An Elastic Net Regression Model for Identifying Long COVID Patients Using Health Administrative Data: A Population-Based Study

Abstract

Background Long COVID patients experience persistent symptoms after acute SARS-CoV-2 infection. Healthcare utilization data could provide critical information on the disease burden of long COVID for service planning, however, not all patients are diagnosed or assigned long COVID diagnostic codes. We developed an algorithm to identify individuals with long COVID using population-level health administrative data from British Columbia (BC), Canada. Methods An elastic net penalized logistic regression model was developed to identify long COVID patients based on demographic characteristics, pre-existing conditions, COVID-19-related data, and all symptoms/conditions recorded 28-183 days after the COVID-19 symptom onset/reported (index) date of known long COVID patients (N = 2,430) and a control group (N = 24,300); selected from all adult COVID-19 cases in BC with an index date on/before October 31, 2021 (N = 168,111). Known long COVID cases were diagnosed in a clinic and/or had the ICD-10-CA code for “Post COVID-19 condition” in their records. Results The algorithm retained known symptoms/conditions associated with long COVID, demonstrating high sensitivity (86%), specificity (86%), and area under the receiver operator curve (93%). It identified 25,220 (18%) long COVID patients among the remaining 141,381 adult COVID-19 cases; over ten times the number of known cases. Known and predicted long COVID patients had comparable demographic and health-related characteristics. Conclusions Our algorithm identified long COVID patients with a high level of accuracy. This large cohort of long COVID patients will serve as a platform for robust assessments on the clinical course of long COVID, and provide much needed concrete information for decision-making.

Heterogeneity of Diagnosis and Documentation of Post-COVID Conditions in Primary Care: A Machine Learning Analysis

“…Prior research found low concordance between the ICD-10 code for PCC and the clinical criteria for the disease, (4) and a right-skewed distribution of PCC diagnoses across clinicians. (5) Prior machine learning-based studies that focused on counterfactual probability of diagnosis had patients visited specialty PCC clinics all found different estimates of the mean risk of PCC following COVID-19, ranging from approximately 20% in Binka, et al to over 40% in Pfaff, et al(7,8) This mirrors an even wider difference in estimates derived from prospective patient surveys, which range from 4.5% to 89% --a difference that has been attributed to heterogeneous case definitions. (27) Our study’s results also agreed with prior research that found PCC diagnoses demographically cluster among non-Hispanic white female patients.…”

“…(7) Binka, et al used ridge regression with administrative data from British Columbia for the same aim. (8) Among the high-prevalence test set of patients attending or diagnosed at a specialty PCC clinic, all achieved good performance, though the generalizability to patients outside that setting is unknown.…”

Cohort profile: the British Columbia COVID-19 Cohort (BCC19C)—a dynamic, linked population-based cohort

“…Initial work focused on the use of machine learning approaches to identify long-COVID patients in administrative datasets. In our first paper, an elastic net regression model leveraging several BCC19C datasets was used to create an algorithm for identifying long-COVID patients ( 23 ). Clinical data on individuals enrolled in post-COVID-recovery clinics in BC were integrated into the BCC19C to develop and validate the algorithm.…”

A survey on the role of artificial intelligence in managing Long COVID

“…Binka et al ( 2022 ) detect Long COVID cases from health administrative data, including demographic features, pre-existing conditions, COVID-19-related data and all symptoms recorded 28 days after the COVID-19 symptom index date and lasted up to 183 days afterwards.…”