CoxKAN: Kolmogorov-Arnold networks for interpretable, high-performance survival analysis

Area of Science:

• Statistics
• Machine Learning
• Bioinformatics

Background:

• Survival analysis is vital in medicine for predicting critical events like death or relapse.
• Traditional models balance interpretability and performance; deep learning excels in performance but lacks transparency.
• The need for interpretable yet high-performance models is critical in medical decision-making.

Purpose of the Study:

• Introduce CoxKAN, a Cox proportional hazards Kolmogorov-Arnold Network for interpretable survival analysis.
• Address the trade-off between performance and interpretability in existing survival models.
• Enable transparent and accurate risk prediction in clinical and genomics data.

Main Methods:

• Developed CoxKAN, integrating Kolmogorov-Arnold Networks with Cox proportional hazards models.
• Evaluated CoxKAN on synthetic datasets for formula recovery and feature selection.
• Tested CoxKAN on nine real-world datasets, including clinical and genomics data.

Main Results:

• CoxKAN accurately recovered interpretable hazard function formulae on synthetic data.
• Outperformed the traditional Cox proportional hazards model by up to 4% in C-index on real datasets.
• Matched or surpassed deep learning model performance while providing symbolic formulae and uncovering variable interactions.

Conclusions:

• CoxKAN provides interpretable, high-performance survival analysis.
• It offers clear insights into biomarker impact, crucial for medical applications.
• The model addresses the critical need for transparency in predictive medicine.