ISSN :2582-9793
Dean Eurich and Darren Lau
Adv. Artif. Intell. Mach. Learn., 5 (1):3565-3604
1. Dean Eurich: School of Public Health, University of Alberta
2. Darren Lau: Department of Medicine, Faculty of Medicine & Dentistry, University of Alberta
DOI: https://dx.doi.org/10.54364/AAIML.2025.51204
Article History: Received on: 30-Oct-24, Accepted on: 23-Mar-25, Published on: 30-Mar-25
Corresponding Author: Dean Eurich
Email: deurich@ualberta.ca
Citation: Dean Eurich, Darren Lau, Weiting Li, Olivia R Weaver, Tanya Joon, Ming Ye, Finlay A McAlister, Padma Kaul, Salim Samanani. (2025). Predicting COVID-19 outcomes among Albertans with diabetes and COVID-19: A machine learning approach.. Adv. Artif. Intell. Mach. Learn., 5 (1 ):3565-3604.
Background:
Certain patients with diabetes and COVID-19
are at high risk of severe outcomes. Identification of risk factors among this group
is required to risk-stratify those
who may benefit from further surveillance. We aimed to develop machine learning (ML) models predicting severe
outcomes among individuals with diabetes and COVID-19 in Alberta, Canada.
Methods:
Patients with diabetes and COVID-19 determined by PCR test administered in
community and/or emergency department (ED) settings (March 2020-March 2021) were
included. Outcomes were ED visit, hospitalization or death for those tested in
the community (“Community cohort”) and hospitalization or death for those
tested in ED (“ED cohort”), and in the combined cohorts (“Community+ED
cohort”). Outcomes and features (socio-demographics, drug/healthcare
utilization, health history) were identified using healthcare administrative
data (2008-2021). Calibration plots, areas under the receiver operating curve,
precision-recall curves (AUC, AUPRC), and threshold analyses were used to
assess the models.
Results: The
Community cohort included 11,247 individuals (1,665 ED visits; 756
hospitalizations; 421 deaths). AUCs for models predicting ED/hospitalization/death
were 0.65/0.70/0.93. The AUCs for predicting death in ED (1,495 individuals;
169 deaths) and Community+ED (12,410 individuals; 582 deaths) cohorts were 0.82 and 0.93. Models predicting
hospitalization in these cohorts performed poorly and are not reported. Of all models, that predicting death
in the Community performed best (sensitivity 0.77, specificity 0.91, positive
predictive value 0.26, negative predictive value 0.99), and improved the
prediction of death at a 10% risk threshold (compared to the pre-test
probability, positive likelihood ratio 9.06 and negative likelihood ratio 0.25).
Conclusion: Identifying
diabetes patients at the highest risk of the worst outcomes would assist in
triaging patients to ensure appropriate resource use in times of high demand.
Overall, the model predicting death among patients with diabetes and COVID-19 in
the community could be useful in identifying who requires additional care.