Foundation Model Artificial Intelligence for Electrocardiogram Detection of Rheumatic Mitral Stenosis

Q: What is the clinical evidence from this study?

Study design: Observational. Population: Rheumatic Mitral Stenosis. Intervention: Fine-tuned Electrocardiogram Foundation Model (ECG-FM) vs. ICD-9/10 diagnosis codes for Rheumatic Mitral Stenosis across 12-Lead, 6-Lead, and 1-Lead ECG formats. Primary outcome: Area under the receiver operating characteristic curve (AUROC) for detecting ICD-coded rheumatic mitral stenosis on 12-lead ECG (AUROC 0.89 (internal), AUROC 0.82 (external), 95% CI 0.88-0.90).

Q: What does this research mean for the field?

Fine-tuned electrocardiogram foundation models can detect rheumatic mitral stenosis across standard 12-lead and reduced-lead formats with external AUROCs ranging from 0.77 to 0.82. Novelty: ClaimNovelty.METHODOLOGICAL. Consensus alignment: ConsensusAlignment.NEUTRAL.

Raghav Sriram; Ivan Nenadic; Zak Loring; Marat Fudim; Anthony Mazella; Venkatesh Locharla Murthy; Anil K. Gehi

doi:10.64898/2026.10.12.26365300

October 13, 2026Open Access

Foundation Model Artificial Intelligence for Electrocardiogram Detection of Rheumatic Mitral Stenosis

View Full Paper

Ask AI

Bookmark

Share

Key result

Fine-tuned ECG models detect rheumatic mitral stenosis with ~0.82 external AUROC.

AUROC 0.89 (internal), AUROC 0.82 (external)
95% CI 0.88-0.90

Why the study?

Rheumatic mitral stenosis disproportionately affects low-resource settings with limited echocardiography availability, prompting evaluation of foundation model feasibility for detection using limited, indirectly labeled ECGs.

Does a fine-tuned electrocardiogram foundation model accurately detect ICD-coded rheumatic mitral stenosis across 12-lead, 6-lead, and 1-lead ECG configurations in US hospital cohorts?

Key Points

To evaluate whether fine-tuning an electrocardiogram foundation model can accurately detect rheumatic mitral stenosis across standard and reduced-lead formats using limited, diagnostic code-derived labels.
A transformer-based electrocardiogram foundation model (ECG-FM) was fine-tuned using five-fold cross-validation on 804 positive and 804 negative ECGs per fold from Massachusetts General Hospital, identified by ICD-9/10 codes.
The model was tested across 12-lead, 6-lead, and 1-lead configurations and externally validated on an Emory University Hospital cohort comprising 529 positive ECGs and 52,349 negative ECGs.
Internal test performance achieved AUROCs of 0.89 [0.88, 0.90] for 12-lead, 0.87 [0.86, 0.88] for 6-lead, and 0.87 [0.86, 0.88] for 1-lead formats, with corresponding AUPRCs of 0.08, 0.07, and 0.08.
External validation demonstrated AUROCs of 0.82 [0.81, 0.83] for 12-lead, 0.80 [0.78, 0.82] for 6-lead, and 0.77 [0.75, 0.79] for 1-lead formats, with consistent performance observed across demographic subgroups.

Study Design

Type

Observational

Multicenter

Yes

Structured PICO

Does a fine-tuned electrocardiogram foundation model accurately detect ICD-coded rheumatic mitral stenosis across 12-lead, 6-lead, and 1-lead ECG configurations in US hospital cohorts?

P

Population

Adults (~52% female, mean age 54.65-58.32 years) from two US hospital cohorts with ICD-coded rheumatic mitral stenosis (ECG within 2 years prior to diagnosis) and controls without rheumatic heart disease.

E

Exposure

Fine-tuned electrocardiogram foundation model (ECG-FM, a transformer-based model with 90,374,017 parameters) applied to 12-lead (12L), 6-lead limb (6L), and 1-lead (Lead I, 1L) ECG configurations

C

Comparator

ICD-9 (394.0) or ICD-10 (I05.0) diagnostic code-derived rheumatic mitral stenosis labels vs control ECGs without coded rheumatic heart disease or rheumatic mitral stenosis

O

Outcome

Diagnostic discrimination for detecting ICD-coded rheumatic mitral stenosis measured by Area Under the Receiver Operating Characteristic curve (AUROC) and Area Under the Precision-Recall Curve (AUPRC)surrogate

Main Result

Effect estimate: AUROC 0.89 (internal), AUROC 0.82 (external) (95% CI 0.88-0.90)

Fine-tuning an ECG foundation model enables moderate-to-high discrimination for detecting ICD-coded rheumatic mitral stenosis across 12-lead, 6-lead, and single-lead ECGs, supporting its potential as an echocardiography triage tool in resource-limited settings.

Limitations

Reliance on ICD-based rheumatic mitral stenosis labels rather than echocardiography-confirmed diagnoses.
Control patients were defined by absence of RHD codes, which may not exclude undiagnosed or miscoded disease.
Reduced specificity in the external validation cohort indicates potential for substantial false-positives in high-sensitivity screening.
Lack of prospective evaluation on low-cost edge devices and lack of validation in RHD-endemic low- or middle-income countries.
Reliance on ICD-based rheumatic mitral stenosis labels rather than echocardiogram-confirmed mitral stenosis
Control patients defined by absence of RHD codes, which may not exclude undiagnosed or miscoded disease
Specificity declined in external validation, indicating potential for substantial false-positives if used with a high-sensitivity threshold
Lack of prospective evaluation on low-cost edge devices to determine computational efficiency, downstream echocardiography burden, and clinical utility
Lack of validation in low- or middle-income countries (LMICs) where disease spectrum, comorbidities, ECG acquisition conditions, and pretest probability differ from US cohorts

Ask about this paper

Answers are grounded in this paper and cite their sources.

Implication

May aid echocardiography triage for rheumatic mitral stenosis in low-resource settings; leaves open prospective validation before clinical use.

Abstract

Background Rheumatic mitral stenosis (RMS) disproportionately affects low-resource settings where echocardiography is limited. We evaluated foundation model (FM) feasibility for RMS detection using limited, indirectly labeled ECGs. Methods An ECG-FM was fine-tuned using 804 RMS-positive ECGs per fold and externally validated across clinical and wearable lead configurations. Results Internal AUROCs were 0.89, 0.87, and 0.87, for 12-Lead, 6-Lead, and 1-Lead respectively. External AUROCs were 0.82, 0.80, and 0.77. Performance remained comparable across demographic subgroups. Conclusions FM fine-tuning enables RMS detection using limited ICD-derived labels, supporting further evaluation of reduced-lead ECG screening and echocardiography triage in resource-constrained settings.

Discussion

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

Cite This Study

Sriram et al. (2026) conducted an observational in Rheumatic Mitral Stenosis. Fine-tuned Electrocardiogram Foundation Model (ECG-FM) vs. ICD-9/10 diagnosis codes for Rheumatic Mitral Stenosis across 12-Lead, 6-Lead, and 1-Lead ECG formats was evaluated on Area under the receiver operating characteristic curve (AUROC) for detecting ICD-coded rheumatic mitral stenosis on 12-lead ECG (AUROC 0.89 (internal), AUROC 0.82 (external), 95% CI 0.88-0.90). Fine-tuned ECG foundation models detected rheumatic mitral stenosis with internal AUROCs of 0.89 (12L), 0.87 (6L), and 0.87 (1L), and external AUROCs of 0.82, 0.80, and 0.77.

synapsesocial.com/papers/6ae28e56cbadf70daf6c41c0 https://doi.org/10.64898/2026.10.12.26365300

View Full Paper

Ask AI

Bookmark

Share

Foundation Model Artificial Intelligence for Electrocardiogram Detection of Rheumatic Mitral Stenosis

Key Points

Study Design

Structured PICO

Main Result

Limitations

Ask about this paper

Implication

Abstract

Discussion

Cite This Study

Foundation Model Artificial Intelligence for Electrocardiogram Detection of Rheumatic Mitral Stenosis

Key Points

Study Design

Structured PICO

Main Result

Limitations

Ask about this paper

Implication

Abstract

Discussion

Cite This Study

Keep reading

Keep reading