New arXiv paper shows 'memorisation bias': medical AI trained on a patient's history can miss new conditions
chaumian · x · 2026-09-16
- Ben Glocker and colleagues introduce memorisation bias: medical AI models that saw a patient's anonymised historical records during training produce significantly different predictions on that patient's unseen future data.
- The bias spans diverse modalities and architectures, and can persist in records acquired decades after the training data.
- In simulated prospective deployment, effects are asymmetric: when a returning patient presents a de novo condition absent from their historical records, diagnostic sensitivity drops significantly; if their health state is unchanged, both sensitivity and specificity are inflated.
- The findings reveal a previously uncharacterised clinical deployment risk for models used on patients whose data was in training.
More from Research
- Memory layer cuts agent context tokens 23-62x and beats full history on 90-day recall — No_Advertising2536 · 2026-09-16
- Mathematician satirizes AI benchmarks: 'we never solved problems for benchmarks' — RexDouglass · 2026-09-16
- Japan's crowdsourced Minna de Honkoku OCR upgrades to v19 with better kanbun recognition — tkasasagi · 2026-09-16
- Reddit essay: LLMs just blur information like a weighted sum — and AGI isn't here — Helpful-Series132 · 2026-09-16
- Jev programming likened to MapReduce for decisions: parallel, mutually unaware queries — cocktailpeanut · 2026-09-16
- CMU's Decoy Direction Optimization blocks refusal-ablation attacks at 30-450x lower cost — CarnegieMellonU · 2026-09-16