Distribution-Free Selection of Low-Risk Oncology Patients for Survival Beyond a Time Horizon

2025-12-19

AI Summary

This study addresses the clinical need for treatment de-escalation and healthcare resource optimization in oncology by proposing a distribution-free method to identify low-risk patients—those with extremely low probability of adverse events (e.g., death or disease progression) within a prespecified time window. Methodologically, it is the first to systematically compare two assumption-free frameworks: high-probability risk control and false discovery rate (FDR) control via conformal p-values, and innovatively incorporates inverse probability censoring weighting (IPCW) to accommodate right-censored survival data. The method seamlessly integrates with any black-box survival model. Experiments on real-world Flatiron oncology data and semi-synthetic datasets demonstrate that both frameworks reliably satisfy the target survival probability constraint; the conformal approach achieves higher selection efficiency, while the high-probability framework offers stronger theoretical guarantees.

Technology Category

• Machine Learning: Calibration & Uncertainty Quantification
• Reasoning under Uncertainty: Stochastic Optimization
• Constraint Satisfaction and Optimization: Distributed CSP/Optimization

Application Category

• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Responsible Web: Consent frameworks and practices on the web

Abstract

We study the problem of selecting a subset of patients who are unlikely to experience an event within a specified time horizon, by calibrating a screening rule based on the output of a black-box survival model. This statistics problem has many applications in medicine, including identifying candidates for treatment de-escalation and prioritizing the allocation of limited medical resources. In this paper, we compare two families of methods that can provide different types of distribution-free guarantees for this task: (i) high-probability risk control and (ii) expectation-based false discovery rate control using conformal $p$-values. We clarify the relation between these two frameworks, which have important conceptual differences, and explain how each can be adapted to analyze time-to-event data using inverse probability of censoring weighting. Through experiments on semi-synthetic and real oncology data from the Flatiron Health Research Database, we find that both approaches often achieve the desired survival rate among selected patients, but with distinct efficiency profiles. The conformal method tends to be more powerful, whereas high-probability risk control offers stronger guarantees at the cost of some additional conservativeness. Finally, we provide practical guidance on implementation and parameter tuning.

Problem

• Selecting low-risk oncology patients for survival beyond a specific time horizon
• Comparing distribution-free methods for risk control and false discovery rate
• Adapting frameworks to time-to-event data using censoring weighting techniques

Innovation

• Calibrating black-box survival model screening rules
• Using distribution-free guarantees with conformal p-values
• Applying inverse probability of censoring weighting to time-to-event data