Kolmogorov-Arnold Network for Atherosclerotic Cardiovascular Disease Risk Prediction
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Background
Assessing the risk of future atherosclerotic cardiovascular disease (ASCVD) is crucial in clinical practice, yet it continues to pose significant challenges in cardiology. The Kolmogorov–Arnold Network (KAN) has recently been proposed as an alternative to traditional multi-layer perceptrons (MLP). This study aims to develop a 10-year ASCVD risk prediction model using KAN and to compare its performance with pre-existing and other machine learning-based techniques.
Methods
We utilized data from 2,116,621 individuals in the Korean National Health Insurance Service cohort for training and validating the models. Alongside KAN, we developed models using MLP, logistic regression, and random forest. Conventional equation models included PREVENT, PCE, Korean Risk Prediction Model, and SCORE2. We evaluated the models using the area under the receiver operating characteristic curve (AUROC) and the area under the precision–recall curve (AUPRC).
Findings
The KAN-based model achieved an AUROC of 0.7765 (95% confidence intervals, 0.7461-0.8068) and an AUPRC of 0.1551, outperforming both conventional equation-based models and other machine learning approaches. KAN offered interpretability by enabling the extraction of symbolic formulas and visualizing the contributions of individual risk factors through pruned network graphs. We further assessed the influence of risk factors on ASCVD prediction using SHAP analysis, ablation studies, and model output visualizations.
Interpretations
The KAN-based model demonstrated superior performance and enhanced interpretability in pre-dicting ASCVD risk. These findings suggest that KAN could be a promising alternative to other machine learning models in medical applications.
Funding
This research was supported by the Bio&Medical Technology Development Program of the National Research Foundation (NRF) funded by the Korean government (MSIT) (NO. RS-2023-00222910) (RS-2025-00517929).
Copyright
This preprint is made available under a CC BY-NC 4.0 license. Copyright remains with the author(s).
Research in context
Evidence before this study
Predicting 10-year atherosclerotic cardiovascular disease (ASCVD) risk is a clinically important task in cardiology. In recent years, numerous studies have shown that machine learning (ML) models outperform conventional equation-based risk prediction models such as the Pooled Cohort Equations (PCE), PREVENT, and SCORE2. However, widely used ML models such as multilayer perceptrons (MLPs) often lack interpretability and involve complex architectures, which makes them difficult to implement in clinical settings. To investigate whether the recently proposed Kolmogorov–Arnold Network (KAN) could offer a more interpretable machine learning–based alternative in 10-year ASCVD task, we conducted a comprehensive literature search up to April 2025 using PubMed, Google Scholar, Open-Review, and arXiv. Search terms included “atherosclerotic cardiovascular disease”, “risk prediction”, “Kolmogorov-Arnold Network”, “machine learning”, “deep learning”, “multilayer perceptrons”, “cardiovascular disease”, and “medical data”. Despite the growing interest in ML–based cardiovascular risk prediction, we found no studies, that applied KAN to ASCVD risk estimation. Existing studies using KAN in cardiology have focused on tasks such as synthetic ECG signal analysis or classification, but not long-term ASCVD risk prediction. This highlights a clear gap in the literature for interpretable machine learning models that can provide equation-like outputs in real-world population-based risk prediction.
Added value of this study
Although multilayer perceptrons (MLPs) have shown strong performance in predicting 10-year ASCVD risk, their lack of interpretability limits clinical adoption, where conventional models like PCE and PREVENT remain dominant. In this study, we propose a Kolmogorov–Arnold Network (KAN)–based model that generates simple and interpretable equations while maintaining high predictive accuracy. Using over 2.1 million individuals from the Korean National Health Insurance Service cohort with 10-year follow-up, the KAN model outperformed both traditional equation models and other ML approaches in AUROC and AUPRC. To our knowledge, this is the first study to apply KAN to long-term ASCVD risk prediction, bridging the gap between interpretability and machine learning performance in a clinically meaningful context.
Implications of all the available evidence
The growing body of evidence supports the use of ML models for predicting future ASCVD risk, yet lack of interpretability makes most models remain impractical for clinical use. Our findings demonstrate that interpretable ML approaches such as KAN can bridge this gap. By producing symbolic formulas with competitive performance, the model may enable clinicians to adopt ML-driven risk scores within familiar clinical workflows. Broader adoption of such approaches could support personalized prevention strategies and improved risk communication in cardiovascular care. We also hope that our study serves not only as a proposal of a novel predictive model, but also as a foundation that inspires future research on clinically applicable prediction models using KAN, further expanding its utility beyond theoretical development.
