CoChat
Sign Up
Cited by 138Machine Learning in HealthcareDomain Adaptation and Few-Shot LearningSepsis Diagnosis and Treatment

The DOI matches its Crossref record; 49 of the 57 references it cites were found in a registry or an index.

Evaluation of domain generalization and adaptation on improving model robustness to temporal dataset shift in clinical medicine

Open access
Guo, Lin Lawrence, Pfohl, Stephen R., Fries, Jason, Johnson, Alistair E. W., Posada, Jose, Aftandilian, Catherine, Shah, Nigam, Sung, Lillian·2022·Scientific Reports
DOI 10.1038/s41598-022-06484-1PMC PMC8854561

Published version of a preprint

Open this paper in CoChat — free accountOpen PDF

Abstract

Temporal dataset shift associated with changes in healthcare over time is a barrier to deploying machine learning-based clinical decision support systems. Algorithms that learn robust models by estimating invariant properties across time periods for domain generalization (DG) and unsupervised domain adaptation (UDA) might be suitable to proactively mitigate dataset shift. The objective was to characterize the impact of temporal dataset shift on clinical prediction models and benchmark DG and UDA algorithms on improving model robustness. In this cohort study, intensive care unit patients from the MIMIC-IV database were categorized by year groups (2008–2010, 2011–2013, 2014–2016 and 2017–2019). Tasks were predicting mortality, long length of stay, sepsis and invasive ventilation. Feedforward neural networks were used as prediction models. The baseline experiment trained models using empirical risk minimization (ERM) on 2008–2010 (ERM[08–10]) and evaluated them on subsequent year groups. DG experiment trained models using algorithms that estimated invariant properties using 2008–2016 and evaluated them on 2017–2019. UDA experiment leveraged unlabelled samples from 2017 to 2019 for unsupervised distribution matching. DG and UDA models were compared to ERM[08–16] models trained using 2008–2016. Main performance measures were area-under-the-receiver-operating-characteristic curve (AUROC), area-under-the-precision-recall curve and absolute calibration error. Threshold-based metrics including false-positives and false-negatives were used to assess the clinical impact of temporal dataset shift and its mitigation strategies. In the baseline experiments, dataset shift was most evident for sepsis prediction (maximum AUROC drop, 0.090; 95% confidence interval (CI), 0.080–0.101). Considering a scenario of 100 consecutively admitted patients showed that ERM[08–10] applied to 2017–2019 was associated with one additional false-negative among 11 patients with sepsis, when compared to the model applied to 2008–2010. When compared with ERM[08–16], DG and UDA experiments failed to produce more robust models (range of AUROC difference, − 0.003 to 0.050). In conclusion, DG and UDA failed to produce more robust models compared to ERM in the setting of temporal dataset shift. Alternate approaches are required to preserve model performance over time in clinical medicine.

Overview

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.1038/s41598-022-06484-1

Cites

49 of 49 references verified0 citation errors · 0 not checkable

Statistics

Figures

Figure 1 Data splitting procedure for baseline, domain generalization (DG) and unsupervised domain adaptation (UDA) experiments. Different shades of the same color indicate that they were used to train or evaluate different models. For instance, in the baseline experiment, the training set of each year group was used to learn models for that year group. In the DG experiment, the training year groups were kept separate to allow DG algorithms to estimate invariance across the year groups. In comparison, in the UDA experiment, data from the training year groups were pooled, and unlabeled samples from the target year group were leveraged for unsupervised distribution matching between training and target year groups. In addition, ERM[8–16] models were learned on pooled data from the training year groups (2008–2016) to be used as ERM comparators for DG and UDA models.
Figure 1 Data splitting procedure for baseline, domain generalization (DG) and unsupervised domain adaptation (UDA) experiments. Different shades of the same color indicate that they were used to train or evaluate different models. For instance, in the baseline experiment, the training set of each year group was used to learn models for that year group. In the DG experiment, the training year groups were kept separate to allow DG algorithms to estimate invariance across the year groups. In comparison, in the UDA experiment, data from the training year groups were pooled, and unlabeled samples from the target year group were leveraged for unsupervised distribution matching between training and target year groups. In addition, ERM[8–16] models were learned on pooled data from the training year groups (2008–2016) to be used as ERM comparators for DG and UDA models.CC BY · PubMed Central
Figure 2 Mean performance (AUROC, AUPRC, and ACE) of models in the baseline experiment. Solid blue lines depict models trained using 2008–2010 (ERM[8–10]) and evaluated in each year group. Dashed lines depict models trained and evaluated in each year group separately (comparators). Error bars indicate 95% confidence interval obtained from 10,000 bootstrap iterations. Black circles indicate statistically significant differences in performance based on the 95% confidence interval of the difference over 10,000 bootstrap iterations when comparing ERM[8–10] and comparators for each year group. The figure shows temporal dataset shift that is larger for Long LOS and Sepsis tasks. ERM: empirical risk minimization; LOS: length of stay; AUROC: area under the receiver operating characteristics curve; AUPRC: area under the precision recall curve; ACE: absolute calibration error.
Figure 2 Mean performance (AUROC, AUPRC, and ACE) of models in the baseline experiment. Solid blue lines depict models trained using 2008–2010 (ERM[8–10]) and evaluated in each year group. Dashed lines depict models trained and evaluated in each year group separately (comparators). Error bars indicate 95% confidence interval obtained from 10,000 bootstrap iterations. Black circles indicate statistically significant differences in performance based on the 95% confidence interval of the difference over 10,000 bootstrap iterations when comparing ERM[8–10] and comparators for each year group. The figure shows temporal dataset shift that is larger for Long LOS and Sepsis tasks. ERM: empirical risk minimization; LOS: length of stay; AUROC: area under the receiver operating characteristics curve; AUPRC: area under the precision recall curve; ACE: absolute calibration error.CC BY · PubMed Central
Figure 3 Difference in mean performance of DG and UDA approaches relative to ERM[8–16] in the target year group (2017–2019). Performance of ERM[8–10] (train set 2008–2010 and test set 2017–2019, dashed line) and ERM[17–19] (train and test sets 2017–2019, solid line) models are also shown for comparison. Error bars indicate 95% confidence interval obtained from 10,000 bootstrap iterations. Here, we show results from three of the four experimental conditions using differing number of unlabelled samples for UDA—we did not observe meaningful differences across the number of unlabelled samples evaluated. Numerical representation of the performance measures relative to ERM[8–16] are presented in Supplementary Table S3. LOS: length of stay; ERM: empirical risk minimization; IRM: invariant risk minimization; AL: adversarial learning; GroupDRO: group distributionally robust optimization; CORAL: correlation alignment; MMD: maximum mean discrepancy; AUROC: area under the receiver operating characteristics curve; AUPRC: area under the precision recall curve; ACE: absolute calibration error; domain generalization: DG; unsupervised domain adaptation: UDA.
Figure 3 Difference in mean performance of DG and UDA approaches relative to ERM[8–16] in the target year group (2017–2019). Performance of ERM[8–10] (train set 2008–2010 and test set 2017–2019, dashed line) and ERM[17–19] (train and test sets 2017–2019, solid line) models are also shown for comparison. Error bars indicate 95% confidence interval obtained from 10,000 bootstrap iterations. Here, we show results from three of the four experimental conditions using differing number of unlabelled samples for UDA—we did not observe meaningful differences across the number of unlabelled samples evaluated. Numerical representation of the performance measures relative to ERM[8–16] are presented in Supplementary Table S3. LOS: length of stay; ERM: empirical risk minimization; IRM: invariant risk minimization; AL: adversarial learning; GroupDRO: group distributionally robust optimization; CORAL: correlation alignment; MMD: maximum mean discrepancy; AUROC: area under the receiver operating characteristics curve; AUPRC: area under the precision recall curve; ACE: absolute calibration error; domain generalization: DG; unsupervised domain adaptation: UDA.CC BY · PubMed Central

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

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

Research works printed but not linked: 3

  1. Varnik, V. Principles of risk minimization for learning theory. In Advances in Neural Information Processing Systems Vol. 4 (eds Moody, J. E. et al.) 831–838 (NeurIPS, 1991).
  2. Efron, B. & Tibshirani, R. An Introduction to the Bootstrap (Chapman & Hall, 1993). · DOI · Paper page
  3. Zhao, H., Combes, R. T. D., Zhang, K. & Gordon, G. On learning invariant representations for domain adaptation. In 36th International Conference on Machine Learning, ICML 2019 7523–7532 (PMLR, 2019).

Cited by

Sorted by citations

See all 138

Metadata

Authors
Guo, L. L., Pfohl, S. R., Fries, J., Johnson, A. E. W., Posada, J., Aftandilian, C., Shah, N., & Sung, L.
Affiliations
  • Hospital for Sick Children (CA)
  • Stanford University (US)
Corresponding author
Sung, Lillian
ORCID
Venue
Scientific Reports
Publisher
Springer Science and Business Media LLC
Published
2022-02-17
Volume
12
Issue
1
Pages
2726-2726
ISSN
2045-2322 (ISSN-L)
DOI
10.1038/s41598-022-06484-1
PMC
PMC8854561
OpenAlex
W4221077687
Cited by
103
References
57
Citations by year
2026: 34 · 2025: 29 · 2024: 15 · 2023: 15 · 2022: 9
FWCI
5.59
Type
journal article
Language
en
Open access
Licence
Locations
  • Scientific Reports (publishedVersion, cc-by)
  • medRxiv (acceptedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Machine Learning in Healthcare
  • Physical Sciences › Computer Science › Artificial Intelligence › Domain Adaptation and Few-Shot Learning
  • Health Sciences › Medicine › Critical Care and Intensive Care Medicine › Sepsis Diagnosis and Treatment
Keywords
domain generalization, unsupervised domain adaptation, model robustness, clinical prediction model, intensive care unit patients, MIMIC-IV, sepsis prediction, mortality prediction, length of stay prediction, empirical risk minimization, clinical decision support systems, model calibration, feedforward neural networks
MeSH terms
Aged, Aged, 80 and over, Female, Humans, Intensive Care Units, Length of Stay, Male, Middle Aged, Models, Biological, Databases, Factual, Neural Networks, Computer, Sepsis

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