TLDR: A new study introduces a non-invasive diagnostic model for coronary heart disease (CHD) using a multilayer perceptron (MLP) neural network. The model is trained on 50 key urinary peptide biomarkers selected via genetic algorithms. By balancing treatment and control groups with SMOTE and employing stratified validation, the MLP achieved high performance metrics: 95.67% precision, sensitivity, and specificity, an F1-score of 0.9565, and an AUC of 0.9748. This demonstrates a reliable and accurate non-invasive tool for CHD detection.
Coronary heart disease (CHD) remains a leading cause of death globally, placing a significant burden on healthcare systems. The need for accurate, non-invasive diagnostic methods is paramount to improve patient outcomes and quality of life. Traditional non-invasive diagnostic tools, such as echocardiography and CT angiography, while valuable, often come with limitations including variability in study quality, cost, and patient-specific factors like claustrophobia or incompatible medical devices. Furthermore, these methods can sometimes lead to false positives in low-risk populations, causing unnecessary anxiety and procedures.
In response to these challenges, researchers are increasingly turning to artificial intelligence (AI) and machine learning (ML) to enhance diagnostic capabilities. A recent study by Jesús Celis-Porras introduces a novel approach: a peptidomic-based prediction model for coronary heart disease using a multilayer perceptron (MLP) neural network. This model leverages the power of proteomics, the large-scale study of proteins, to identify specific biomarkers in urine samples that can indicate the presence of CHD.
Proteomics offers a detailed window into biological processes, allowing for the identification of proteins whose presence, absence, or alteration correlates with specific pathological states. By integrating this rich biomedical information with advanced machine learning techniques, particularly neural networks, the study aims to develop an automated, non-invasive diagnostic system.
How the Model Was Developed
The research involved a comprehensive dataset of peptides obtained from urine samples of patients across various international studies. The initial dataset included 82 cases of coronary artery disease and 345 control individuals. To address the imbalance between the number of cases and controls, a technique called Synthetic Minority Over-sampling Technique (SMOTE) was employed, which effectively increased the number of cases to match the controls, ensuring a balanced dataset for training.
A crucial step in developing the model was the selection of the most informative biomarkers. From a vast pool of 5,605 identified peptides, genetic algorithms (GA) were used to select 50 key urinary peptide biomarkers. Genetic algorithms are optimization metaheuristics inspired by natural selection, capable of efficiently exploring many combinations to find the most predictive features.
The core of the diagnostic tool is a multilayer perceptron (MLP) neural network. This network was designed with three hidden layers, each containing 60 neurons, and utilized hyperbolic tangent activation functions. To ensure the model’s robustness and prevent overfitting—a common issue where a model performs well on training data but poorly on new, unseen data—several strategies were implemented. These included stratified 10-fold cross-validation, which maintains class proportions in each training and test set, and early stopping, which halts training if the model’s performance on a validation set does not improve for a set number of epochs.
Impressive Performance and Reliability
The results of the study demonstrate the model’s exceptional performance. The MLP model achieved an overall accuracy of 95.67% in predicting coronary heart disease. Key metrics highlighting its reliability include a sensitivity of 95.67% (its ability to correctly identify positive cases) and a specificity of 95.67% (its ability to correctly identify negative cases). The F1-score, which balances precision and recall, was 0.9565.
Furthermore, the area under the ROC curve (AUC), a measure of the model’s ability to distinguish between classes, reached 0.9748 for both classes, indicating a high probability that the classifier ranks a positive instance above a negative one. The Matthews correlation coefficient (MCC) and Cohen’s kappa coefficient, robust measures of agreement between predicted and true labels that account for class imbalance, were 0.9134 and 0.9131, respectively. Values close to 1 for these metrics signify near-perfect alignment between predictions and actual diagnoses, ruling out random chance.
The learning curves during the training phase showed that the model learned efficiently and stably, with accuracy rapidly increasing and stabilizing above 90%, and loss steadily decreasing. The close alignment of training and validation curves confirmed minimal overfitting, indicating that the network learned the underlying patterns in the data rather than simply memorizing it.
Also Read:
- Adaptive ECG Anomaly Detection: A Framework for New Heart Rhythm Identification
- New AI Model Predicts Cardiovascular Biomarkers with Uncertainty During Pandemics
Future Implications
This research underscores the significant potential of combining urinary proteomics, genetic algorithm-based feature selection, and deep MLP networks to create a robust, non-invasive diagnostic approach for CHD. The high values across all evaluation metrics—AUC, MCC, and Kappa—demonstrate the model’s strong reliability and generalizability. This suggests that such a model could serve as a valuable complementary diagnostic tool in clinical settings, potentially reducing the reliance on more invasive procedures while maintaining high accuracy.
For more detailed information, you can read the full research paper here.
