New research flags "administrative-cutoff leakage" in survival models

A new research paper identifies a potential pitfall in survival models that use time-indexed inputs, particularly in clinical prediction. The study highlights a phenomenon called "administrative-cutoff leakage," where models may learn to predict the acquisition date of data rather than the actual risk of an event. This occurs because more recent data inherently has less potential follow-up time than older data, leading to biased predictions. The paper proposes methods to detect this leakage and suggests a design principle for survival prediction models to mitigate this bias. AI

IMPACT Highlights a potential bias in AI models used for clinical prediction, urging careful data handling and model design.

RANK_REASON The cluster contains a single academic paper published on arXiv discussing a technical issue in survival models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• administrative-cutoff leakage
• arXiv
• electronic health records
• Harrell's C-index
• Mammography Cohort
• Survival models with Time-Indexed Inputs

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →