificial intelligence-enabled electrocardiography detects aortic stenosis with high accuracy and negative predictive value, capturing diagnostic electrocardiographic features beyond conventional left ventricular hypertrophy voltage criteria. Novelty: ClaimNovelty.INCREMENTAL. Consensus alignment: ConsensusAlignment.NEUTRAL.

Takenobu Shimada; Satoshi Kodera; Shinnosuke Sawano; JingChuan Guan; Wataru Saitoh; Shiho Wakasa; Shinji Ito; Tomoya Yanagishita; Yusuke Hayashi; Atsushi Shibata; Asahiro Ito; Kenichiro Otsuka; Yasutomi Higashikuni; Hiroshi Okamura; Koichi Node; Osamu Yamaguchi; Hisaki Makimoto; Tomoyuki Kabutoya; Yasushi Imai; Masaharu Nakayama; Hisahiko Sato; Hideo Fujita; Takahide Kohro; Tetsuya Matoba; Norihiko Takeda; Daiju Fukuda; Ryozo Nagai

doi:10.1253/circrep.cr-26-0209

Key result

AI-ECG detects aortic stenosis with ~0.85 AUC, outperforming conventional LVH voltage criteria.

AUC 0.849
95% CI 0.832-0.865
n=51,713

Why the study?

Routine echocardiographic screening for aortic stenosis is impractical, and although AI-based ECG models show promise for AS detection, whether they primarily reflect conventional LVH voltage criteria or capture additional ECG features remained unclear.

Does a deep learning 12-lead ECG model improve the detection of aortic stenosis compared to conventional left ventricular hypertrophy voltage criteria in hospitalized patients?

Key Points

To determine whether artificial intelligence detects aortic stenosis from electrocardiograms purely via conventional left ventricular hypertrophy voltage criteria or by capturing broader electrical features.
Retrospective deep learning model development and testing using 244,816 ECGs from 51,713 patients across six academic institutions in Japan (CLIDAS database).
AS diagnoses were defined using inpatient Diagnosis Procedure Combination (DPC) codes.
Model performance was tested in an independent test cohort of 7,758 ECGs (one per patient) and compared against Sokolow–Lyon and Cornell LVH voltage criteria, alongside Grad-CAM attention mapping.
The AI model achieved an AUC of 0.849 (95% CI 0.832–0.865) in the test cohort, outperforming Sokolow–Lyon (AUC 0.706) and Cornell voltage (AUC 0.692) criteria.
At a probability threshold of 0.1, the AI model yielded a sensitivity of 79.1%, a specificity of 73.9%, and a negative predictive value of 98.0%.
Adding LVH voltage criteria to the AI model provided no incremental discriminative benefit (AUC 0.849 vs. 0.847), and Grad-CAM confirmed attention focused heavily on limb lead QRS complexes.

Study Design

Type

Observational (n=51,713)

Multicenter

Yes

Structured PICO

Does a deep learning 12-lead ECG model improve the detection of aortic stenosis compared to conventional left ventricular hypertrophy voltage criteria in hospitalized patients?

P

Population

51,713 hospitalized patients (mean age 67.4 years, 38.1% female) across 6 Japanese academic institutions with 244,816 digital 12-lead ECGs recorded within 28 days of admission linked to inpatient diagnosis records.

E

Exposure

2-dimensional convolutional neural network (CNN) artificial intelligence (AI) model applied to standard 12-lead ECG signals (12 leads x 5,000 samples) to predict the probability of aortic stenosis.

C

Comparator

Conventional voltage-based ECG criteria for left ventricular hypertrophy (LVH): Sokolow–Lyon voltage criterion alone, Cornell voltage criterion alone, a logistic regression model combining both voltage criteria, and a combined model incorporating both the AI prediction score and LVH voltage indices.

O

Outcome

Discriminative performance for detecting aortic stenosis (defined as ICD-10 code I35.0 in inpatient DPC diagnostic fields) measured by the area under the receiver operating characteristic curve (AUC) in the independent test cohort.surrogate

Main Result

Effect estimate: AUC 0.849 (95% CI 0.832-0.865)

Absolute Event Rate: 0.849% vs 0.723%

Limitations

Aortic stenosis labels were derived from inpatient administrative diagnosis codes (ICD-10 code I35.0) rather than echocardiographic confirmation, lacking severity stratification.
The study population was derived exclusively from inpatients at academic institutions in Japan, limiting direct generalizability to community-dwelling or primary care settings.
Validation was performed across institutions within the same research project rather than in fully independent external cohorts.
Grad-CAM analysis was applied qualitatively to selected representative cases and was not quantified across the entire cohort.
Comparison with conventional ECG criteria was restricted to voltage-based indices (Sokolow-Lyon and Cornell), excluding other markers like QRS duration, axis deviation, or strain patterns.
Subgroups of patients with left and right bundle branch block were small in size.
AS labels were derived from inpatient DPC diagnosis codes (ICD-10 code I35.0) rather than echocardiographic confirmation, lacking severity stratification and risking coding inaccuracies.
The study population was derived exclusively from inpatients at academic institutions in Japan, which may not be representative of community-dwelling or primary care populations.
All 6 participating institutions were part of the same research project, lacking validation in fully independent external cohorts.
Grad-CAM was applied only to selected representative cases for illustrative purposes rather than quantified across the cohort, and comparisons were limited to voltage-based indices without evaluating QRS duration, axis deviation, or strain patterns.
Subgroups of patients with LBBB and RBBB were limited in size.

Ask about this paper

Answers are grounded in this paper and cite their sources.

Implication

This multicenter deep learning ECG model accurately detects aortic stenosis by capturing waveform features beyond traditional LVH voltage criteria, offering a 98.0% negative predictive value that supports its utility as a clinical rule-out screening tool.

Abstract

Background Aortic stenosis (AS) is a progressive valvular disease associated with poor prognosis once symptoms develop, yet routine echocardiographic screening is impractical. While artificial intelligence (AI)-based electrocardiogram (ECG) models have shown promise for AS detection, it remains unclear whether they primarily reflect conventional left ventricular hypertrophy (LVH) voltage criteria or capture additional ECG features. Methods and Results We developed a deep learning model using 244,816 ECGs from 51,713 patients across six academic institutions in Japan (CLIDAS database). AS labels were derived from inpatient Diagnosis Procedure Combination (DPC) codes. The model achieved an area under the receiver operating characteristic curve (AUC) of 0.849 (95% confidence interval 0.832–0.865) in the independent test cohort, with consistent performance across institutions, sex, and age. At a threshold of 0.1, sensitivity was 79.1%, specificity was 73.9%, and negative predictive value (NPV) was 98.0%. Conventional LVH voltage criteria (Sokolow–Lyon AUC 0.706; Cornell AUC 0.692) showed lower performance, and adding them to the AI model conferred no incremental benefit (AUC 0.849 vs. 0.847). Gradient-weighted class activation mapping (Grad-CAM) revealed predominant attention around QRS complexes in limb leads, beyond regions typically assessed in LVH evaluation. Conclusions This multicenter AI-ECG model demonstrated strong discrimination for AS and captured ECG features beyond conventional LVH voltage criteria. The high NPV supports its use as a rule-out pre-screening tool.

Discussion

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

Cite This Study

Shimada et al. (2026) conducted an observational in Aortic stenosis (n=51,713). Deep learning AI-ECG model vs. Conventional left ventricular hypertrophy (LVH) voltage criteria (Sokolow-Lyon and Cornell) was evaluated on Discriminative performance (AUC) for detecting aortic stenosis (ICD-10 code I35.0) in the independent test cohort (AUC 0.849, 95% CI 0.832-0.865). An AI-ECG model detected aortic stenosis with an AUC of 0.849 (95% CI 0.832-0.865) and 98.0% NPV, outperforming conventional LVH voltage criteria (AUC 0.723).

synapsesocial.com/papers/6ae92596a8a6b2adf62f5fc5 https://doi.org/10.1253/circrep.cr-26-0209

View Full Paper

Ask AI

Bookmark

Share

Does Electrocardiogram-Based Artificial Intelligence Detect Aortic Stenosis Solely Through Conventional Left Ventricular Hypertrophy Voltage Criteria? ― An Analysis of the Clinical Deep Data Accumulation System (CLIDAS) Database ―

Key Points

Study Design

Structured PICO

Main Result

Limitations

Ask about this paper

Implication

Abstract

Discussion

Cite This Study

Does Electrocardiogram-Based Artificial Intelligence Detect Aortic Stenosis Solely Through Conventional Left Ventricular Hypertrophy Voltage Criteria? ― An Analysis of the Clinical Deep Data Accumulation System (CLIDAS) Database ―

Key Points

Study Design

Structured PICO

Main Result

Limitations

Ask about this paper

Implication

Abstract

Discussion

Cite This Study

Keep reading

Keep reading