Memorisation bias in medical AI

• URL: http://arxiv.org/abs/2609.17223v1
• Date: Tue, 15 Sep 2026 14:07:01 GMT
• Title: Memorisation bias in medical AI
• Abstract summary: We show that predictions on a patient's unseen future data can change significantly if a model observed that same patient's anonymised historical data during training.<n>We demonstrate that this bias exists across diverse data modalities and model architectures.<n>In simulated prospective deployment, memorisation bias has asymmetric effects on the diagnostic accuracy of returning data contributors.
• Score: 33.76985328787124
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Medical AI models hold immense potential to improve patient outcomes, but they are also known to unintentionally memorise individual records from their training datasets. While such memorisation has been linked to targeted privacy attacks, its consequences for clinical deployment, where patients may be assessed by a model that saw their historical data during training, remain poorly understood. Here we show that predictions on a patient's unseen future data can change significantly if a model observed that same patient's anonymised historical data during training, a phenomenon we term "memorisation bias". We demonstrate that this bias exists across diverse data modalities and model architectures, and over prolonged time spans: in some cases, memorisation bias persists on future records acquired decades after the historical records used for training. Moreover, in simulated prospective deployment, memorisation bias has asymmetric effects on the diagnostic accuracy of returning data contributors. When a patient returned with a de novo condition absent from their historical records in the training dataset, diagnostic sensitivity decreased significantly compared to an otherwise identical model not trained on their historical data. Conversely, when their health state was unchanged, both sensitivity and specificity were significantly inflated. Our findings reveal a previously uncharacterised risk in medical AI that arises when a model is deployed on patients who contributed to its training data. This exposes a shortcoming of current model development practice: the de-identification measures designed to protect patients' privacy make it difficult to identify returning contributors and exclude them from the AI-assisted interpretation of their own future data. Mitigating memorisation risks may thus require changes to current model training and deployment protocols.

Related papers

• Pitfalls of Administrative Censoring in Survival Models with Time-Indexed Inputs [9.373139684950957]
  Survival models can model time-to-event outcomes using partially observed data.<n>Recent models often rely on rich inputs collected at a specific clinical encounter.<n>This creates a potential failure mode when outcomes are observed only up to a fixed study end date.
  arXiv  Detail & Related papers  (2026-07-11T20:10:16Z)
• Deep State-Space Generative Model For Correlated Time-to-Event Predictions [54.3637600983898]
  We propose a deep latent state-space generative model to capture the interactions among different types of correlated clinical events. Our method also uncovers meaningful insights about the latent correlations among mortality and different types of organ failures.
  arXiv  Detail & Related papers  (2024-07-28T02:42:36Z)
• Unconditional Latent Diffusion Models Memorize Patient Imaging Data: Implications for Openly Sharing Synthetic Data [2.04850174048739]
  We train latent diffusion models on CT, MR, and X-ray datasets for synthetic data generation.<n>We then detect the amount of training data memorized utilizing our novel self-supervised copy detection approach.<n>Our findings show a surprisingly high degree of patient data memorization across all datasets.
  arXiv  Detail & Related papers  (2024-02-01T22:58:21Z)
• IGNITE: Individualized GeNeration of Imputations in Time-series Electronic health records [6.630372114304835]
  We propose a novel deep-learning model that learns the underlying patient dynamics to generate personalized values conditioning on an individual's demographic characteristics and treatments.<n>Our proposed model, IGNITE, utilise a conditional dual-variational autoencoder augmented with dual-stage attention to generate missing values for an individual.<n>We show that IGNITE outperforms state-of-the-art approaches in missing data reconstruction and task prediction.
  arXiv  Detail & Related papers  (2024-01-09T07:57:21Z)
• MedDiffusion: Boosting Health Risk Prediction via Diffusion-based Data Augmentation [58.93221876843639]
  This paper introduces a novel, end-to-end diffusion-based risk prediction model, named MedDiffusion. It enhances risk prediction performance by creating synthetic patient data during training to enlarge sample space. It discerns hidden relationships between patient visits using a step-wise attention mechanism, enabling the model to automatically retain the most vital information for generating high-quality data.
  arXiv  Detail & Related papers  (2023-10-04T01:36:30Z)
• Learning and DiSentangling Patient Static Information from Time-series Electronic HEalth Record (STEER) [3.079694232219292]
  Recent work in machine learning for healthcare has raised concerns about patient privacy and algorithmic fairness. Here we systematically investigated the ability of time-series electronic health record data to predict patient static information. We found that not only the raw time-series data, but also learned representations from machine learning models, can be trained to predict a variety of static information.
  arXiv  Detail & Related papers  (2023-09-20T14:54:48Z)
• Private, fair and accurate: Training large-scale, privacy-preserving AI models in medical imaging [47.99192239793597]
  We evaluated the effect of privacy-preserving training of AI models regarding accuracy and fairness compared to non-private training. Our study shows that -- under the challenging realistic circumstances of a real-life clinical dataset -- the privacy-preserving training of diagnostic deep learning models is possible with excellent diagnostic accuracy and fairness.
  arXiv  Detail & Related papers  (2023-02-03T09:49:13Z)
• Reconstructing Training Data from Model Gradient, Provably [68.21082086264555]
  We reconstruct the training samples from a single gradient query at a randomly chosen parameter value. As a provable attack that reveals sensitive training data, our findings suggest potential severe threats to privacy.
  arXiv  Detail & Related papers  (2022-12-07T15:32:22Z)
• Textual Data Augmentation for Patient Outcomes Prediction [67.72545656557858]
  We propose a novel data augmentation method to generate artificial clinical notes in patients' Electronic Health Records. We fine-tune the generative language model GPT-2 to synthesize labeled text with the original training data. We evaluate our method on the most common patient outcome, i.e., the 30-day readmission rate.
  arXiv  Detail & Related papers  (2022-11-13T01:07:23Z)
• Integrated Convolutional and Recurrent Neural Networks for Health Risk Prediction using Patient Journey Data with Many Missing Values [9.418011774179794]
  This paper proposes a novel end-to-end approach to modeling EHR patient journey data with Integrated Convolutional and Recurrent Neural Networks. Our model can capture both long- and short-term temporal patterns within each patient journey and effectively handle the high degree of missingness in EHR data without any imputation data generation.
  arXiv  Detail & Related papers  (2022-11-11T07:36:18Z)
• Practical Challenges in Differentially-Private Federated Survival Analysis of Medical Data [57.19441629270029]
  In this paper, we take advantage of the inherent properties of neural networks to federate the process of training of survival analysis models. In the realistic setting of small medical datasets and only a few data centers, this noise makes it harder for the models to converge. We propose DPFed-post which adds a post-processing stage to the private federated learning scheme.
  arXiv  Detail & Related papers  (2022-02-08T10:03:24Z)
• DeepRite: Deep Recurrent Inverse TreatmEnt Weighting for Adjusting Time-varying Confounding in Modern Longitudinal Observational Data [68.29870617697532]
  We propose Deep Recurrent Inverse TreatmEnt weighting (DeepRite) for time-varying confounding in longitudinal data. DeepRite is shown to recover the ground truth from synthetic data, and estimate unbiased treatment effects from real data.
  arXiv  Detail & Related papers  (2020-10-28T15:05:08Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.