CoChat

Empirical data drift detection experiments on real-world medical imaging data

Open access
Kore, Ali, Abbasi Bavil, Elyar, Subasri, Vallijah, Abdalla, Moustafa, Fine, Benjamin, Dolatabadi, Elham, Abdalla, Mohamed · 2024 · Nature Communications

The DOI matches its Crossref record; 63 of the 63 references it cites were found in a registry or an index; 13 of the 28 sources its claims cite were checked for support.

Abstract

While it is common to monitor deployed clinical artificial intelligence (AI) models for performance degradation, it is less common for the input data to be monitored for data drift – systemic changes to input distributions. However, when real-time evaluation may not be practical (eg., labeling costs) or when gold-labels are automatically generated, we argue that tracking data drift becomes a vital addition for AI deployments. In this work, we perform empirical experiments on real-world medical imaging to evaluate three data drift detection methods’ ability to detect data drift caused (a) naturally (emergence of COVID-19 in X-rays) and (b) synthetically. We find that monitoring performance alone is not a good proxy for detecting data drift and that drift-detection heavily depends on sample size and patient features. Our work discusses the need and utility of data drift detection in various scenarios and highlights gaps in knowledge for the practical application of existing methods.

Overview

CoChat Summary

TL;DR

This study empirically evaluates data drift detection on real-world chest X-ray data, including natural drift from COVID-19 and synthetic demographic/pathology shifts. It finds that monitoring model performance alone is a poor proxy for data drift, while data-based drift detection is sensitive to sample size and the specific feature being enriched.

Key findings

  • Tracking aggregate model performance failed to capture clinically obvious COVID-19 drift: AUROC remained relatively stable across the first wave, whereas TAE + BBSD detected drift (with a delay). Aggregate scores changed less than 3% in the synthetic hospital-site shift shown in Fig. 4.
  • Drift-detection sensitivity depended strongly on the enriched feature: a 5% increase in patients aged 18–35 was detected by TAE + BBSD, but a 30% increase in inpatients or patients aged 65+ was needed for significance. The authors note patients 18–35 were 6% of the population, patients 65+ were 62%, inpatients were 57%, and outpatients were 19%.
  • Sample size strongly affected sensitivity: with 500 images the p-value never dropped below 0.1, while with 4000 images the p-value dropped below 0.05 at a 40% increase in male patients.
  • TAE + BBSD was generally more sensitive than TAE alone, and BBSD was nearly as sensitive as TAE + BBSD.

Methods

  • Retrospective study using 239,235 frontal chest radiographs from 78,542 patients over age 18 at Trillium Health Partners (Mississauga, Ontario, Canada), January 2016–December 2020; labels extracted with CheXpert NLP for 14 pathologies.
  • Compared four approaches: model-performance tracking, image data-based drift detection (TorchXRayVision AutoEncoder, TAE), model output-based drift detection (Black Box Shift Detection, BBSD), and combined TAE + BBSD. Drift was tested with a two-sided multivariate maximum mean discrepancy (MMD) statistical test.
  • Natural drift used COVID-19 emergence with rolling 30-day buckets and 500 images per source/target sample; synthetic categorical drifts enriched target datasets from 5% to 50% across sex, institution, ICU status, age, and patient class, repeated 10 times, generally using 4000 images per source and target unless testing sample size.

Limitations

  • The paper states findings are limited to computer vision algorithms on chest radiographs and should be evaluated on other data types; synthetic drift was limited to common patient attributes, and ethnicity/race were not explored because those data are not collected at Trillium Health Partners.
  • The paper states only a small selection of metrics was tested, with a focus on discriminative performance; calibration methods and other drift-detection methods, dimensionality reductions, and two-sample tests were not studied. It also notes no trustworthy automated way to explain what causes drift, and reports a hint of inequity because detection sensitivity differed across patient demographics.
  • Evident from the text: the study is retrospective and from a single three-site hospital system, so the natural COVID-19 drift and synthetic shifts may not generalize to other institutions, populations, or deployment contexts.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1038/s41467-024-46142-w

Cites

63 of 63 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Fig. 1 Data drift as detected by TAE + BBSD (image-and-output-based) drift detection over time on real-world data. The AUC is plotted in light blue and corresponds to the left y-axis. The p-value output from the TAE + BBSD approach (green) is also read against the left y-axis. p-values were calculated using a two-sided multivariate maximum mean discrepancy (MMD) statistical test with no correction for multiple comparison. The number of daily COVID hospitalizations at Trillium Health Partners is overlaid in red and can be read using the right y-axis. The horizontal line represents the value of 0.05 on the left y-axis. TAE trained autoencoder, BBSD black box shift detector, AUROC area under the curve – receiver operating characteristic.
Fig. 1 Data drift as detected by TAE + BBSD (image-and-output-based) drift detection over time on real-world data. The AUC is plotted in light blue and corresponds to the left y-axis. The p-value output from the TAE + BBSD approach (green) is also read against the left y-axis. p-values were calculated using a two-sided multivariate maximum mean discrepancy (MMD) statistical test with no correction for multiple comparison. The number of daily COVID hospitalizations at Trillium Health Partners is overlaid in red and can be read using the right y-axis. The horizontal line represents the value of 0.05 on the left y-axis. TAE trained autoencoder, BBSD black box shift detector, AUROC area under the curve – receiver operating characteristic.CC BY · PubMed Central
Fig. 2 A comparison between TAE (light blue), TAE + BBSD (red), BBSD (green), and AUROC (purple) monitoring methods for various synthetic drifts. For each subplot, the distribution of source and target dataset is different across the following representative categories: sex, institution, is_icu, patient age, and patient class. The p-values (outputs from the TAE, TAE + BBSD, BBSD) are plotted against the left y-axis while the AUROC is plotted against the right y-axis. p-values were calculated using a two-sided multivariate maximum mean discrepancy (MMD) statistical test with no correction for multiple comparison. Data are presented as mean values +/− 1 standard deviation. TAE trained autoencoder, BBSD black box shift detector, AUROC area under the curve – receiver operating characteristic.
Fig. 2 A comparison between TAE (light blue), TAE + BBSD (red), BBSD (green), and AUROC (purple) monitoring methods for various synthetic drifts. For each subplot, the distribution of source and target dataset is different across the following representative categories: sex, institution, is_icu, patient age, and patient class. The p-values (outputs from the TAE, TAE + BBSD, BBSD) are plotted against the left y-axis while the AUROC is plotted against the right y-axis. p-values were calculated using a two-sided multivariate maximum mean discrepancy (MMD) statistical test with no correction for multiple comparison. Data are presented as mean values +/− 1 standard deviation. TAE trained autoencoder, BBSD black box shift detector, AUROC area under the curve – receiver operating characteristic.CC BY · PubMed Central
Fig. 3 Experiment exploring the effect of sample size on data drift detection. Synthetic categorical shift for a target category of sex(M) with the TAE + BBSD method for a series of different sample sizes. p-values were calculated using a two-sided multivariate maximum mean discrepancy (MMD) statistical test with no correction for multiple comparison.
Fig. 3 Experiment exploring the effect of sample size on data drift detection. Synthetic categorical shift for a target category of sex(M) with the TAE + BBSD method for a series of different sample sizes. p-values were calculated using a two-sided multivariate maximum mean discrepancy (MMD) statistical test with no correction for multiple comparison.CC BY · PubMed Central
Fig. 4 Comparison of aggregate vs non-aggregate metrics for synthetic categorical shift for a target category of [Hospital Site 1] (4A) and [Hospital Site 2] (4B). The plots on the left show the absolute change in score for aggregate metrics (i.e., AUROC, Precision, Recall, F1 score, and Brier Score averaged all classes). The plots on the right show the breakdown across pathologies for the F1 score. Aggregate scores change less than 3%. p-values were calculated using a two-sided multivariate maximum mean discrepancy (MMD) statistical test with no correction for multiple comparison. Data are presented as mean values +/− 1 standard deviation. Eng. Card. stands for enlarged cardiomediastinum. AUROC area under the curve – receiver operating characteristic.
Fig. 4 Comparison of aggregate vs non-aggregate metrics for synthetic categorical shift for a target category of [Hospital Site 1] (4A) and [Hospital Site 2] (4B). The plots on the left show the absolute change in score for aggregate metrics (i.e., AUROC, Precision, Recall, F1 score, and Brier Score averaged all classes). The plots on the right show the breakdown across pathologies for the F1 score. Aggregate scores change less than 3%. p-values were calculated using a two-sided multivariate maximum mean discrepancy (MMD) statistical test with no correction for multiple comparison. Data are presented as mean values +/− 1 standard deviation. Eng. Card. stands for enlarged cardiomediastinum. AUROC area under the curve – receiver operating characteristic.CC BY · PubMed Central
Fig. 5 Illustration of sampling methodology and key parameters for COVID drift detection. We split acquired data into 30-day buckets. For example, S1, S2, and S3 represent different samples taken at different timepoints to explore the presence of data drift. For each sample we have two sub-samples: the source dataset and target dataset. Each dataset has a defined window_size (how many 30-day buckets can be used to draw patients), and sample_size (how many patients to draw from the selected window_size). Note, neither the window_size nor the sample_size of the source and target datasets need to match; the figure illustrates a difference in window_size. To increase the stability of comparisons, and avoid patients occurring in both source and target data, the stride parameter describes how many 30-day buckets separate the source and target datasets. Data are presented as mean values +/− 1 standard deviation.
Fig. 5 Illustration of sampling methodology and key parameters for COVID drift detection. We split acquired data into 30-day buckets. For example, S1, S2, and S3 represent different samples taken at different timepoints to explore the presence of data drift. For each sample we have two sub-samples: the source dataset and target dataset. Each dataset has a defined window_size (how many 30-day buckets can be used to draw patients), and sample_size (how many patients to draw from the selected window_size). Note, neither the window_size nor the sample_size of the source and target datasets need to match; the figure illustrates a difference in window_size. To increase the stability of comparisons, and avoid patients occurring in both source and target data, the stride parameter describes how many 30-day buckets separate the source and target datasets. Data are presented as mean values +/− 1 standard deviation.CC BY · PubMed Central
Fig. 6 The flowchart for the TAE + BBSD method. CXR images from the source and target sets are fed into both a trained autoencoder (TAE) and a black box shift detector (BBSD) to create a feature vector of each image with reduced dimensionality. The feature vector from the bottleneck of the TAE and the softmax probabilities from the classifier are concatenated and passed as one feature vector to the two-sample statistical test to determine if drift has occurred. Data are presented as mean values +/− 1 standard deviation.
Fig. 6 The flowchart for the TAE + BBSD method. CXR images from the source and target sets are fed into both a trained autoencoder (TAE) and a black box shift detector (BBSD) to create a feature vector of each image with reduced dimensionality. The feature vector from the bottleneck of the TAE and the softmax probabilities from the classifier are concatenated and passed as one feature vector to the two-sample statistical test to determine if drift has occurred. Data are presented as mean values +/− 1 standard deviation.CC BY · PubMed Central

Verification

Claims the paper makes

10 Valid · 1 Not valid · 6 Not checked · 47 Own results and interpretations

21 more claims

Create a free account to see all 64 claims and to verify the 4 references still open.

See all claims — free account
  • Probably not a substantial claim. 25
  • It may state two separate claims as one. 19
  • The quote only partly states this claim. 1

Create a free account to see these claims.

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

2 of 2

Citations

Cites

Show all 63 references

Cited by

See all 124

Metadata

Authors
Kore, A., Abbasi Bavil, E., Subasri, V., Abdalla, M., Fine, B., Dolatabadi, E., & Abdalla, M.
Affiliations
  • Vector Institute (CA)
  • University Health Network (CA)
  • Harvard University (US)
  • University of Toronto (CA)
  • York University (CA)
Corresponding author
  • Abdalla, Moustafa
  • Abdalla, Mohamed
Venue
Nature Communications
Published
2024-02-29
Volume
15
Issue
1
Pages
1887-1887
ISSN
2041-1723 (ISSN-L)
OpenAlex
W4392283037
Cited by
124
References
50
Citations by year
2026: 66 · 2025: 52 · 2024: 6
FWCI
18.22
Type
journal article
Language
en
Open access
Licence
Locations
  • Nature Communications (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • RePEc: Research Papers in Economics (submittedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Data Stream Mining Techniques
  • Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
  • Physical Sciences › Computer Science › Artificial Intelligence › Domain Adaptation and Few-Shot Learning
Keywords
data drift detection, diagnostic imaging, clinical AI, X-rays, sample size, field deployment
MeSH terms
COVID-19, Artificial Intelligence, Diagnostic Imaging, Humans, Radiography

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-09, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-09 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol and qwen3.8-omni-flash).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Empirical data drift detection experiments on real-world medical imaging data. https://app.cochat.ai/papers/10.1038/s41467-024-46142-w
The paper
Kore, A., Abbasi Bavil, E., Subasri, V., Abdalla, M., Fine, B., Dolatabadi, E., & Abdalla, M. (2024). Empirical data drift detection experiments on real-world medical imaging data. Nature Communications. https://doi.org/10.1038/s41467-024-46142-w

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.