Data Domain Change and Feature Selection to Predict Cardiac Pathology with a 2D Clinical Dataset and Convolutional Neural Networks (Student Abstract)

Mario Tasso Ribeiro Serra Neto, Marco A. F. Mollinetti, Inês DutraAAAI 2021 · DOI

Abstract

This study presents the Map-Optimize-Learn (MOL) pipeline, a novel framework designed to transform tabular clinical data into synthetic images. By converting patient records into a visual format, the researchers leveraged the spatial pattern recognition capabilities of Convolutional Neural Networks (CNNs) to predict cardiac pathology in pediatric and adolescent patients. The MOL-CNN approach achieved a test accuracy of 92.5%, significantly outperforming traditional machine learning models such as Artificial Neural Networks (ANNs) and Logistic Regression (LR).

Introduction

Convolutional Neural Networks (CNNs) are the state-of-the-art for image recognition due to their ability to detect complex spatial hierarchies. However, a vast majority of critical medical data is stored in "2D" tabular formats (rows and columns). While standard neural networks can process tabular data, they fail to exploit the spatial relationships that CNNs utilize.

The primary challenge addressed in this research is the effective representation of non-image data as images. Simply arranging features into a grid is insufficient; for a CNN to learn effectively, the resulting "image" must possess a meaningful structure and consistent visual patterns across different samples. The authors propose a structured mapping and optimization strategy to bridge this gap.

Methodology: The Map-Optimize-Learn (MOL) Pipeline

The MOL pipeline consists of three distinct phases designed to prepare tabular data for deep learning:

  1. Map: Features are sorted based on statistical similarity. In this study, Euclidean distance was employed to ensure that related features are placed in close proximity within the synthetic image, mimicking the spatial correlation found in natural images.
  2. Optimize: Using Evolutionary Particle Swarm Optimization (EPSO), the researchers determined the optimal image dimensions (e.g., 6x1) and the most predictive subset of features. The optimization objective was to minimize the Mean Squared Error (MSE) between consecutive instances, creating a consistent "visual texture" for similar clinical cases.
  3. Learn: The optimized images are processed by a CNN. The architecture used in this study featured a single convolutional layer with 32 filters, followed by a max-pooling layer and a small hidden layer for final classification.

Experimental Setup and Data

The study utilized a real-world dataset from a specialized cardiovascular hospital in northeastern Brazil, consisting of 17,874 patient records. After data cleaning and the removal of missing values, 9,484 samples remained.

Key clinical features included Weight, Height, BMI, Age, Blood Pressure (PPA), Heart Murmur (HM), and History Emergency Level (HEL). Following the EPSO optimization phase, the model identified a 6x1 image configuration representing six critical features: Second heart sound (B2), History Emergency Level (HEL), Heart Murmur, Gender, Blood Pressure (PPA), and Visit Reason.

Key Insights and Visual Representation

A significant finding of the study was the visual consistency of the generated images. Instances without cardiac pathology (Outcome 0) tended to produce darker visual representations, whereas those with pathology (Outcome 1) appeared lighter and contained more distinct gradients. While these differences may be subtle to the human eye, the CNN successfully identifies these patterns as classification signals.

Figure 1: Comparison of synthetic images for non-pathological (Outcome 0) and pathological (Outcome 1) cases.

By sorting features by distance, the researchers ensured that the CNN's filters analyzed related data points in proximity, effectively creating a "spatial correlation" where none previously existed in the raw table.

Key Results

The MOL-CNN approach was evaluated against a standard Artificial Neural Network (ANN) and Logistic Regression (LR) using 10-fold cross-validation. The results demonstrate that the image-based representation captured nuances in the clinical data that traditional models overlooked.

Table 1: Performance comparison of the MOL-CNN approach against traditional models.
Algorithm Training Accuracy Test Accuracy
CNN (MOL images) 0.936 ± 0.007 0.925
ANN (Original data) 0.914 ± 0.007 0.904
LR (Original data) 0.894 ± 0.069 0.872

Conclusion and Future Work

The preliminary results of this student abstract suggest that data domain transformation is a viable path for improving clinical predictions. However, several avenues for future research remain:

  • Architectural Complexity: Testing deeper CNN architectures (e.g., ResNet) may further improve accuracy.
  • Mapping Techniques: Exploring alternative mapping strategies, such as t-SNE or manifold learning, could yield more informative visual structures.
  • Interpretability: Implementing visualization tools like Grad-CAM could help clinicians understand which specific clinical features in the "patient image" triggered a pathology alert, increasing trust in the model's decisions.

Summary updated

BibTeX
@inproceedings{neto2021data,
  author = {Mario Tasso Ribeiro Serra Neto and Marco A. F. Mollinetti and Inês Dutra},
  title = {{Data Domain Change and Feature Selection to Predict Cardiac Pathology with a 2D Clinical Dataset and Convolutional Neural Networks (Student Abstract)}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.1609/aaai.v35i18.17938},
}
Figure