CoxFNN: Interpretable machine learning method for survival analysis

Survival analysis plays a pivotal role in healthcare, particularly in analyzing time-to-event data such as in disease progression, treatment efficacy, and drug development. Traditional methods in survival analysis often face a trade-off: they either make linear assumptions, which are interpretable but may be overly simplistic, or they capture complex, non-linear relationships but lack clarity and ease of understanding. To overcome these challenges, we develop a novel machine-learning approach. This method is an extension of the Cox proportional hazards model. Unlike its traditional version, our method can model non-linear relationships between various features and the associated risks without relying on specific distribution assumptions. More importantly, it also adds the ability to learn rules from the dataset that are comprehensible to humans. Our proposed model demonstrated comparable performance in our experimental evaluations relative to other existing survival analysis models while also successfully identifying potential high-risk factors and clinical rules, demonstrating its practical utility in real-world healthcare settings. This balance of sophisticated analysis and interpretability makes it promising in the field of survival analysis.