cs.LGMar 30, 2026
SaveDetecting low left ventricular ejection fraction from ECG using an interpretable and scalable predictor-driven framework
Organizations: Department of Information Center, Fuwai Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing, China · The Tsinghua Shenzhen International Graduate School, Tsinghua University, Shenzhen, China · Institute of Statistics and Big Data, Renmin University of China, Beijing, China · Cardiac Arrhythmia Center, Fuwai Hospital, National Center for Cardiovascular Diseases, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing, China · Function Test Center, Fuwai Hospital, National Center for Cardiovascular Diseases, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing, China
Abstract
Low left ventricular ejection fraction (LEF) frequently remains undetected until progression to symptomatic heart failure, underscoring the need for scalable screening strategies. Although artificial intelligence-enabled electrocardiography (AI-ECG) has shown promise, existing approaches rely solely on end-to-end black-box models with limited interpretability or on tabular systems dependent on commercial ECG measurement algorithms with suboptimal performance. We introduced ECG-based Predictor-Driven LEF (ECGPD-LEF), a structured framework that integrates foundation model-derived diagnostic probabilities with interpretable modeling for detecting LEF from ECG. Trained on the benchmark EchoNext dataset comprising 72,475 ECG-echocardiogram pairs and evaluated in predefined independent internal (n=5,442) and external (n=16,017) cohorts, our framework achieved robust discrimination for moderate LEF (internal AUROC 88.4%, F1 64.5%; external AUROC 86.8%, F1 53.6%), consistently outperforming the official end-to-end baseline provided with the benchmark across demographic and clinical subgroups. Interpretability analyses identified high-impact predictors, including normal ECG, incomplete left bundle branch block, and subendocardial injury in anterolateral leads, driving LEF risk estimation. Notably, these predictors independently enabled zero-shot-like inference without task-specific retraining (internal AUROC 75.3-81.0%; external AUROC 71.6-78.6%), indicating that ventricular dysfunction is intrinsically encoded within structured diagnostic probability representations. This framework reconciles predictive performance with mechanistic transparency, supporting scalable enhancement through additional predictors and seamless integration with existing AI-ECG systems.
Figures & tables
Figure 1: Flowchart of model development and evaluation. Model development is based on a traditional AI-ECG interpretation model, where a large unlabeled ECG dataset is used for pre-training and a smaller labeled ECG dataset is used for enhanced post-training. The automatic ECG diagnosis model outputs diagnostic probabilities for each ECG finding and serves as the predictor extractor. Based on the predictor extractor, we developed both a single-predictor approach and a multi-predictor approach. The single-predictor approach enables zero-shot-like inference without further training to detect LEF and can serve as an important indicator. The multi-predictor approach further trains a tabular model using ECG-ECHO pairs and provides both local- and global-level explanations based on SHAP values. Model evaluation is conducted on the EchoNext test set and an independently collected ECHO-Note pairs dataset. Multidimensional model dissection is performed, including overall performance evaluation, model interpretability analysis, and subgroup analyses across diverse populations. LEF, low left ventricular ejection fraction; No-LEF, absence of low left ventricular ejection fraction; SHD, structural heart disease; VHD, valvular heart disease, SHAP, Shapley Additive exPlanations.
| ECG–ECHO | ECG–Note | |||
| Training set | Validation set | Internal test set | External test set | |
| Patients (n) | 26,218 | 4,626 | 5,442 | 16,017 |
| ECGs (n) | 72,475 | 4,626 | 5,442 | 16,017 |
| Age groups | ||||
| 18–59 | 29,783 (41.1%) | 1,787 (38.6%) | 2,124 (39.0%) | 4,270 (26.7%) |
| 60–69 | 18,745 (25.9%) | 1,093 (23.6%) | 1,318 (24.2%) | 3,637 (22.7%) |
Table 1: Baseline demographic and clinical characteristics in the ECG-ECHO and external ECG-Note cohorts.
| Predictor | Traditional ECG Model | LEF Detection (Single-predictor) | ||||||
| AUROC | AUPRC | F1 Score | Thresh | AUROC | AUPRC | F1 Score | Thresh | |
| NORM | 94.9 (94.1–95.7) | 92.8 (91.4–94.2) | 85.1 (83.5–86.7) | 0.370 | 81.0 (79.6–82.4) | 47.4 (44.2–50.9) | 51.4 (49.3–53.8) | 0.003641 |
| ILBBB | 90.9 (65.3–99.7) | 31.6 ( 8.1–70.3) | 30.0 ( 0.0–55.6) | 0.163 | 80.0 (78.4–81.5) | 48.5 (45.3–52.0) | 50.5 (48.0–53.1) | 0.000349 |
| INJAL | 98.6 (97.3–99.6) | 52.2 (27.0–79.8) | 47.6 (19.0–72.0) | 0.500 | 75.3 (73.7–76.8) | 34.8 (32.4–37.6) | 45.1 (42.8–47.4) | 0.000386 |
| ISCLA | 92.3 (85.5–97.6) | 19.0 ( 5.6–45.9) | 12.5 ( 0.0–36.4) | 0.248 | 75.9 (74.5–77.4) | 38.9 (36.2–42.3) | 43.9 (41.6–46.3) | 0.001918 |
| ANEUR | 96.7 (91.7–99.2) | 15.0 ( 6.0–37.3) | 11.8 ( 0.0–33.3) | 0.294 | 74.8 (73.1–76.6) | 38.9 (35.9–42.3) | 42.9 (40.4–45.4) | 0.001924 |
Table 2: Performance of traditional AI-ECG predictors and LEF detection using a single-predictor approach.
| Method | Tabular Model | Predictor Extractor | AUROC | AUPRC | F1 Score |
| Internal test set | |||||
| Columbia mini model | – | – | 85.2 (83.9–86.5) | 59.9 (56.6–63.3) | 57.9 (55.4–60.3) |
| ECGPD-LEF | Logistic Regression | Transformer | 84.5 (83.1–85.9) | 56.8 (53.4–60.3) | 57.0 (54.6–59.7) |
| Transformer-PT | 84.9 (83.6–86.3) | 57.8 (54.6–61.2) | 58.5 (56.2–60.8) | ||
| XGBoost | Transformer | 84.9 (83.5–86.2) | 58.5 (55.1–61.7) | 57.6 (55.1–60.1) | |
| Transformer-PT | 88.4 (87.1–89.5) | 69.2 (66.2–72.0) | 64.5 (62.2–66.7) | ||
Table 3: Performance comparison of the proposed ECGPD-LEF framework and baseline Columbia mini model for LEF detection.
Figure 2: Model performance across different evaluation settings. The first row shows receiver operating characteristic (ROC) curves (a) and precision–recall (PR) curves (b) for five methods using the full feature set (71 predictors). The second row presents the performance of four models evaluated with an increasing number of predictors (1–71), in terms of AUROC (c), AUPRC (d) and F1 score (e). The gray dashed line indicates the Columbia mini model and serves as a reference baseline.
Figure 3: Global-level explanation of ECG model predictions across all predictors. The first row shows (a) the cumulative absolute SHAP contributions of all 71 predictors, expressed as percentages of the total contribution, and (b) a SHAP beeswarm plot for the top 10 predictors ranked by F1 score obtained from the single-predictor method. The second and third rows show (c-h) the relationships between individual predictors (NORM, ILBBB, ISCLA, ANEUR, INJIL and ASMI) and their SHAP values. Point density was estimated using a Gaussian kernel density on the log-transformed values. Each plot includes two vertical reference lines: the first (LEF threshold) indicates the threshold that achieves the optimal F1 score using a single-predictor method, and the second (Diagnosis threshold) indicates the positivity threshold defined by an independent ECG diagnosis model.
Figure 4: Local-level explanation of model predictions for a positive and a negative case. The first row shows SHAP waterfall plots for a positive case (a) and a negative case (b), illustrating the top ten ECG predictors contributing to the prediction, with the remaining predictors aggregated as “62 other features”. Feature contributions are shown in the log-odds space and sum to the final model output. The second row presents the corresponding predictor–probability relationships for the same cases (c,d). Vertical reference lines indicate positivity thresholds derived from an independent ECG diagnosis model.
| Columbia mini model | ECGPD-LEF (Multi-predictor) | |||||||
| Subgroup | n | Prevalence (%) | AUROC | AUPRC | F1 Score | AUROC | AUPRC | F1 Score |
| Age groups | ||||||||
| 18–59 | 2124 | 13.6 | 85.2 (82.8–87.5) | 53.5 (48.1–60.0) | 53.8 (49.0–58.3) | 87.4 (85.0–89.8) | 64.6 (59.4–69.9) | 59.0 (54.7–63.6) |
| 60–69 | 1318 | 16.6 | 83.8 (80.9–86.6) | 57.6 (51.1–64.6) | 54.7 (49.0–59.7) | 87.2 (84.2–90.2) | 70.6 (64.9–76.3) | 63.3 (58.1–68.5) |
| 70–79 | 1154 | 21.7 | 86.2 (83.7–88.7) | 65.8 (59.5–72.2) | 63.7 (58.9–67.9) | 89.7 (87.2–91.8) | 74.7 (69.5–80.0) | 70.5 (65.8–74.8) |
| 80+ | 846 | 24.1 | 83.5 (80.4–86.5) | 65.4 (59.5–71.8) | 59.5 (54.7–64.4) | 87.3 (84.8–89.7) | 67.7 (60.9–74.1) | 66.1 (61.4–70.8) |
Table 4: Subgroup performance of ECGPD-LEF and the Columbia mini model in the internal test set.
Appendix figures & tables14 assets
Supplementary material from the paper’s appendix.
Appendix
| ECG-ECHO | ECG–Note | |||
|---|---|---|---|---|
| Training set | Validation set | Test set | External Set | |
| Estimated predictor | ||||
| NORM | 19,259 (26.6%) | 1,485 (32.1%) | 1,806 (33.2%) | 3,343 (20.9%) |
| ILBBB | 161 (0.2%) | 8 (0.2%) | 8 (0.1%) | 74 (0.5%) |
| INJAL | 666 (0.9%) | 48 (1.0%) | 48 (0.9%) | 108 (0.7%) |
| ISCLA | 109 (0.2%) | 3 (0.1%) | 6 (0.1%) | 24 (0.1%) |
Table A.1: Estimated patient characteristics across the training, validation, and test splits of the EchoNext dataset and the external cohort.
Figure A.1: Construction of the external test set. (a) Flowchart illustrating the data selection process, including exclusion criteria applied to the MIMIC-IV database to derive the final high-quality testing set. (b) The prompt template used for the Large Language Model (LLM) Clinical Note Extraction Module, designed to extract quantitative ejection fraction(EF) values from unstructured clinical notes into a structured JSON format.
| Training set | Validation set | Test set | |
| Echocardiographic findings | |||
| LVWT 1.3cm | 17,667 (24.4%) | 877 (19.0%) | 1,061 (19.5%) |
| Aortic stenosis | 2,919 (4.0%) | 252 (5.4%) | 286 (5.3%) |
| Aortic regurgitation | 878 (1.2%) | 62 (1.3%) | 66 (1.2%) |
| Mitral regurgitation | 6,137 (8.5%) | 282 (6.1%) | 337 (6.2%) |
| Tricuspid regurgitation | 7,707 (10.6%) | 305 (6.6%) | 353 (6.5%) |
Table B.1: Baseline echocardiographic characteristics across the training, validation, and test splits of the ECG-ECHO dataset.
Figure B.1: Model performance stratified by other structural heart disease (SHD). ROC (a,c) and PR (b,d) curves for individuals with other SHD (a,b) and without other SHD (c,d). Other SHD was defined as the presence of at least one predefined echocardiographic abnormality independent of left ventricular ejection fraction (see Appendix B.1 ). Transformer-PT-XGBoost is compared with the Columbia mini model. AUROC and AUPRC are indicated in each panel.
Figure B.2: Model performance stratified by valvular heart disease (VHD). ROC (a,c) and PR (b,d) curves for individuals with moderate or greater VHD (a,b) and without moderate or greater VHD (c,d). VHD was defined as moderate or greater valvular heart disease on echocardiography (see Appendix B.1 ). Transformer-PT-XGBoost is compared with the Columbia mini model. AUROC and AUPRC are indicated in each panel.
| Columbia mini model | ECGPD-LEF (Multi-predictor) | |||||||
| Subgroup | n | Prevalence (%) | AUROC | AUPRC | F1 Score | AUROC | AUPRC | F1 Score |
| Age groups | ||||||||
| 18–59 | 4270 | 12.2 | 83.0 (80.9–84.9) | 46.0 (41.5–50.8) | 49.1 (45.4–52.4) | 89.5 (88.1–90.9) | 62.1 (57.4–66.4) | 57.3 (54.2–60.4) |
| 60–69 | 3637 | 15.7 | 82.0 (80.2–83.9) | 47.7 (43.7–52.6) | 50.1 (46.9–53.1) | 88.8 (87.3–90.3) | 65.1 (61.4–69.2) | 56.3 (53.4–59.3) |
| 70–79 | 3761 | 18.0 | 81.1 (79.2–82.8) | 51.6 (47.7–55.9) | 50.6 (48.0–53.3) | 85.5 (84.0–86.9) | 56.6 (53.0–60.9) | 54.4 (51.6–57.0) |
| 80+ | 4349 | 17.2 | 77.9 (76.1–79.7) | 42.9 (39.5–46.8) | 43.5 (40.9–45.9) | 83.5 (82.1–85.1) | 49.2 (45.6–53.2) | 49.6 (47.2–52.1) |
Table B.2: Subgroup performance of ECGPD-LEF and the Columbia mini model in the external test set.
| Grouped | Original Categories |
| Asian | ASIAN, ASIAN - ASIAN INDIAN, ASIAN - CHINESE, ASIAN - KOREAN, ASIAN - SOUTH EAST ASIAN |
| Black | BLACK/AFRICAN, BLACK/AFRICAN AMERICAN, BLACK/CAPE VERDEAN, BLACK/CARIBBEAN ISLAND |
| Hispanic | HISPANIC OR LATINO, CENTRAL AMERICAN, COLUMBIAN, CUBAN, DOMINICAN, GUATEMALAN, HONDURAN, MEXICAN, PUERTO RICAN, SALVADORAN, SOUTH AMERICAN |
| White | WHITE, WHITE - BRAZILIAN, EASTERN EUROPEAN, OTHER EUROPEAN, RUSSIAN, PORTUGUESE |
| Other | OTHER, AMERICAN INDIAN/ALASKA NATIVE, MULTIPLE RACE/ETHNICITY, PACIFIC ISLANDER |
| Unknown | UNKNOWN, UNABLE TO OBTAIN, DECLINED |
Table B.3: Mapping of original race/ethnicity categories to aggregated subgroups.
| Predictor | Traditional ECG Model | LEF Detection (One Predictor) | ||||||
|---|---|---|---|---|---|---|---|---|
| AUROC | AUPRC | F1 Score | Thresh | AUROC | AUPRC | F1 Score | Thresh | |
| LAFB | 98.8 (98.3–99.2) | 87.0 (82.1–91.2) | 78.9 (74.2–83.5) | 0.274 | 70.0 (68.2–71.7) | 29.1 (26.9–31.6) | 39.0 (36.8–41.1) | 0.001226 |
| ALMI | 97.3 (95.0–99.1) | 61.3 (43.6–77.1) | 58.3 (40.0–73.2) | 0.257 | 68.6 (66.8–70.6) | 35.0 (32.3–38.3) | 38.6 (36.1–41.1) | 0.002106 |
| ABQRS | 87.5 (85.5–89.5) | 57.4 (51.7–63.5) | 54.4 (50.5–58.2) | 0.053 | 68.3 (66.5–70.1) | 29.8 (27.5–32.5) | 38.0 (35.9–40.1) | 0.012123 |
| CLBBB | 99.8 (99.6–100.0) | 93.0 (85.9–98.4) | 89.3 (82.8–94.7) | 0.053 | 67.9 (66.0–69.7) | 35.4 (32.5–38.8) | 38.0 (35.5–40.2) | 0.000006 |
| ILMI | 95.7 (91.9–98.5) | 63.4 (50.6–75.7) | 61.1 (48.9–71.2) | 0.157 | 67.6 (65.7–69.4) | 31.7 (29.0–34.8) | 37.9 (35.9–40.1) | 0.000172 |
Table C.1: Performance of traditional AI-ECG predictors and LEF detection using a single-predictor approach (remaining 61 predictors)
| Predictor | LEF Detection (One Predictor) | |||
|---|---|---|---|---|
| AUROC | AUPRC | F1 Score | Thresh | |
| NORM | 78.8 (77.8-79.6) | 36.2 (34.5-38.1) | 42.3 (41.0-43.5) | 0.003641 |
| ILBBB | 71.6 (70.5-72.6) | 33.2 (31.5-35.0) | 38.2 (36.8-39.6) | 0.000349 |
| INJAL | 73.4 (72.4-74.3) | 27.9 (26.6-29.2) | 40.9 (39.6-42.1) | 0.000386 |
| ISCLA | 67.7 (66.6-68.8) | 26.7 (25.4-28.4) | 34.0 (32.6-35.2) | 0.001918 |
| ANEUR | 71.0 (70.0-72.0) | 32.1 (30.4-33.8) | 38.0 (36.7-39.4) | 0.001924 |
Table C.2: Performance of LEF detection using the single-predictor approach on the external test cohort (ECG-Note) for all 71 ECG predictors.
Figure C.1: Operating thresholds of individual predictors for LEF detection. a , Left column: thresholds for the first half of predictors (AUROC 60%). b , Right column: thresholds for the remaining predictors. For each predictor, three types of thresholds are shown: 1) Diagnosis threshold : threshold for determining whether the ECG shows a positive signal for that predictor; 2) F1-max threshold : threshold for predicting LEF positivity using that predictor, chosen to maximize F1 score on the validation set; 3) Recall-based threshold : threshold for predicting LEF positivity using that predictor, set to achieve recall 90%. Only predictors with AUROC 60% are shown for clarity; complete results for all predictors are provided in Tables 2 and C.1 .
| NORM | ILBBB | |||||||
| Subgroup | n | Prevalence (%) | AUROC | AUPRC | F1 Score | AUROC | AUPRC | F1 Score |
| Age groups | ||||||||
| 18–59 | 2124 | 13.6 | 80.9 (78.2–83.8) | 41.3 (36.0–48.3) | 47.1 (42.6–51.6) | 79.3 (76.2–82.1) | 44.1 (38.7–49.9) | 44.4 (39.6–49.3) |
| 60–69 | 1318 | 16.6 | 81.6 (78.3–84.8) | 49.3 (42.2–56.6) | 50.4 (45.5–55.0) | 78.1 (74.7–81.1) | 44.8 (37.9–51.4) | 45.8 (40.1–51.0) |
| 70–79 | 1154 | 21.7 | 80.6 (77.6–83.5) | 53.6 (47.7–60.5) | 55.4 (51.0–59.9) | 81.1 (78.0–84.0) | 55.7 (49.5–61.8) | 58.5 (53.4–62.9) |
| 80+ | 846 | 24.1 | 77.0 (73.4–80.2) | 47.9 (41.9–54.8) | 52.8 (48.2–57.1) | 77.8 (74.5–81.3) | 50.7 (44.8–58.2) | 53.8 (48.6–58.9) |
Table C.3: Subgroup performance of the single-predictor approach for the NORM and ILBBB predictors on the internal test set.
| INJIL | ISCLA | |||||||
| Subgroup | n | Prevalence (%) | AUROC | AUPRC | F1 Score | AUROC | AUPRC | F1 Score |
| Age groups | ||||||||
| 18–59 | 2124 | 13.6 | 77.4 (74.6–80.2) | 32.4 (28.4–37.9) | 40.8 (36.6–45.1) | 79.0 (76.4–81.8) | 36.6 (31.8–43.0) | 43.3 (39.3–47.4) |
| 60–69 | 1318 | 16.6 | 75.4 (72.0–78.8) | 34.0 (29.0–40.1) | 45.5 (40.6–50.2) | 74.4 (70.9–78.0) | 37.0 (30.8–44.3) | 43.0 (38.1–48.0) |
| 70–79 | 1154 | 21.7 | 73.9 (70.6–77.1) | 39.5 (34.5–45.6) | 48.0 (43.1–52.7) | 73.9 (70.6–77.2) | 42.5 (36.9–49.3) | 45.8 (40.9–50.1) |
| 80+ | 846 | 24.1 | 69.2 (65.1–72.9) | 36.3 (31.2–42.8) | 47.5 (42.3–52.5) | 70.8 (67.2–74.5) | 43.3 (37.2–50.1) | 43.8 (38.5–48.8) |
Table C.4: Subgroup performance of the single-predictor method for INJIL and ISCLA in the internal test set.
| NORM | ILBBB | |||||||
| Subgroup | n | Prevalence (%) | AUROC | AUPRC | F1 Score | AUROC | AUPRC | F1 Score |
| Age groups | ||||||||
| 18–59 | 4270 | 12.2 | 81.5 (79.4–83.3) | 37.1 (33.1–41.5) | 44.6 (41.6–47.7) | 73.7 (71.3–76.0) | 33.7 (30.3–38.0) | 40.4 (37.7–43.2) |
| 60–69 | 3637 | 15.7 | 78.5 (76.5–80.3) | 40.8 (36.9–45.3) | 46.2 (43.4–48.9) | 71.6 (69.3–73.9) | 37.6 (33.9–41.8) | 43.4 (40.5–46.1) |
| 70–79 | 3761 | 18.0 | 77.7 (76.1–79.5) | 39.3 (36.1–43.3) | 43.7 (41.3–46.0) | 70.5 (68.4–72.7) | 36.0 (33.1–39.7) | 39.9 (37.4–42.5) |
| 80+ | 4349 | 17.2 | 74.9 (73.1–76.7) | 34.2 (31.3–37.6) | 38.4 (36.4–40.4) | 67.0 (64.9–69.2) | 29.3 (26.8–32.4) | 36.1 (33.9–38.5) |
Table C.5: Subgroup performance of the single-predictor method for NORM and ILBBB in the external test set.
| INJIL | ISCLA | |||||||
| Subgroup | n | Prevalence (%) | AUROC | AUPRC | F1 Score | AUROC | AUPRC | F1 Score |
| Age groups | ||||||||
| 18–59 | 4270 | 12.2 | 73.1 (71.0–75.1) | 25.7 (22.4–29.6) | 42.2 (39.1–45.5) | 61.0 (58.6–63.4) | 23.2 (20.3–26.6) | 32.4 (30.0–34.8) |
| 60–69 | 3637 | 15.7 | 72.4 (70.4–74.5) | 28.9 (25.6–32.8) | 45.0 (42.4–47.6) | 61.3 (59.0–63.6) | 26.2 (23.4–29.6) | 35.6 (33.2–38.1) |
| 70–79 | 3761 | 18.0 | 71.6 (69.8–73.5) | 29.4 (26.9–32.4) | 43.5 (41.1–45.9) | 62.7 (60.6–64.9) | 25.9 (23.6–29.1) | 33.2 (30.7–35.5) |
| 80+ | 4349 | 17.2 | 69.9 (68.0–71.7) | 27.3 (25.1–30.2) | 39.2 (36.9–41.3) | 62.2 (60.2–64.4) | 24.7 (22.7–27.4) | 32.0 (29.8–34.3) |
Table C.6: Subgroup performance of the single-predictor method for INJIL and ISCLA in the internal test set.