CoChat
Sign Up
Cited by 122Machine Learning in HealthcareArtificial Intelligence in HealthcareDiabetes Management and Research

The DOI matches its Crossref record; 91 of the 93 references it cites were found in a registry or an index.

Predicting adverse outcomes due to diabetes complications with machine learning using administrative health data

Open access
Ravaut, Mathieu, Sadeghi, Hamed, Leung, Kin Kwan, Volkovs, Maksims, Kornas, Kathy, Harish, Vinyas, Watson, Tristan, Lewis, Gary F., Weisman, Alanna, Poutanen, Tomi, Rosella, Laura·2021·npj Digital Medicine
DOI 10.1038/s41746-021-00394-8PMC PMC7881135
Open this paper in CoChat — free accountOpen PDF

Abstract

Across jurisdictions, government and health insurance providers hold a large amount of data from patient interactions with the healthcare system. We aimed to develop a machine learning-based model for predicting adverse outcomes due to diabetes complications using administrative health data from the single-payer health system in Ontario, Canada. A Gradient Boosting Decision Tree model was trained on data from 1,029,366 patients, validated on 272,864 patients, and tested on 265,406 patients. Discrimination was assessed using the AUC statistic and calibration was assessed visually using calibration plots overall and across population subgroups. Our model predicting three-year risk of adverse outcomes due to diabetes complications (hyper/hypoglycemia, tissue infection, retinopathy, cardiovascular events, amputation) included 700 features from multiple diverse data sources and had strong discrimination (average test AUC = 77.7, range 77.7–77.9). Through the design and validation of a high-performance model to predict diabetes complications adverse outcomes at the population level, we demonstrate the potential of machine learning and administrative health data to inform health planning and healthcare resource allocation for diabetes management.

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/s41746-021-00394-8

Cites

91 of 91 references verified0 citation errors · 0 not checkable

Statistics

Figures

Fig. 1 Overview of our end-to-end prediction pipeline. (Step 1) Events from multiple administrative health datasets are ordered chronologically and mapped onto patient timeline. (Steps 2 and 3a) Patient timeline is partitioned into patient-time instances and each instance is assigned an observation, buffer and target window. In this study, observation window is 2 years, buffer is 3 years, target window is 3 months and we use nonoverlapping target windows. To illustrate this partitioning, we consider the test period which runs from Jan 1, 2016 to Dec 31, 2016 as an example. In this period a given patient has four instances at Jan 1, 2016, Apr 1, 2016, Jul 1, 2016, and Oct 1, 2016. The first instance has observation window [Jan 1, 2011–Dec 31, 2012], buffer [Jan 1, 2013–Dec 31, 2015], and target window [Jan 1, 2016–Mar 31, 2016]. The second instance has observation window [Apr 1, 2011–Mar 31, 2013], buffer [Apr 1, 2013–Mar 31, 2016], and target window [Apr 1, 2016–Jun 30, 2016], and so on. (Step 3b) Events from the observation window are used to extract features that summarize patient’s health history up to the end of the observation window. (Step 4) Extracted features are passed to the machine learning model to generate adverse outcome predictions. The goal of the model is to accurately predict which instances will have an adverse outcome from each of the six complications in the target window.
Fig. 1 Overview of our end-to-end prediction pipeline. (Step 1) Events from multiple administrative health datasets are ordered chronologically and mapped onto patient timeline. (Steps 2 and 3a) Patient timeline is partitioned into patient-time instances and each instance is assigned an observation, buffer and target window. In this study, observation window is 2 years, buffer is 3 years, target window is 3 months and we use nonoverlapping target windows. To illustrate this partitioning, we consider the test period which runs from Jan 1, 2016 to Dec 31, 2016 as an example. In this period a given patient has four instances at Jan 1, 2016, Apr 1, 2016, Jul 1, 2016, and Oct 1, 2016. The first instance has observation window [Jan 1, 2011–Dec 31, 2012], buffer [Jan 1, 2013–Dec 31, 2015], and target window [Jan 1, 2016–Mar 31, 2016]. The second instance has observation window [Apr 1, 2011–Mar 31, 2013], buffer [Apr 1, 2013–Mar 31, 2016], and target window [Apr 1, 2016–Jun 30, 2016], and so on. (Step 3b) Events from the observation window are used to extract features that summarize patient’s health history up to the end of the observation window. (Step 4) Extracted features are passed to the machine learning model to generate adverse outcome predictions. The goal of the model is to accurately predict which instances will have an adverse outcome from each of the six complications in the target window.CC BY · PubMed Central
Fig. 2 Test set AUC and calibration curves for all six complications. The average test AUC is 77.74 (77.7–77.9). The model is retrained five times with random restarts. The reported AUC results are averaged across the restarts. Corresponding ranges are also shown, low variance signifies that the results are stable. Calibration curves are computed using 20 bins of identical size, well-calibrated models have curves close to the identity line. Incidence rates vary significantly across complications (from 0.04 to 1.01%) so calibration curves have correspondingly different ranges.
Fig. 2 Test set AUC and calibration curves for all six complications. The average test AUC is 77.74 (77.7–77.9). The model is retrained five times with random restarts. The reported AUC results are averaged across the restarts. Corresponding ranges are also shown, low variance signifies that the results are stable. Calibration curves are computed using 20 bins of identical size, well-calibrated models have curves close to the identity line. Incidence rates vary significantly across complications (from 0.04 to 1.01%) so calibration curves have correspondingly different ranges.CC BY · PubMed Central
Fig. 3 Top eight features for each complication. For each feature we show the corresponding administrative health dataset where it is derived from, and the magnitude of the contribution to the model. The contribution is measured using the mean absolute Shapley values (see Methods) over a large random sample of 10,000 test instances. The feature contributions here do not represent causal effects, and only indicate correlation with the target predicted adverse outcomes as captured by the model.
Fig. 3 Top eight features for each complication. For each feature we show the corresponding administrative health dataset where it is derived from, and the magnitude of the contribution to the model. The contribution is measured using the mean absolute Shapley values (see Methods) over a large random sample of 10,000 test instances. The feature contributions here do not represent causal effects, and only indicate correlation with the target predicted adverse outcomes as captured by the model.CC BY · PubMed Central
Fig. 4 Adverse outcome cost analysis and high-risk statistics across complications. The total annual cost for adverse outcomes across all five complications is estimated to be ~$3.9B. a Annual cost for most at-risk patients predicted by our model. For each complication, we sort all patients by the predicted likelihood of adverse outcome, then compute total cost for each percentage of patients in the sorted list from top 1–100%. b A detailed sub-view from panel a for top 5% of predicted patients with the total cost across all six complications. c Statistics for the top 1% of most at risk patients predicted by our model. We analyze age, sex, immigration status, and HbA1c values, and compare them to the full cohort. All statistics are computed at the end of the observation window for each patient when the model makes its prediction. d Breakdown of the contribution to the total annual cost by complication.
Fig. 4 Adverse outcome cost analysis and high-risk statistics across complications. The total annual cost for adverse outcomes across all five complications is estimated to be ~$3.9B. a Annual cost for most at-risk patients predicted by our model. For each complication, we sort all patients by the predicted likelihood of adverse outcome, then compute total cost for each percentage of patients in the sorted list from top 1–100%. b A detailed sub-view from panel a for top 5% of predicted patients with the total cost across all six complications. c Statistics for the top 1% of most at risk patients predicted by our model. We analyze age, sex, immigration status, and HbA1c values, and compare them to the full cohort. All statistics are computed at the end of the observation window for each patient when the model makes its prediction. d Breakdown of the contribution to the total annual cost by complication.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

Show all 93 references

Research works printed but not linked: 2

  1. Bojer, C. & Meldgaard, J. Learnings from Kaggle’s Forecasting Competitions. arXiv, https://arxiv.org/ftp/arxiv/papers/2009/2009.07701.pdf (2020).
  2. Vaswani, A. et al. in Advances in Neural Information Processing Systems (eds. Guyon, I. et al.) 5998–6008 (Curran Associates, Inc., 2017).

Cited by

Sorted by citations

Related

See all

Metadata

Authors
Ravaut, M., Sadeghi, H., Leung, K. K., Volkovs, M., Kornas, K., Harish, V., Watson, T., Lewis, G. F., Weisman, A., Poutanen, T., & Rosella, L.
Affiliations
  • University of Toronto (CA)
  • Mount Sinai Hospital (CA)
Corresponding author
Rosella, Laura
ORCID
Venue
npj Digital Medicine
Publisher
Springer Science and Business Media LLC
Published
2021-02-12
Volume
4
Issue
1
Pages
24-24
ISSN
2398-6352 (ISSN-L)
DOI
10.1038/s41746-021-00394-8
PMC
PMC7881135
OpenAlex
W3127218785
Cited by
121
References
93
Citations by year
2026: 20 · 2025: 30 · 2024: 18 · 2023: 28 · 2022: 18 · 2021: 7
FWCI
6.27
Type
journal article
Language
en
Open access
Licence
Locations
  • npj Digital Medicine (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
  • Canada Research Chairs
  • Gouvernement du Canada | Canadian Institutes of Health Research
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Machine Learning in Healthcare
  • Health Sciences › Health Professions › Health Information Management › Artificial Intelligence in Healthcare
  • Health Sciences › Medicine › Endocrinology, Diabetes and Metabolism › Diabetes Management and Research
Keywords
diabetes complications, adverse outcome prediction, administrative health data, gradient boosting decision tree, tissue infection, diabetic retinopathy, cardiovascular events, amputation, single-payer health care, Ontario, healthcare resource allocation, health planning

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