Pipeline for Interpretable Survival Analysis (PISA) providing multiple complexity accuracy trade-off models

Accurate time-to-event prediction is central to clinical research, guiding prognosis, treatment decisions, and resource allocation. For survival analysis models to be relevant in healthcare, accuracy and interpretability are critical: predictions must be traceable to patient-specific characteristics, and risk factors should be identifiable to generate actionable insights for clinicians. Traditional survival models have limited accuracy as they often fail to capture non-linear interactions, while modern deep learning approaches, though powerful, are not interpretable. We propose a Pipeline for Interpretable Survival Analysis (PISA) that provides multiple survival analysis models that trade off complexity and performance. Using multiple-feature, multi-objective feature engineering, combined with a traditional survival analysis technique, PISA transforms patient characteristics and time-to-event data into multiple survival analysis models. Every model is converted into simple patient stratification flowcharts supported by Kaplan–Meier curves. While PISA is survival-technique agnostic, we illustrate its flexibility through applications of Cox regression and shallow survival trees. We apply PISA to predict overall survival in two clinical benchmark datasets concerning 1546 and 696 patients with node positive breast cancer (GBSG), 7098 hospitalised seriously ill patients (SUPPORT), and 1043 and 339 patients with symptomatic spinal bone metastases pertaining to a prior departmental study. On all datasets, PISA finds multiple survival models of varying complexity that outperform the survival analysis techniques when directly applied to the data. Further, PISA transforms all models into patient stratifications without compromising on performance. Revisiting a prior departmental study demonstrates PISA’s capacity to automate survival analysis workflows in real-world clinical research. PISA produces interpretable survival models and intuitive stratification flowcharts, whilst achieving state-of-the-art performance and providing valuable insights into survival prediction tasks. Predicting when a patient may experience an event, e.g. relapse or death, is important for guiding medical decisions and planning care. However, existing methods face a trade-off: simpler models are easier to understand but are often less accurate, while more advanced approaches are too complex to interpret. We propose a Pipeline for Interpretable Survival Analysis (PISA). PISA returns multiple prediction models that trade off complexity and accuracy. This allows clinicians and researchers to choose a model that best fits their needs. Each model is transformed into a simple flowchart that groups patients and shows their expected outcomes. Tested on clinical datasets, PISA finds interpretable prediction models, whilst retaining performance. PISA has the potential to markedly reduce, if not eliminate, the manual effort traditionally required to create accurate, interpretable survival prediction models for clinical practice. ED SUM: Schlender et al. propose a pipeline for interpretable survival analysis that provides multiple survival analysis models with varying complexity and accuracy trade-offs, as well as automatically generated patient stratifications per model. They show that these methods outperform traditional survival analysis models.