Abstract
Extensive clinical evidence suggests that preventive screening for coronary heart disease (CHD) at an earlier stage can significantly reduce mortality rates. This study presents a machine learning approach that integrates 64 two-dimensional speckle tracking echocardiography (2D-STE) features and seven clinical risk factors to predict CHD. By developing a generalized two-step stacking method, we improved diagnostic accuracy from approximately 70% to 87.7% on the testing set. The proposed method achieved a sensitivity of 0.903, a specificity of 0.843, and an AUC of 0.904, significantly outperforming individual classification models. This work establishes a foundation for deploying non-invasive, STE-based screening tools in clinical settings.
Introduction
Coronary heart disease (CHD) remains a leading cause of global mortality. While early detection is critical for improving patient outcomes, the current "gold standards" for diagnosis—computed tomography angiography (CTA) and coronary angiography (CA)—present significant barriers. These methods are highly accurate but are also expensive, invasive, and require specialized hospital infrastructure that may not be accessible to all populations.
Two-dimensional speckle tracking echocardiography (2D-STE) offers a promising, non-invasive alternative. By tracking the motion of heart tissue, 2D-STE can identify subtle functional abnormalities. However, the resulting data is often heterogeneous and "noisy," making it difficult for individual clinical indices or simple machine learning models to achieve the precision required for clinical decision-making. This research addresses this gap by introducing a novel ensemble machine learning framework designed to extract high-precision diagnostic value from echocardiographic and clinical data.
Methodology: The Two-Step Stacking Approach
The core innovation of this study is a two-step stacking ensemble approach. In traditional machine learning, "stacking" involves training multiple base learners and using a meta-model to combine their predictions. This study generalizes that concept into a hierarchical process to better handle the complexities of medical data.
Feature Engineering
The model utilizes 71 total predictors:
- 64 Strain-based Features: Derived from 2D-STE, these numerical features measure how various heart muscle segments stretch and contract. Principal Component Analysis (PCA) was employed to condense these complex heart-segment data points into manageable dimensions.
- 7 Clinical Factors: These include essential patient metrics such as age, smoking status, and the presence of diabetes.
The Two-Step Framework
- Step 1 (Initial Ensemble): The researchers trained 14 different types of classifiers, ranging from Logistic Regression to Neural Networks. These diverse models were aggregated into a "first-step" stack to synthesize their individual predictive strengths.
- Step 2 (Noise Reduction): To combat the inherent noise in clinical data, the first step was repeated 10 times using different random partitions of the training data. These 10 sets of results were then stacked again using a Random Forest algorithm to determine the final weights for the prediction.
This hierarchical structure acts as a sophisticated filter, amplifying the strengths of diverse models while neutralizing the errors of individual poor-performing ones.
Dataset and Clinical Trial
The study utilized data from a retrospective clinical trial (NCT03905200) involving 424 patients at Beijing Hospital. The cohort was divided into:
- Case Group: 217 CHD-positive patients, confirmed via angiography with ≥ 50% stenosis.
- Control Group: 207 CHD-negative patients.
The data was partitioned into a training set (288), a validation set (72) for tuning the stacking weights, and a dedicated testing set (64) to evaluate final performance.
Key Insights and Results
Performance Metrics
The two-step stacking method demonstrated a substantial leap in diagnostic performance compared to both individual models and traditional stacking techniques:
- Accuracy: Reached 87.7%, a significant improvement over the ~70% accuracy of individual models and the 81.8% achieved by traditional one-step stacking.
- AUC (Area Under the Curve): Achieved 0.904, indicating excellent diagnostic power.
- Sensitivity and Specificity: The model reached a sensitivity of 0.903 and a specificity of 0.843.
For comparison, using only the standard Global Longitudinal Peak Strain (GLPS) index—a common clinical metric—resulted in an accuracy of only 63.3%. This confirms that an ensemble "big data" approach is necessary to capture the subtle patterns indicative of CHD.
Clinical and Physiological Findings
- Diversity Over Individual Strength: The study found that a "committee" of diverse models outperforms even the best single model. An ensemble of 14 models significantly outperformed a 3-model ensemble consisting only of the top performers.
- The Importance of Longitudinal Strain: Statistical analysis confirmed that longitudinal strain (the heart's contraction from top to bottom) is a more sensitive indicator of CHD than radial strain. This aligns with heart physiology, as sub-endocardial fibers—the most sensitive to blood flow reductions—are oriented longitudinally.
Limitations and Future Directions
While the results are promising, several hurdles remain for widespread clinical deployment. The study was conducted at a single center, meaning the models may require recalibration to account for inter-vendor variability in ultrasound hardware. Additionally, current speckle tracking requires manual adjustment of "regions of interest," which can introduce subjective bias. The model also faces challenges with low-quality images where heart borders are poorly defined.
Future work is focused on developing automated image quality control and AI-driven heart-tracing techniques. These advancements aim to remove the "human in the loop," making the screening process entirely objective, real-time, and ready for broad clinical application.
Summary updated