Zeineb Bouzid

Aims: AI in electrocardiography (ECG) has diverged into two paths: traditional signal processing with machine learning, and deep learning of raw waveforms. The former preserves physiological interpretability but may miss novel patterns, while the latter overlooks a century of ECG science. No framework yet integrates both. This study introduces a fusion model combining handcrafted feature-based random forest and waveform-based convolutional neural network to improve the diagnosis of acute coronary events. Methods and results: This prospective, multicentre cohort study included pre-hospital ECGs recorded on emergency medical services from patients with a chief complaint of chest pain. Using this dataset, we developed four fusion frameworks that integrate ECG features with waveform information: (1) decision-level fusion combining predictions from random forest and convolutional neural network (CNN); (2) decision-level fusion with retraining; (3) feature-level fusion by concatenating CNN embeddings with ECG features; and (4) feature-level fusion combining CNN embeddings with tree-level outputs from the random forest. This study comprised of 10 393 ECGs from 7397 patients. On both outcomes, acute coronary syndrome and occlusion myocardial infarction, the proposed decision fusion with retraining (Approach 2) achieved an AUROC of 0.878 (95% CI, 0.837–0.915) and 0.961 (95% CI, 0.934–0.983) on the test set and an average precision score of 0.749 (95% CI, 0.681–0.810) and 0.850 (95% CI, 0.763–0.918), respectively, outperforming all other proposed fusion and individual models. Conclusion: This study demonstrates the significance of fusion models that integrate complementary ECG representations to improve diagnostic accuracy in acute coronary events, highlighting a promising approach for real-time AI-augmented cardiovascular disease detection across diverse use cases.

Objective: Acute coronary syndrome (ACS) is a life-threatening condition requiring accurate diagnosis for better outcomes. However, variability in signs and symptoms among racial subgroups could cause disparities in diagnostic accuracy. In this study, we use machine learning models to diagnose ACS, focusing on mitigating disparities and ensuring fairness between Black and non-Black populations. Methods: We built on a state-of-the-art random forest classifier to compare three mitigation strategies. The first two approaches involved resampling or partitioning the data prior to training, while the third approach proposed an innovative framework called adversarial debiasing. To evaluate our model performance, we used the receiver operating characteristic (ROC) curve and an operating point at 80% specificity for clinical importance. Results: After mitigation with adversarial debiasing, the difference in sensitivities between the two subgroups decreased from 9.8% to 1.3%. Specifically, this approach achieved areas under the ROC of 0.810 and 0.817, and sensitivities of 70.1% and 71.4%, respectively for Black and non-Black subgroups. Conclusion: The proposed adversarial debiasing model outperformed the other two methods in both diagnostic accuracy and effectiveness in minimizing disparities. Significance: We expect this framework to achieve fair diagnostic models across diverse demographic populations globally and be generalizable to other outcomes.

Background and Aims: The importance of risk stratification in patients with chest pain extends beyond diagnosis and immediate treatment. This study sought to evaluate the prognostic value of electrocardiogram feature-based machine learning models to risk-stratify all-cause mortality in those with chest pain. Methods: This was a prospective observational cohort study of consecutive, non-traumatic patients with chest pain. All-cause death was ascertained from multiple sources, including the CDC National Death Index registry. Six machine learning models were trained for survival analysis using 73 morphological electrocardiogram features (80% training with 10-fold cross-validation and 20% testing), followed by a variational Bayesian Gaussian mixture model to define distinct risk groups. The resulting classification performance was compared against the HEART score. Results: The derivation cohort included 4015 patients (age 59 ± 16 years, 47% women). The mortality rate was 20.3% after a median follow-up period of 3.05 years (interquartile range 1.75–5.32). Extra Survival Trees outperformed other forecasting models, and the derived risk groups successfully classified patients into low-, moderate-, and high-risk groups (log-rank test statistic = 121.14, P < .001). This model outperformed the HEART score, reducing the rate of missed events by >90% with a negative predictive value and sensitivity of 93.4% and 85.9%, compared to 89.0% and 75.0%, respectively. In an independent external testing cohort (N = 3095, age 59 ± 15 years, 44% women, 30-day mortality 3.5%), patients in the moderate [odds ratio 3.62 (1.35–9.74)] and high [odds ratio 6.12 (2.38–15.75)] risk groups had significantly higher odds of mortality compared to those in the low-risk group. Conclusions: The externally validated machine learning-based model, exclusively utilizing features from the 12-lead electrocardiogram, outperformed the HEART score in stratifying the mortality risk of patients with acute chest pain. This may have the potential to impact the precision of care delivery and the allocation of resources to those at highest risk of adverse events.

Patients with occlusion myocardial infarction (OMI) and no ST-elevation on presenting electrocardiogram (ECG) are increasing in numbers. These patients have a poor prognosis and would benefit from immediate reperfusion therapy, but, currently, there are no accurate tools to identify them during initial triage. Here we report, to our knowledge, the first observational cohort study to develop machine learning models for the ECG diagnosis of OMI. Using 7,313 consecutive patients from multiple clinical sites, we derived and externally validated an intelligent model that outperformed practicing clinicians and other widely used commercial interpretation systems, substantially boosting both precision and sensitivity. Our derived OMI risk score provided enhanced rule-in and rule-out accuracy relevant to routine care, and, when combined with the clinical judgment of trained emergency personnel, it helped correctly reclassify one in three patients with chest pain. ECG features driving our models were validated by clinical experts, providing plausible mechanistic links to myocardial injury.

Patients with occlusion myocardial infarction (OMI) and no ST-elevation on presenting ECG are increasing in numbers. These patients have a poor prognosis and would benefit from immediate reperfusion therapy, but we currently have no accurate tools to identify them during initial triage. Herein, we report the first observational cohort study to develop machine learning models for the ECG diagnosis of OMI. Using 7,313 consecutive patients from multiple clinical sites, we derived and externally validated an intelligent model that outperformed practicing clinicians and other widely used commercial interpretation systems, significantly boosting both precision and sensitivity. Our derived OMI risk score provided superior rule-in and rule-out accuracy compared to routine care, and when combined with the clinical judgment of trained emergency personnel, this score helped correctly reclassify one in three patients with chest pain. ECG features driving our models were validated by clinical experts, providing plausible mechanistic links to myocardial injury.