[Paper Review] On Deep Neural Networks for Detecting Heart Disease

Nathalie-Sofia Tomov, Stanimire Tomov|arXiv (Cornell University)|Jan 1, 2018

Artificial Intelligence in Healthcare8 references20 citations

TL;DR

This paper proposes HEARO-5, a novel five-layer deep neural network (DNN) architecture optimized for detecting heart disease using routine clinical data. By applying hyperparameter tuning, L2 regularization, and robust handling of missing data and outliers, HEARO-5 achieves 99% accuracy and 0.98 Matthews correlation coefficient (MCC), significantly outperforming existing methods on the Cleveland heart disease dataset.

ABSTRACT

Heart disease is the leading cause of death, and experts estimate that approximately half of all heart attacks and strokes occur in people who have not been flagged as "at risk." Thus, there is an urgent need to improve the accuracy of heart disease diagnosis. To this end, we investigate the potential of using data analysis, and in particular the design and use of deep neural networks (DNNs) for detecting heart disease based on routine clinical data. Our main contribution is the design, evaluation, and optimization of DNN architectures of increasing depth for heart disease diagnosis. This work led to the discovery of a novel five layer DNN architecture - named Heart Evaluation for Algorithmic Risk-reduction and Optimization Five (HEARO-5) -- that yields best prediction accuracy. HEARO-5's design employs regularization optimization and automatically deals with missing data and/or data outliers. To evaluate and tune the architectures we use k-way cross-validation as well as Matthews correlation coefficient (MCC) to measure the quality of our classifications. The study is performed on the publicly available Cleveland dataset of medical information, and we are making our developments open source, to further facilitate openness and research on the use of DNNs in medicine. The HEARO-5 architecture, yielding 99% accuracy and 0.98 MCC, significantly outperforms currently published research in the area.

Motivation & Objective

• To improve the accuracy of heart disease diagnosis, especially for asymptomatic or misdiagnosed patients.
• To address the limitations of linear risk models like the AHA guidelines, which oversimplify complex, non-linear interactions among risk factors.
• To design and optimize deep neural network architectures that can learn complex, non-linear relationships in clinical data for improved diagnostic prediction.
• To develop a robust, open-source DNN framework (HEARO) that handles missing data and outliers effectively, enhancing generalization and clinical applicability.

Proposed method

• The study employs a flexible, fully connected DNN framework with variable depth (2 to 10 layers), optimized via exhaustive hyperparameter search.
• The architecture uses L2 regularization with hyperparameter α ∈ {0, 0.7, 1} to reduce overfitting and improve generalization.
• A comprehensive search space is defined over layer sizes (1–13 neurons), learning rates (λ ∈ {0.001, 0.01, 0.1}), L2 regularization strength (α), batch size (nb), and training epochs (6000).
• The model is trained and validated using k-way cross-validation and evaluated using Matthews correlation coefficient (MCC) for balanced performance assessment.
• The framework is implemented in Python with NumPy for linear algebra, enabling portability to optimized BLAS/GEMM backends.
• The final HEARO-5 architecture is selected based on empirical optimization, with the configuration [5, 9, 7, 5, 3, 1, 1, 1, 1, 1, 2, 0.01, 0.7, 200, 6000] yielding optimal results.

Experimental results

Research questions

• RQ1 Can deep neural networks significantly improve the accuracy of heart disease diagnosis beyond traditional linear risk models using routine clinical data?
• RQ2 How does regularization impact generalization and performance on test data in small, noisy medical datasets?
• RQ3 What is the optimal DNN architecture depth and width for heart disease detection on the Cleveland dataset?
• RQ4 Can hyperparameter tuning and robust handling of missing data/outliers lead to state-of-the-art performance in diagnostic prediction?
• RQ5 Does the proposed HEARO-5 model generalize well across different data splits, as confirmed by cross-validation and MCC?

Key findings

• HEARO-5 achieved 99% test accuracy and a Matthews correlation coefficient (MCC) of 0.98, indicating highly balanced and accurate classification.
• The model outperformed previously published methods, including a 76% accurate neural network and a 95.4% accurate random forest, by a significant margin.
• Regularization with α = 0.7 reduced overfitting, increasing test accuracy from 93% (unregularized) to 99%, demonstrating improved generalization.
• The unregularized model showed 99% training accuracy but only 93% test accuracy, indicating overfitting to noise and outliers.
• K-fold cross-validation confirmed that HEARO-5’s performance was stable and consistent, with cross-validated accuracy closely matching test accuracy.
• The HEARO-5 architecture effectively handled missing data and outliers through regularization, maintaining high performance without data imputation.