Transfer learning for a tabular-to-image approach: A case study for cardiovascular disease prediction

Area of Science:

• Machine learning and artificial intelligence in healthcare
• Medical image analysis and computational pathology
• Data science and predictive modeling for disease risk assessment

Background:

• Machine learning (ML) models are widely used for tabular data classification.
• Transforming tabular data into images enables the use of powerful convolutional neural networks (CNNs).
• Existing tabular-to-image methods struggle with datasets having few samples and mixed data types.

Purpose of the Study:

• To evaluate the 'low mixed-image generator for tabular data' (LM-IGTD) method for converting tabular data into images.
• To assess the effectiveness of transfer learning and fine-tuning CNNs for improving predictions on tabular data.
• To advance cardiovascular disease (CVD) diagnosis through enhanced predictive modeling.

Main Methods:

• Two public CVD datasets (Framingham and Steno) were utilized.
• Tabular data were converted into images using the LM-IGTD method.
• CNN models were trained on the Framingham dataset and then fine-tuned on the Steno dataset using transfer learning to predict CVD risk.

Main Results:

• The CNN model with transfer learning achieved the highest Area Under the ROC Curve (AUCORC) of 0.855 on the Steno dataset.
• This approach demonstrated a 2% improvement in accuracy compared to the best traditional model, TabPFN.
• The CNN-based transfer learning method outperformed other ML models including decision trees, K-nearest neighbors, LASSO, and support vector machine.

Conclusions:

• This is the first study to evaluate transfer learning and fine-tuning on tabular data via tabular-to-image conversion.
• The proposed method leverages CNNs' predictive power for improved CVD risk prediction.
• This framework supports early clinical intervention and decision-making in CVD diagnosis.