Source-linked AI summary
An Efficient Convolutional Neural Network for Coronary Heart Disease Prediction
Aniruddha Dutta, Tamal Batabyal, Meheli Basu, Scott T. Acton
TL;DR
CHD prediction remains challenging in highly imbalanced clinical data, particularly when maintaining accurate detection of the minority CHD class. The paper proposes a shallow convolutional network with LASSO-based feature selection and a simulated annealing-like training routine, reporting correlations among selected clinical variables and a resilience claim for the architecture. Limitations include the lack of a theoretical explanation for the network’s resistance to imbalance and the need to test whether LASSO reflects true variable importance.
Problem
The study addresses CHD prediction in clinical data where severe class imbalance makes accurate minority-class detection difficult.
Method
The paper combines LASSO feature selection and voting with homogenization through a fully connected layer before shallow convolutional stages and uses a simulated annealing-like training routine.
Results
0.79 correlation was found between glycohemoglobin and glucose, while 0.77 was found between serum ALT and AST and 0.89 between BMI and a related variable.
Takeaways & Limitations
The authors report that a special-purpose shallow convolutional network exhibits resilience to class imbalance for CHD prediction in the NHANES data.
Takeaways & Limitations
The study does not theoretically explain how the network resists data imbalance, and it identifies testing whether LASSO reflects true variable importance as future work.
Abstract
from arXiv · showhide
This study proposes an efficient neural network with convolutional layers to classify significantly class-imbalanced clinical data. The data are curated from the National Health and Nutritional Examination Survey (NHANES) with the goal of predicting the occurrence of Coronary Heart Disease (CHD). While the majority of the existing machine learning models that have been used on this class of data are vulnerable to class imbalance even after the adjustment of class-specific weights, our simple two-layer CNN exhibits resilience to the imbalance with fair harmony in class-specific performance. In order to obtain significant improvement in classification accuracy under supervised learning settings, it is a common practice to train a neural network architecture with a massive data and thereafter, test the resulting network on a comparatively smaller amount of data. However, given a highly imbalanced dataset, it is often challenging to achieve a high class 1 (true CHD prediction rate) accuracy as the testing data size increases. We adopt a two-step approach: first, we employ least absolute shrinkage and selection operator (LASSO) based feature weight assessment followed by majority-voting based identification of important features. Next, the important features are homogenized by using a fully connected layer, a crucial step before passing the output of the layer to successive convolutional stages. We also propose a training routine per epoch, akin to a simulated annealing process, to boost the classification accuracy. Despite a 35:1 (Non-CHD:CHD) ratio in the NHANES dataset, the investigation confirms that our proposed CNN architecture has the classification power of 77% to correctly classify the presence of CHD and 81.8% the absence of CHD cases on a testing data, which is 85.70% of the total dataset. ( (<1920 characters)Please check the paper for full abstract)
1. Introduction
The introduction motivates CHD prediction from clinical data and proposes a shallow convolutional neural network designed for imbalanced datasets. The approach combines LASSO-based feature selection, feature voting, and a simulated annealing-like training schedule.
- Heart disease is a leading cause of death, and timely diagnosis is described as crucial for reducing health risk.
- Prior studies used statistical, machine learning, neural-network, and deep-learning methods for cardiovascular prediction, but often involved only a few thousand patients and limited risk factors.
- The study proposes an efficient convolutional neural network using NHANES clinical, laboratory, and examination data to predict CHD occurrence.
- LASSO regression, feature voting, and elimination are used for preprocessing before the network is trained and compared with traditional machine learning models.
- The proposed shallow convolutional architecture is intended to address scarce class-specific data and improve CHD prediction in unbalanced datasets.
- A simulated annealing-like training schedule is proposed to minimize the generalization error between training and testing losses.
2. Data Preprocessing
The study compiles NHANES data from 1999–2000 through 2015–2016 across demographic, examination, laboratory, and questionnaire sources. The resulting predictors include continuous and categorical variables, with selected redundant or uncorrelated variables excluded before analysis.
- The dataset combines demographic, examination, laboratory, and questionnaire data from NHANES survey cycles spanning 1999–2000 through 2015–2016.
- Figure 1 organizes the data into Demography, Examination, and Laboratory categories containing quantitative and qualitative variables.
- The feature set includes demographic, examination, laboratory, and questionnaire variables such as age, gender, BMI, blood pressure, cholesterol, glucose, and diabetes indicators.
- Thirty continuous and six categorical independent variables are used to predict the likelihood of CHD.
- CHD is treated as a dichotomous dependent variable, with awareness defined by a yes response to whether participants had been told they had CHD.
- Variables judged linearly dependent or uncorrelated, including annual family income, height, and several laboratory measures, are excluded from subsequent processing and analysis.
3. Proposed Architecture
The proposed architecture combines LASSO-based feature selection and majority voting with a shallow convolutional neural network for mixed-type, imbalanced clinical classification. Selected variables are homogenized through fully connected layers before convolutional representation learning and Softmax classification.
- LASSO Shrinkage and Majority Voting: LASSO regression performs variable selection by shrinking some coefficients to zero through an absolute-value penalty controlled by λ.Variables with zero γ are treated as unimportant; increasing λ increases shrinkage and eliminates more coefficients.
- LASSO Shrinkage and Majority Voting: Repeated LASSO fits on balanced random subsamples use majority voting to identify variables selected as nonzero across iterations.The procedure counts nonzero coefficients over N subsampled datasets and applies a manually set selection threshold.
- Motivation and Design: The model addresses severely imbalanced binary clinical classification where augmentation may be infeasible and misclassification carries significant medical risk.The paper specifically targets prediction of subjects with and without CHD under these constraints.
- CNN Architecture: The proposed network is a shallow feedforward CNN with convolution layers sandwiched between fully connected layers.The architecture uses a sequential one-input-one-output design and applies convolutional representation learning after the initial fully connected block.
- CNN Architecture: The input contains 50 active phenotypes, and 64-neuron dense layers linearly combine and homogenize the variables before nonlinear transformation.The network then applies ReLU, dropout, convolution, pooling, further dense layers, and Softmax classification.
- CNN Architecture: The final dense layers before Softmax can be retrained for transfer learning when new data become available.The architecture description also identifies dropout probability, neuron counts, activation functions, and convolution parameters as training choices.
- Motivation and Design: Synthetic augmentation is treated cautiously because biologically generated measurements may be implausible or difficult to authenticate.The paper notes that augmented biological variables may not correspond to possible subject readings.
4.1 Summary Statistics
The selected predictor variables are examined for pairwise correlation after LASSO-based processing. Several variable pairs remain moderately or highly correlated, while most pairs are described as fairly weakly correlated.
- Correlation Analysis: 30 continuous predictor variables are investigated for correlations before further variable selection for CHD classification.The analysis uses correlation assessment alongside LASSO regression to determine predictors for subsequent modeling.
- Correlated Predictors: 0.77 is the reported correlation between serum ALT and AST, while AST is retained as a major CHD risk factor.The cited discussion notes higher AST levels in CHD patients than controls and their potential use as biochemical markers.
- Correlated Predictors: 0.89 is the reported correlation between body-mass-index and weight, and 0.74 between hemoglobin and red blood cells.The paper characterizes the body-mass-index–weight association as expected and notes limited evidence linking hemoglobin with clinically recognized CHD.
- Correlated Predictors: 0.79 is the reported correlation between glycohemoglobin and glucose.The paper connects these variables with prior evidence concerning elevated blood glucose and CHD risk in non-diabetic patients.
- Correlated Predictors: 0.41 is the reported correlation between lactate dehydrogenase and AST.The discussion relates increased LDH in active populations to low CHD risk based on previous studies.
- Correlation Analysis: The correlation table identifies glucose–glycohemoglobin, red blood cells–hemoglobin, ALT–AST, and weight–body-mass-index as the principal correlated pairs.The remaining pairs show fairly low correlation values, which the paper interprets as sufficient decorrelation after LASSO-majority voting.
4.2 Model results
The CNN achieves high overall and class-specific performance through LASSO-based feature selection, subsampling, and class-weight tuning on imbalanced CHD data. The reported results indicate balanced generalization across CHD and Non-CHD cases.
- A majority-voting threshold of 16.67–20 appears suitable for balanced per-class accuracy across 100 LASSO instances.
- LASSO reduces the coefficients of BMI, glucose, ALT, and one categorical variable to zero.
- 83.51% maximum training accuracy and 0.489 minimum training loss occur with a 1300:4000 CHD:Non-CHD sampling ratio and 3:1 misclassification penalty.
- 82.32% test accuracy is obtained on 31,779 remaining samples, representing 85.70% of the whole dataset.
- 77% CHD accuracy and 81% Non-CHD accuracy are maintained on the large testing cohort.
- The model reports an AUC of 0.767, a positive likelihood ratio of 4.27, and a negative likelihood ratio of 0.27.
- 79.5% balanced accuracy is reported for the CNN classifier.
4.3. Comparison of ML models
Compared with conventional classifiers and sampling strategies, the proposed LASSO-CNN combines competitive recall with higher specificity, test accuracy, and balanced class-wise performance. Its advantage is most evident when evaluation retains a large, highly imbalanced test cohort.
- Logistic regression and AdaBoost achieve the highest test accuracies but have low recall values for CHD detection.
- 79.5% balanced accuracy is reported for the CNN, exceeding the individual accuracies of SVM and random forest classifiers.
- 0.77 recall, 0.81 specificity, and 0.82 test accuracy are reported for the CNN in model comparisons.
- 79.42% vanilla-CNN test accuracy is approximately 2% below the average reported for LASSO(6)-CNN, corresponding to about 635 additional correctly labeled samples.
- 79%, 79.14%, and 75.5% are the ADASYN test, Non-CHD, and CHD accuracies, with class score variance of 3.06.
- 74.59% overall test accuracy accompanies 74% Non-CHD and 87% CHD accuracy for near-miss version 3.
- Undersampling can reduce majority-class accuracy because the reduced data may insufficiently capture population variation and generalize to test data.
4.4. Validation on stroke data
The CNN is also evaluated on an NHANES stroke dataset containing mixed-type measurements and substantial sample size. It correctly labels almost 80% of the test cases while other models trade accuracy against class-balance differences.
- The stroke dataset contains 37,177 subjects and 36 mixed-type measurement variables.
- Almost 80% of test cases are correctly labeled by the proposed CNN.
- 74.44% and 74.74% overall test accuracies are reported for random forest and SVM, respectively.
4.5. Notes on the resilience to data imbalance
The experiments examine resilience to severe imbalance under constraints that make common sampling approaches infeasible and misclassification costly. Convolutional layers reduce class-wise accuracy disparities and lessen training destabilization relative to dense architectures.
- The study targets severely imbalanced binary mixed-type data where augmentation methods are infeasible and misclassification risks are significant.
- 2.04%, 2.4%, and 0.07% class-wise accuracy differences are obtained by Conv-I, Conv-II, and Conv-III, respectively.
- 27.14% class-wise accuracy difference occurs for MLP-I, while MLP-II reduces it to 5.73% with substantially more parameters and overfitting risk.
- Multiple convolution layers diminish the degeneration in which minority-class accuracy plummets during scheduled training.
5. Conclusion, Limitations and Future Research
The paper proposes a modular, shallow convolutional pipeline for CHD prediction under severe class imbalance, combining repeated LASSO-based selection, majority voting, and convolutional modeling. On NHANES data, it reports balanced class-wise performance, while identifying limitations and future extensions involving feature selection, dietary variables, and transfer learning.
- Conclusion: The proposed model is simple, modular, and moderately resilient to data imbalance compared with commonly used classifiers.The paper specifically reports balanced class-wise classification in contrast to Logistic Regression and Adaboost.
- Method: Repeated LASSO on randomly subsampled data followed by majority voting selects significant variables and reduces dimensionality before CNN classification.The procedure is intended to control false-negative effects in variable selection and filter highly correlated or unimportant variables.
- Scope and rationale: The approach targets clinical data with mixed qualitative and quantitative attributes, severe imbalance, infeasible augmentation, and substantial misclassification risk.The authors avoid conventional linear or nonlinear dimension-reduction methods when their derived combinations reduce interpretability.
- Limitations: LASSO selection remains limited by its linear estimator, possible uncertainty about whether shrinkage reflects true variable importance, and insufficient representation of the majority class during 1:1 subsampling.The paper also notes that densely interspersed, non-convex classes can make synthetic augmentation biologically implausible or incorrectly labeled.
- Future research: Future work includes retaining LASSO-voting weights as CNN priors, adding NHANES dietary variables, and retraining dense layers for transfer learning.Dietary predictors may introduce endogeneity because clinical variables can themselves be affected by dietary habits.