Preventing dataset shift from breaking machine-learning biomarkers
Open accessThe DOI matches its Crossref record; 30 of the 30 references it cites were found in a registry or an index.
Abstract
Machine learning brings the hope of finding new biomarkers extracted from cohorts with rich biomedical measurements. A good biomarker is one that gives reliable detection of the corresponding condition. However, biomarkers are often extracted from a cohort that differs from the target population. Such a mismatch, known as a dataset shift, can undermine the application of the biomarker to new individuals. Dataset shifts are frequent in biomedical research, e.g., because of recruitment biases. When a dataset shift occurs, standard machine-learning techniques do not suffice to extract and validate biomarkers. This article provides an overview of when and how dataset shifts break machine-learning–extracted biomarkers, as well as detection and correction strategies.
Overview
The claims and statistics of this paper haven’t been checked yet.
Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
The claims and statistics of this paper haven’t been checked yet.
Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
The self-check needs the paper’s full text.
Citations
Cites
Sudlow, C., Gallacher, J., Allen, N., Beral, V., Burton, P., Danesh, J., Downey, P., Elliott, P., Green, J., Landray, M., et al. (2015) · PLOS Medicine · doi:10.1371/journal.pmed.1001779 · Cited by 12,728 · cited 2×
“The same behavior is visible on real epidemiologic data with age shifts, such as predicting the smoking status of participants in the UKBiobank study [23], as shown in Figure 2.”
BackgroundAustin, P. C. (2011) · Multivariate Behavioral Research · doi:10.1080/00273171.2011.568786 · Cited by 12,242 · cited 1×
“Indeed, an over-constrained model may be able to fit the data well only in some regions of the input feature space (Fig.”
He, H., & Garcia, E. A. (2009) · IEEE Transactions on Knowledge and Data Engineering · doi:10.1109/tkde.2008.239 · Cited by 10,359 · cited 1×
“Even more frequently, prior probability shift arises when one rare class is over-represented in the training data so that the dataset is more balanced, as when extracting a biomarker from a case-control cohort, or when the dataset is resampled as a strategy to handle the class imbalance problem [61].”
Adversarial Discriminative Domain Adaptation
Tzeng, E., Hoffman, J., Saenko, K., & Darrell, T. (2017) · 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2017.316 · Cited by 5,154 · cited 1×
“Therefore in predictive settings, where the goal is machine-learning models that generalize well, large and diverse datasets are desirable.”
Learning Transferable Features with Deep Adaptation Networks
Long, M., Cao, Y., Wang, J., & Jordan, M. I. (2015) · arXiv · doi:10.48550/arxiv.1502.02791 · Cited by 4,505 · cited 1×
“…methods strive to extract features that are predictive of the target while having similar distributions in the source and target domains [e.g. Long et al., 2015], or while preventing an adversary to distinguish source and target data [“domain-adversarial” learning, e.g. Tzeng et al., 2017].”
MethodA survey of cross-validation procedures for model selection
Arlot, S., & Celisse, A. (2010) · Statistics Surveys · doi:10.1214/09-ss054 · Cited by 3,522
Deo, R. C. (2015) · Circulation · doi:10.1161/circulationaha.115.001593 · Cited by 3,466
Henrich, J., Heine, S. J., & Norenzayan, A. (2010) · Nature · doi:10.1038/466029a · Cited by 3,441
Wynants, L., Van Calster, B., Bonten, M. M., Collins, G. S., Debray, T. P., De Vos, M., Haller, M. C., Heinze, G., Moons, K. G., Riley, R. D., et al. (2020) · doi:10.1101/2020.03.24.20041020 · Cited by 3,070
A structural approach to selection bias
Hernán, M. A., Hernández-Díaz, S., & Robins, J. M. (2004) · Epidemiology · doi:10.1097/01.ede.0000135174.63482.43 · Cited by 2,844 · cited 1×
“However the usual (unweighted) estimator is not consistent for an over-constrained (misspecified) model.”
Cited by
Crosby, D., Bhatia, S., Brindle, K. M., Coussens, L. M., Dive, C., Emberton, M., Esener, S., Fitzgerald, R. C., Gambhir, S. S., Kuhn, P., et al. (2022) · Science · doi:10.1126/science.aay9040 · Cited by 1,379
“AI andMLmodels are often developed (or trained) on datasets derived from selected populations that do not represent the real population where the AI-derived test would be used, so they cannot be extrapolated to real-life conditions (91).”
Machine learning for medical imaging: methodological failures and recommendations for the future
Varoquaux, G., & Cheplygina, V. (2022) · npj Digital Medicine · doi:10.1038/s41746-022-00592-y · Cited by 656
Self-citationData drift in medical machine learning: implications and potential remedies
Sahiner, B., Chen, W., Samala, R. K., & Petrick, N. (2023) · The British Journal of Radiology · doi:10.1259/bjr.20220878 · Cited by 279
Evaluating Machine Learning Models and Their Diagnostic Value
Varoquaux, G., & Colliot, O. (2023) · Neuromethods · doi:10.1007/978-1-0716-3195-9_20 · Cited by 242
Self-citationData leakage inflates prediction performance in connectome-based machine learning models
Rosenblatt, M., Tejavibulya, L., Jiang, R., Noble, S., & Scheinost, D. (2024) · Nature Communications · doi:10.1038/s41467-024-46150-w · Cited by 236
Generative models improve fairness of medical classifiers under distribution shifts
Ktena, I., Wiles, O., Albuquerque, I., Rebuffi, S.-A., Tanno, R., Roy, A. G., Azizi, S., Belgrave, D., Kohli, P., Cemgil, T., et al. (2024) · Nature Medicine · doi:10.1038/s41591-024-02838-6 · Cited by 165
“This is because of data scarcity 16 , challenges in the acquisition strategies of evaluation datasets (for example, different imaging or screening protocols 10,17,18 ) and the limitations of evaluation metrics 10 .”
Generative models improve fairness of medical classifiers under distribution shifts
Wiles, O., Albuquerque, I., Rebuffi, S.-A., Tanno, R., Roy, A. G., Azizi, S., Belgrave, D., Kohli, P., Cemgil, T., Karthikesalingam, A., et al. (2023) · doi:10.21203/rs.3.rs-2976332/v1 · Cited by 165
Guiding principles for the responsible development of artificial intelligence tools for healthcare
Badal, K., Lee, C. M., & Esserman, L. J. (2023) · Communications Medicine · doi:10.1038/s43856-023-00279-9 · Cited by 77
Statistical quantification of confounding bias in machine learning models
Spisak, T. (2022) · GigaScience · doi:10.1093/gigascience/giac082 · Cited by 34
Eichinski, P., Alexander, C., Roe, P., Parsons, S., & Fuller, S. (2022) · Frontiers in Ecology and Evolution · doi:10.3389/fevo.2022.810330 · Cited by 29
“A mismatch between training data and the unlabeled inference data encountered at deployment is an issue encountered in many machine learning scenarios and is known as dataset shift or domain shift (Dockès et al., 2021; Kouw and Loog, 2021; Stacke et al., 2021).”
BackgroundMellot, A., Collas, A., Rodrigues, P. L. C., Engemann, D., & Gramfort, A. (2023) · Imaging Neuroscience · doi:10.1162/imag_a_00040 · Cited by 16
Low, D. M., Rao, V., Randolph, G., Song, P. C., & Ghosh, S. S. (2024) · PLOS Digital Health · doi:10.1371/journal.pdig.0000516 · Cited by 16
Metadata
- Authors
- Dockès, J., Varoquaux, G., & Poline, J.-B.
- Venue
- GigaScience
- Publisher
- Oxford University Press (OUP)
- Published
- 2021-09
- Volume
- 10
- Issue
- 9
- ISSN
- 2047-217X (ISSN-L)
- OpenAlex
- W3183267169
- Cited by
- 86
- References
- 30
- FWCI
- 4.94
- Type
- journal article
- Language
- en
- Open access
- gold
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.
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
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.