22 What’s next for machine learning survival analysis?
Thank you for reading our book. What remains is to reflect on what the future may hold for machine learning survival analysis.
It is tempting to point to currently popular topics in machine learning as the next frontier for survival analysis, such as large language models (LLMs). Indeed, recent work has explored the use of LLMs to extract data from Kaplan-Meier curves, to generate or interrogate survival plots for model checking (such as assessing the proportional hazards assumption), and to assist in the development of parametric survival models by first inferring underlying event-time distributions. Whilst these directions may be interesting, they do not address the more fundamental question of when machine learning methods are genuinely useful in survival analysis.
Survival analysis is often applied in high-stakes domains such as engineering (component reliability), finance (customer churn), and healthcare (patient survival or treatment effectiveness). In such settings, many practitioners continue to rely on relatively simple models, prioritizing inference and interpretability over predictive accuracy (e.g., Gorrod et al. 2019). This helps to explain why machine learning methods are not yet widely adopted in applied survival analysis. Moreover, even when prediction is the primary goal, it has been repeatedly shown that machine learning models can be out-performed by the Cox PH model in real-world settings (e.g., Burk et al. 2024; Zhang et al. 2021; Beaulac et al. 2020).
Machine learning approaches are most useful in settings where less complex methods struggle. One common advantage is the ability to model high-dimensional data or complex, highly non-linear relationships, where flexible learners can capture structure that is difficult to specify or estimate parametrically. A further advantage is the ability of machine learning methods to handle competing risks in a flexible and scalable manner, either directly through an algorithm’s ‘native’ methodology or via reductions. The Fine-Gray model is perhaps the most widely-used competing risks model; however, as discussed in this book, it has well-known limitations. In contrast, machine learning frameworks can often accommodate competing risks more naturally and with fewer modelling restrictions.
Finally, while machine learning methods may still be outperformed by classical models, even in high-dimensional settings (Spooner et al. 2020), this typically occurs when available features are weak or uninformative. When high-dimensional features are genuinely useful (such as being clinically meaningful in healthcare) then machine learning methods become essential for capturing complex patterns and for supporting large-scale inference through tools such as feature importance. In these settings, rather than merely offering alternative models, machine learning enables analyses that would otherwise be difficult or impractical to carry out with other statistical approaches.
As such, knowing how and when to apply machine learning to survival analysis is an important and valuable skill, provided these methods are used thoughtfully and with a clear understanding of their strengths and weaknesses; an understanding we hope this book has provided.