Comparative Study of Graph Neural Networks and Traditional Models in Disease Prediction

Abstract

In recent years, the emergence of graph neural networks (GNNs) has significantly advanced the field of disease prediction by leveraging the inherent graph structure present in biomedical data. This study provides a comprehensive comparative analysis of GNNs and traditional machine learning models in the context of disease prediction, focusing on their efficacy, scalability, and interpretability. Traditional models, such as logistic regression and support vector machines, have been the cornerstone of predictive analytics, often relying on feature engineering and linear or kernel-based transformations. However, these models may struggle to capture the complex, non-linear relationships intrinsic to biological networks.

Graph neural networks, on the other hand, offer a robust framework for integrating topological and node feature information, providing a more holistic representation of biological systems. This paper employs a series of experiments across multiple disease datasets, evaluating the predictive performance of both GNNs and traditional models. Key metrics, including accuracy, precision, recall, and F1-score, are utilized to benchmark the models. Furthermore, the scalability of each model is assessed by analyzing computational efficiency and memory utilization, especially in large-scale data scenarios.

Our findings indicate that GNNs consistently outperform traditional models in terms of predictive accuracy and robustness, particularly in datasets characterized by intricate interdependencies among biological entities. Moreover, GNNs demonstrate superior scalability, making them well-suited for large datasets typical in genomics and epidemiology. However, traditional models still hold value in scenarios where interpretability and simplicity are prioritized, as GNNs often require complex architectures and substantial computational resources.

In conclusion, this study underscores the potential of graph neural networks to revolutionize disease prediction, while also acknowledging the enduring relevance of traditional models. This dual perspective offers valuable insights for researchers and practitioners seeking to enhance predictive models in biomedical applications.