Research map / Large language models / AI and society
Statistically Valid Post-Deployment Monitoring Should Be Standard for AI-Based Digital Health
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Position paper arguing that post-deployment monitoring of clinical AI should be framed as statistically valid, label-efficient two-sample hypothesis testing for data shift, performance degradation, and subgroup impact.
Problem
Post-deployment monitoring of clinical AI models is manual, sporadic, and reactive, lacks statistical guarantees, and is hampered by scarce ground-truth labels, so performance degradation and data shifts go undetected.
Approach
Reframe monitoring tasks (covariate shift, label shift, concept drift, performance degradation, prediction-correctness shift, subgroup identification) as two-sample hypothesis tests with explicit error-rate control, and outline open problems in label-efficient testing, attribution, and subgroup discovery.
Key results
- Cites a review finding that only 9% of FDA-registered AI-based healthcare tools include a post-deployment surveillance plan.
- Proposes a permutation-based two-sample test procedure that turns divergences such as Jensen-Shannon, energy distance, or Wasserstein into valid hypothesis tests.
- Lists open problems in label scarcity, drift mechanism attribution, and impacted-subgroup identification without empirical experiments.
Limitations
- No empirical validation of the proposed tests on clinical data.
- Assumes locally i.i.d. data within pre- and post-deployment windows and does not handle temporal dependence.
- Excludes generative models such as LLMs and causal failure attribution.
Experimental setup
Baselines: continual learning Bayesian online change-point detection conformal prediction monitoring
Builds on: black box shift estimation permutation testing maximum mean discrepancy prediction aided by surrogate training
Closest papers by meaning
- Reliably detecting model failures in deployment without labels (NeurIPS 2025 · similarity 0.86)
- WATCH: Adaptive Monitoring for AI Deployments via Weighted-Conformal Martingales (ICML 2025 · similarity 0.82)
- DDx-TRACE: A Benchmark for Medical Diagnostic Trajectories in VLMs (NeurIPS 2026 · similarity 0.82)
- $\chi$-Bench: Can AI Agents Automate End-to-End, Long-Horizon, Policy-Rich Healthcare Workflows? (NeurIPS 2026 · similarity 0.82)
- Timely Clinical Diagnosis through Active Test Selection (NeurIPS 2025 · similarity 0.82)
- PhysicianBench: Evaluating LLM Agents in Real-World EHR Environments (NeurIPS 2026 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.