In a recent paper submitted to the arxiv* server, researchers presented a new R package called survex to explain the predictions of machine learning survival models. The software provides insights into model operations and decisions by applying explainable AI techniques. This transparency can promote trust and accountability in sensitive areas like healthcare, where survival models are being increasingly adopted.
*Important notice: arXiv publishes preliminary scientific reports that are not peer-reviewed and, therefore, should not be regarded as definitive, used to guide development decisions, or treated as established information in the field of artificial intelligence research.
Background
Survival analysis focuses on modeling time-to-event data subject to censoring. While traditional statistical approaches still prevail, the superior flexibility of machine learning survival models is driving integration. However, their black-box nature raises valid concerns over reliability and fairness.
Survex addresses this by generating visual explanations concerning variable effects and importance. The model-agnostic package supports diverse survival modeling frameworks in R, like ranger and survival. Tailored answers are offered for entire models and individual predictions. Unified interfaces enable evaluating performance and generating predictions across models.
The package implements methods like SHAP values adapted for survival functions to show variable attributions over time. Model components reveal predictors’ impacts on performance metrics. Diagnostics include residual analysis for detecting systematic errors. Interactive visualizations empower stakeholders to understand model behaviors thoroughly.
Applications span explaining biases in hospital length of stay prediction to analyzing influencers of cancer prognosis models. By promoting transparency in biomedical research and healthcare, Survex exemplifies the responsible application of machine learning for survival analysis. Ongoing developments include enhancing supported model types and explanation functionalities.
R offers statistical packages like Survival and machine learning packages like Ranger for Survival Analysis. Generic model explanation packages support regression and classification. But survival models warrant tailored descriptions handling time-dependencies. Survex fills this gap with a range of specifically designed answers.
Healthcare applications warrant particular scrutiny in deploying AI responsibly. Survival models are increasingly utilized for tasks like predicting disease risk and outcomes. Lack of transparency can undermine trust in model reliability. Explainable AI methods can address this by revealing the rationale behind predictions. However, bespoke solutions are needed for the unique structure of survival data. By explicitly tailoring explanations to survival models, Survex exemplifies the practical implementation of AI ethics.
survex Overview
The unifying component of survex is the explainer wrapper handling model prediction interfaces. For popular R survival packages, explainers are automatically created. Otherwise, prediction functions can be defined manually. The predict() method generates predictions as survival functions, risk scores, or hazards.
Explanations are categorized as global concerning the entire model or local to a particular prediction. Model_*() functions provide global insights like variable importance and effects. Predict_*() produces local explanations revealing the contributions of variables to a prediction. Interactive visualizations empower users to analyze models thoroughly.
The modular architecture enables conveniently adding new model types and explanations. By unifying explanation workflows across models, survex simplifies the model inspection and comparison process.
Key Explanation Types
model_parts() reveals predictors’ impacts on performance metrics through permutation importance. model_performance() supports evaluating models on time-dependent metrics like the Brier score.
model_diagnostics() enables residual analysis and ROC curves to detect systematic errors. model_profile() illustrates global variable effects, complemented by local insights from predict_profile().
model_survshap() and predict_parts() provide global and local variable attributions using adapted SHAP values. Alternatively, SurvLIME generates local explanations through Cox models.
Tailored visuals accompany each explanation, with extensive customization options. Collectively, these interactive insights empower stakeholders to inspect models from diverse angles.
Applications in Healthcare
Recent work demonstrates Survex’s value for responsible AI in healthcare. It offered insights into predictors’ relative importance in cancer prognosis models. Analyzing body composition-mortality relationships helped validate predictive signals.
Ongoing applications help explain biases in hospital length of stay forecasting, driving improvements. By promoting transparency in high-stakes biomedical research, survex enables trustworthy integration of machine learning. Furthermore, interactive visual explanations can empower clinical researchers to thoroughly inspect the rationale of models informing medical decisions. Domain experts can probe the validity of predictive patterns leveraged by models.
By facilitating scrutiny of survival models in healthcare, survex promotes accountability and helps ensure patient well-being remains the top priority in clinical AI adoption. Its transparency safeguards against risks like unchecked biases.
Future Directions
Upcoming extensions will expand supported model types like competing risks and incorporate new explanations. The ability to handle multiple event types will broaden applicability across domains. Operationalizing AI ethics requires translating principles into practical software tools. By generating tailored interactive insights into survival models, survex provides a blueprint for responsible machine learning in sensitive real-world analysis.
*Important notice: arXiv publishes preliminary scientific reports that are not peer-reviewed and, therefore, should not be regarded as definitive, used to guide development decisions, or treated as established information in the field of artificial intelligence research.
Journal reference:
- Preliminary scientific report.
Spytek, M., Krzyziński, M., Langbein, S. H., Baniecki, H., Wright, M. N., & Biecek, P. (2023, August 30). survex: an R package for explaining machine learning survival models. ArXiv.org. https://doi.org/10.48550/arXiv.2308.16113, https://arxiv.org/abs/2308.16113