Abstract
Background
The COVID-19 pandemic placed immense strain on Canada’s healthcare system and nosocomial infections increased risk for both healthcare workers and high-risk patients. Nosocomial COVID-19 surveillance typically relies on time-based case definitions that aim to balance sensitivity against specificity, yielding inconsistent estimates of relative mortality risk. We developed a latent class analysis approach to probabilistically classify healthcare-associated infections using multiple indicator variables and evaluated mortality risk in hospitalized COVID-19 patients.
Methods
We analyzed COVID-19 surveillance data from Ontario’s Case Contact and Management System and COVaxON vaccine registry (March 17, 2020 to September 4, 2022). Latent class analysis (LCA) integrated five binary indicators (nosocomial flag, outbreak linkage, and three time intervals from admission to positive test) to classify 53,191 hospitalized cases by likelihood of healthcare-acquired infection. We estimated mortality odds and survival by classification using multivariable logistic regression and time-varying Cox models, adjusting for age, comorbidities, immunocompromised status, long-term care residence, vaccination, and pandemic wave.
Results
LCA identified three classes: unlikely (n = 46,819), moderately likely (n = 2,687), and likely (n = 3,685) healthcare-acquired infection. Compared to unlikely, moderately likely cases had elevated mortality (OR: 1.26, 95% CI: 1.14–1.40); likely cases showed no elevation in cumulative mortality (OR: 1.05, 95% CI: 0.96–1.15). Time-varying survival models revealed substantially higher mortality risk early in the period following infection: hazard ratios were 3.25 (95% CI: 2.99–3.52) for moderately likely and 4.88 (95% CI: 4.53–5.25) for likely healthcare-acquired infection. Single-indicator variable sensitivities ranged from 34% to 70%.
Conclusion
Defining nosocomial infections early in the COVID-19 pandemic was challenging due to the volume of cases and number of patients hospitalized. Several different characteristics can indicate a case’s likelihood of being healthcare-acquired, and using a single variable to define a nosocomial case often presents trade-offs in terms of sensitivity or specificity. These findings underscore the limitations of single-indicator definitions and the risk of misclassifying cases in a way that would underestimate the relative severity of hospital-acquired infection by biasing effects toward the null. Latent class modeling addresses sensitivity-specificity trade-offs inherent in time-based nosocomial definitions and reveals elevated mortality risk that strict definitions miss.
Supplementary Information
The online version contains supplementary material available at 10.1186/s12879-026-13564-4.
Keywords: COVID-19, SARS-CoV-2, Epidemiology, Public health, Pandemics, Nosocomial infection
Introduction
The coronavirus disease 2019 (COVID-19) pandemic exposed fundamental challenges in how we measure and understand healthcare-associated infections. While nosocomial severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) transmission has been widely recognized as a threat to the health and safety of both patients and healthcare workers [1–3], our ability to quantify its burden remains constrained by measurement approaches that prioritize administrative convenience over epidemiological validity. Current surveillance definitions of healthcare-associated COVID-19 typically rely on time elapsed between hospital admission and positive test, with cutoffs ranging from 3 to 15 days. These single-indicator, threshold-based approaches create an inherent tension: restrictive definitions (e.g., ≥ 15 days post-admission) achieve high specificity but miss many true nosocomial cases, while more inclusive definitions (e.g., ≥ 3 days) capture more cases at the cost of reduced specificity.
This trade-off has consequences for both surveillance accuracy and research validity. Using restrictive definitions may systematically underestimate the frequency and severity of healthcare-associated infection through misclassification bias, while more permissive definitions may result in overestimation. The resulting heterogeneity in measurement approaches has produced a fragmented and sometimes contradictory literature on the relative severity of nosocomial COVID-19 [4–6]. Some studies report lower mortality among hospital-acquired cases, while others find elevated risk. These discordant findings may reflect differences in populations or care settings, but they may also arise from differences in case classification.
We hypothesized that a probabilistic classification approach using latent class analysis (LCA) could address the measurement limitations of single-indicator definitions by integrating multiple nosocomial infection indicators simultaneously. LCA estimates the sensitivity and specificity of each indicator while assigning individuals to latent classes based on covariance patterns across indicators, avoiding the need for arbitrary thresholds. We further hypothesized that this refined classification would reveal mortality patterns obscured by misclassification in studies using restrictive definitions.
In this study, we used population-based surveillance data from Ontario, Canada to: (1) develop and validate a latent class analysis approach for classifying healthcare-associated COVID-19 that integrates multiple indicators with varying sensitivity and specificity profiles; and (2) estimate mortality risk associated with healthcare-acquired infection compared to community-acquired infection among hospitalized patients, adjusting for demographic and clinical confounders.
Methods
We utilized COVID-19 surveillance data from Ontario, Canada, including laboratory-confirmed case counts, outbreak information, hospitalizations, and vaccination records. Data were obtained from the Ontario Case and Contact Management (CCM) System and the CoVaxON vaccine information system. The CCM system aggregates case reports from Ontario’s 34 Public Health Units (PHUs) and includes laboratory confirmation, epidemiological linkage data, hospitalization status, and clinical outcomes [7]. CoVaxON vaccination data were linked to CCM case data through a unique identifier used in both databases to associate cases with their individual vaccination status [8]. Data spanned March 17, 2020, through September 4, 2022, encompassing six pandemic waves defined by dominant SARS-CoV-2 variants. Population denominators were obtained from Statistics Canada.
We restricted our analysis to cases that were hospitalized at any point during their illness. This exclusion criterion means that individuals who may have acquired COVID-19 in healthcare settings but were never admitted (e.g., healthcare workers, visitors, or those exposed in outpatient or emergency department settings) were not included in our analysis. Cases with incomplete data on key covariates were excluded from multivariable analyses. For some variables, a categorical response of “missing” was available in the dataset, and these individuals were retained in the analyses.
Classification of healthcare-associated infection using latent class analysis
Rationale for latent class approach
Accurate classification of healthcare-associated infections requires balancing sensitivity and specificity. Time-based definitions (e.g., positive test ≥ 3, ≥8, or ≥ 15 days after admission) vary in their performance characteristics: more restrictive cutoffs increase specificity but reduce sensitivity, while inclusive cutoffs capture more true cases at the cost of false positives. Administrative flags (e.g., “nosocomial” designation in surveillance systems) and outbreak linkage provide additional information but also have imperfect concordance with true infection source. Rather than selecting a single definition and accepting its inherent limitations, we employed latent class analysis (LCA) to integrate multiple indicators simultaneously. LCA is a probabilistic classification method that identifies unobserved subgroups (latent classes) based on covariance patterns across observed binary or categorical indicators [9]. By modeling the joint distribution of multiple imperfect measures, LCA assigns individuals to classes based on their posterior probability of membership. This approach has been widely used in infectious disease epidemiology to address diagnostic uncertainty and improve case classification when no gold standard exists [9–11].
Indicator variables
We defined five binary indicators of healthcare-associated SARS-CoV-2 infection:
Nosocomial flag: Cases classified as nosocomial in the CCM system by reporting PHUs.
Hospital outbreak linkage: Cases linked to a hospital-associated outbreak by PHU investigation.
≥ 3-day definition: Positive SARS-CoV-2 test occurring 3 or more days after hospital admission.
≥ 8-day definition: Positive test occurring 8 or more days after admission.
≥ 15-day definition: Positive test occurring 15 or more days after admission.
The time-based definitions reflect cutoffs commonly used in nosocomial surveillance, with the 3-day threshold representing a moderately inclusive approach, the 8-day threshold balancing sensitivity and specificity, and the 15-day threshold prioritizing specificity. The nosocomial flag in the CCM dataset represents classification by local PHUs based on individual case investigation and likely source of exposure. Classification criteria were not standardized across PHUs or over time throughout the pandemic. Hospital outbreaks were recorded by local PHUs according to the Ontario Ministry of Health provincial guidelines. A confirmed outbreak was defined when two or more individuals were epidemiologically linked (e.g., within a specified area/unit/floor/ward) with laboratory-confirmed SARS-CoV-2 infection within a 7-day period, where both cases reasonably acquired their infection in the healthcare setting. Outbreak declaration and case linkage were determined by local public health units based on epidemiologic investigation and facility-specific circumstances [12]. Of hospitalized cases who had their first positive test 8 or more days following hospital admission, 46.2% were captured as nosocomial cases in the CCM, and 67.6% were linked to a hospital outbreak, indicating that these administrative flags alone have a high potential to miss some infections acquired in-hospital.
Latent class model specification and selection
We fit LCA models assuming 2, 3, and 4 latent classes and compared model fit using the Bayesian Information Criterion (BIC), which balances model likelihood against complexity. Lower BIC values indicate better fit. The LCA model estimates the probability that each case belongs to each latent class, conditional on the observed pattern of indicator variables. Each case was assigned to the class for which it had the highest posterior probability of membership.
We conducted LCA using the poLCA package in R. The three-class solution was selected as optimal based on lowest BIC and clinical interpretability. We labeled the three classes as: (1) unlikely healthcare-acquired infection, (2) moderately likely healthcare-acquired infection, and (3) likely healthcare-acquired infection, based on the conditional probabilities of positive indicators within each class. It is important to note that the “likely” class is heavily weighted toward patients with positive tests ≥ 15 days after admission, meaning these individuals have necessarily survived hospitalization for at least two weeks before acquiring and being diagnosed with SARS-CoV-2. This conditioning on prolonged survival introduces the potential for survivorship bias when comparing mortality outcomes across classes.
Validation of indicator performance
Using the three-class solution, we calculated the estimated sensitivity (probability of a positive indicator given healthcare-associated infection) and specificity (probability of a negative indicator given community-acquired infection) of each indicator variable, treating the moderately likely and likely healthcare-acquired classes combined as the “true positive” reference group and the unlikely class as the “true negative” reference group. This approach provides empirical estimates of indicator performance in the absence of a gold standard for infection source.
Analysis of mortality risk by healthcare-acquisition status
We estimated odds of in-hospital death associated with healthcare-acquired COVID-19 using multivariable binomial logistic regression. The primary exposure was latent class membership (unlikely healthcare-acquired as reference, moderately likely, and likely as separate indicator variables). Covariates were selected a priori based on established associations with COVID-19 mortality and included: age, sex, presence of any documented comorbidity, immunocompromised status, long-term care home residence, vaccination status (≥ 2 doses received ≥ 14 days prior to positive test), and pandemic wave. A directed acyclic graph visualizing covariate relationships is presented in supplementary Fig. 1. Comorbidity data were recorded by public health units using standardized case investigation protocols established at the onset of the pandemic. Documented comorbidities included cardiovascular disease, diabetes, renal disease, chronic obstructive pulmonary disease, chronic liver disease, malignancy, and other chronic conditions, consistent with definitions used in prior Ontario COVID-19 mortality analyses [13]. We coded the presence of any documented comorbidity as a binary variable to reflect overall burden of chronic disease. Immunocompromised status was captured as a binary variable in the CCM system based on public health unit case investigation. PHU investigators determined immunocompromised status during case interviews, typically based on reported medical history including conditions such as HIV/AIDS, active malignancy, organ or stem cell transplant, primary immunodeficiency, and use of immunosuppressive medications. While specific criteria were not standardized across PHUs, this classification has demonstrated strong predictive validity for COVID-19 mortality in Ontario surveillance data [13]. Pandemic waves were defined as follows: wave 1 (before August 31, 2020; wild-type), wave 2 (September 1, 2020 to February 28, 2021; wild-type), wave 3 (March 1 to June 30, 2021; Alpha/Beta/Gamma), wave 4 (July 1 to December 25, 2021; Delta), wave 5 (December 26, 2021 to March 19, 2022; Omicron), and wave 6 (after March 20, 2022; Omicron). This classification follows Mitchell et al. [14] and accounts for temporal changes in viral transmissibility, severity, and immune landscape. Individual-level measures including socioeconomic status, race, reason for initial hospitalization, and frailty were not available in our dataset. Facility-level factors such as hospital size, number of admissions, or infection prevention practices in place were also unavailable. Cases that were never hospitalized were excluded from analysis by design. In addition to the primary logistic regression model analyzing odds of mortality across all three latent classes, we conducted a secondary analysis restricting comparisons to only the unlikely and moderately likely healthcare-acquired groups to minimize the impact of survivorship bias in the group with the most patients infected 15 or more days after admission. The logistic regression model was performed as described above, restricting to only the unlikely and moderately likely healthcare-acquired groups.
Survival analysis
To assess whether nosocomial infection likelihood was associated with in-hospital survival, a Cox proportional hazards model was fit to estimate the association between LCA class and time from positive SARS-CoV-2 test to mortality. Follow-up time was defined from first positive test to in-hospital death or discharge. Individuals without a valid discharge date, or whose positive test was recorded after death/discharge were excluded. The date of first positive test was used as the time origin to ensure correct temporal ordering between infection source (LCA class) and survival, although it is important to note that because LCA classification incorporates time from initial hospital admission to diagnosis, it is partially conditioned on pre-infection/pre-diagnosis survival. The model was adjusted for age, sex, presence of any documented comorbidity, immunocompromised status, long-term care home residence, vaccination status, and pandemic wave as above. The proportional hazards assumption was evaluated using Schoenfeld residuals. This assessment indicated that the effect of likelihood of nosocomial infection on mortality was not constant over time, violating the proportional hazards assumption. To account for this violation, a second Cox proportional hazards model with time-varying effects for LCA class was fit. An interaction term between LCA class and log(time) since positive test allowed for the hazard ratio associated with nosocomial infection likelihood to change over time. Finally, a third Cox model was fit using LCA class as an ordinal exposure, with unlikely healthcare-acquired as the reference, moderately likely healthcare-acquired as the second level, and likely healthcare-acquired as the third level. The ordinal model included the same covariates listed above and an interaction term between LCA class and log(time) since positive test. The time-varying Cox model was considered as the primary survival analysis due to violation of the proportional hazards assumption in the standard Cox model.
Statistical software and reproducibility
Latent class analyses were conducted using the poLCA package (version 1.6.0) in R version 4.3.1. Logistic and Cox regression models were fit using R version 4.3.1 and Stata version 15.1 (StataCorp, College Station, TX). Code and aggregated data are available from the corresponding author upon reasonable request, subject to data use agreements with provincial data custodians.
Ethics approval
This study was approved by the University of Toronto Health Sciences Research Ethics Board (Protocol #46160). Informed consent was waived because the study used pre-collected anonymized datasets, and neither patients nor the public were involved in the conduct of this research. This study was carried out in accordance with institutional guidelines and the principles of the Declaration of Helsinki. No tissue samples were analyzed, and no experiments were performed.
Results
Study population characteristics
Between March 17, 2020, and September 4, 2022, a total of 53,665 laboratory-confirmed COVID-19 cases were hospitalized in Ontario. After excluding 474 cases with incomplete covariate data, 53,191 cases were included in our analysis. Overall, in-hospital mortality was 16.6% (n = 8,852) (Table 1).
Table 1.
Characteristics of hospitalized COVID-19 cases based on latent class classification
| Characteristics | Unlikely healthcare-acquired | Moderately likely healthcare-acquired | Likely healthcare-acquired | Total hospitalized cases |
|---|---|---|---|---|
| Total | 46,819 | 2687 | 3685 | 53,191 |
| Died, N (%) | ||||
| No | 39,395 (84.1%) | 2062 (76.7%) | 2882 (78.2%) | 44,339 (83.4%) |
| Yes | 7424 (15.9%) | 625 (23.3%) | 803 (21.8%) | 8852 (16.6%) |
| Hospital outbreak-associated, N (%) | ||||
| No | 46,819 (100%) | 412 (15.3%) | 1194 (32.4%) | 48,425 (91.0%) |
| Yes | 0 (0.0%) | 2275 (84.7%) | 2491 (67.6%) | 4766 (9.0%) |
| Mean age (± SD) (years) | 64.4 (21.5) | 73.9 (15.8) | 74.3 (16.3) | 65.5 (21.1) |
| Sex, N (%) | ||||
| Male | 21,724 (46.4%) | 1269 (47.2%) | 1715 (46.5%) | 24,708 (46.5%) |
| Female | 22,821 (48.7%) | 1320 (49.1%) | 1820 (49.4%) | 25,961 (48.8%) |
| Missing | 2274 (4.9%) | 98 (3.6%) | 150 (4.1%) | 2522 (4.7%) |
| Any comorbidities, N (%) | ||||
| No | 32,318 (69.0%) | 1836 (68.3%) | 2135 (57.9%) | 36,289 (68.2%) |
| Yes | 14,501 (31.0%) | 851 (31.7%) | 1550 (42.1%) | 16,902 (31.8%) |
| Immunocompromised, N (%)* | ||||
| No | 44,657 (95.4%) | 2562 (95.3%) | 3540 (96.1%) | 50,759 (95.4%) |
| Yes | 2162 (4.6%) | 125 (4.7%) | 145 (3.9%) | 2432 (4.6%) |
| LTC resident, N (%) | ||||
| No | 44,116 (94.2%) | 2611 (97.2%) | 3581 (97.2%) | 50,308 (94.6%) |
| Yes | 2696 (5.8%) | 75 (2.8%) | 103 (2.8%) | 2874 (5.4%) |
| Missing | 7 (0.0%) | 1 (0.0%) | 1 (0.0%) | 9 (0.0%) |
| Pandemic wave, N (%) | ||||
| 1 | 4455 (9.5%) | 219 (8.2%) | 366 (9.9%) | 5040 (9.5%) |
| 2 | 9057 (19.3%) | 778 (29.0%) | 1135 (30.8%) | 10,970 (20.6%) |
| 3 | 11,083 (23.7%) | 419 (15.6%) | 523 (14.2%) | 12,025 (22.6%) |
| 4 | 4021 (8.6%) | 135 (5.0%) | 170 (4.6%) | 4326 (8.1%) |
| 5 | 9401 (20.1%) | 545 (20.3%) | 766 (20.8%) | 10,712 (20.1%) |
| 6 | 8802 (18.8%) | 591 (22.0%) | 725 (19.7%) | 10,118 (19.0%) |
| Vaccinated (2 + doses), N (%) | ||||
| No | 26,613 (56.8%) | 1529 (56.9%) | 2185 (59.3%) | 30,327 (57.0%) |
| Yes | 20,206 (43.2%) | 1158 (43.1%) | 1500 (40.7%) | 22,864 (43.0%) |
Abbreviations: LTC = long-term care; PCR = polymerase chain reaction; SARS-CoV-2 = severe acute respiratory syndrome coronavirus 2
*Immunocompromised status was captured as a binary variable in the CCM system based on public health unit case investigation. PHU investigators determined immunocompromised status during case interviews, typically based on reported medical history including conditions such as HIV/AIDS, active malignancy, organ or stem cell transplant, primary immunodeficiency, and use of immunosuppressive medications. Specific criteria were not standardized across PHUs, but this classification has demonstrated strong predictive validity for COVID-19 mortality in Ontario surveillance data as noted in the Methods section
Public health units identified 1,838 hospital-associated outbreaks and 17,785 hospital outbreak-linked cases. Hospital outbreak incidence varied substantially across Ontario’s 34 PHUs, ranging from 0.8 to 18.4 outbreaks per 100,000 population (Fig. 1). A timeline of case counts including reported, hospital outbreak-associated, and test-adjusted cases is presented in Fig. 2.
Fig. 1.
Map of Ontario showing cumulative hospital-associated SARS-CoV-2 outbreak count per 100,000 residents by Public Health Unit from March 17, 2020, through September 4, 2022
Fig. 2.
Weekly total reported, test-adjusted, and hospital outbreak-associated SARS-CoV-2 case counts in Ontario, Canada from March 17, 2020, through September 4, 2022. Further details on test-adjustment methods are detailed in Bosco et al. [15]
Latent class analysis: model selection and classification
We compared LCA models with 2, 3, and 4 latent classes. The 3-class solution provided the best fit based on BIC (BIC₂= 86937; BIC₃= 80879; BIC₄= 80941) and yielded clinically interpretable class profiles. Details of 2-, 3-, and 4-latent class model performance metrics are detailed in supplementary Table 1.
The three latent classes showed distinct patterns across the five indicator variables and were labeled as:
Unlikely healthcare-acquired (n = 46,819; 88.0%): Low probability of positive indicators across all five measures. This class predominantly includes patients who tested positive at or shortly after admission, consistent with community acquisition.
Moderately likely healthcare-acquired (n = 2,687; 5.1%): Moderate-to-high probability of shorter-lag indicators (nosocomial flag, outbreak linkage, ≥ 3-day definition) but lower probability of longer-lag indicators (≥ 8-day, ≥ 15-day definitions). This class captures cases diagnosed within the first 1–2 weeks of hospitalization, representing earlier nosocomial acquisition without the requirement of prolonged pre-diagnosis survival.
Likely healthcare-acquired (n = 3,685; 6.9%): High probability of positive indicators across all five measures, including the restrictive ≥ 15-day definition. Patients in this class typically remained hospitalized for at least two weeks before testing positive for SARS-CoV-2, representing definite nosocomial acquisition but requiring survival through extended hospitalization before COVID-19 diagnosis.
Figure 3 presents the intersection of the five indicator variables across all 53,191 hospitalized cases, illustrating the heterogeneity in indicator patterns that the LCA model integrates.
Fig. 3.
Intersection between five indicator variables used in latent class analysis of 53,191 total patients hospitalized with SARS-CoV-2. Orange bars represent the “intersection size”, or the number of patients with each combination of indicator variables indicated by the black dots. The blue bars represent the “set size”, or the number of patients who had a positive result for each single indicator variable
Performance characteristics of individual indicators
Using the LCA classification as the reference standard (moderately likely and likely classes combined as “cases”), we estimated the sensitivity and specificity of each indicator for detecting healthcare-acquired infection (Table 2). All five indicator variables demonstrated high specificity but low sensitivity, which could result in many true nosocomial cases being missed. The indicator variable with the greatest sensitivity was the nosocomial flag in the CCM system. Critically, the ≥ 15-day definition commonly used in strict nosocomial surveillance could miss up to two-thirds of cases identified as healthcare-associated by more lenient measures. Moreover, this definition systematically excludes patients who deteriorate rapidly after acquiring nosocomial infection, as they must survive at least two weeks of hospitalization to meet the case definition. This creates a fundamental bias in any analysis relying solely on late-onset definitions.
Table 2.
Sensitivity and specificity of nosocomial SARS-CoV-2 indicator variables as estimated through latent class analysis
| Indicator variable | Sensitivity | Specificity |
|---|---|---|
| Nosocomial classification in CCM | 0·705 | 1·000 |
| Hospital outbreak-associated | 0·486 | 1·000 |
| Positive test 3 + days after admission | 0·634 | 0·978 |
| Positive test 8 + days after admission | 0·545 | 1·000 |
| Positive test 15 + days after admission | 0·345 | 1·000 |
Abbreviations: CCM = Ontario COVID-19 case and contact management system; SARS-CoV-2 = severe acute respiratory syndrome coronavirus 2
Mortality risk by healthcare-acquisition likelihood
After adjusting for age, sex, comorbidities, immunocompromised status, long-term care residence, vaccination status, and pandemic wave, cases moderately likely to have acquired COVID-19 in hospital had significantly elevated odds of in-hospital death compared to unlikely healthcare-acquired cases (OR: 1.26, 95% CI: 1.14–1.40, p < 0.001) (Table 3). In contrast, cases with high likelihood of healthcare-acquired infection did not show significantly elevated mortality (OR: 1.05, 95% CI: 0.96–1.15, p = 0.25). Results of the mortality analysis restricted to only the unlikely and moderately likely healthcare-acquired groups were consistent with the results from the analysis with all three groups included. Moderately likely healthcare-acquired cases had elevated odds of in-hospital mortality (OR: 1.27, 95% CI: 1.15–1.40, p < 0.001). Full results of the restricted logistic regression model are detailed in supplementary Table 2.
Table 3.
Binomial logistic regression results for odds of in-hospital mortality in COVID-19 patients in Ontario. Likelihood of healthcare-acquisition of SARS-CoV-2 was determined through latent class analysis (LCA)
| Variable | Odds Ratio | 95% Confidence Interval | p-value | |
|---|---|---|---|---|
| LCA Group (ref = unlikely healthcare-acquired) | ||||
| Moderately likely healthcare-acquired | 1.26 | 1.14 | 1.40 | < 0.001 |
| Likely healthcare-acquired | 1.05 | 0.96 | 1.15 | 0.2525 |
| Age | 1.05 | 1.05 | 1.05 | < 0.001 |
| Comorbidities reported (ref = No) | ||||
| Yes | 1.33 | 1.23 | 1.40 | < 0.001 |
| Immunocompromised status (ref = No) | ||||
| Yes | 1.34 | 1.19 | 1.51 | < 0.001 |
| Long-term care resident (ref = No) | ||||
| Yes | 2.23 | 2.04 | 2.43 | < 0.001 |
| Unknown | 0.62 | 0.03 | 3.82 | 0.6644 |
| Sex (ref = Female) | ||||
| Male | 0.99 | 0.95 | 1.05 | 0.8708 |
| Unknown | 1.05 | 0.93 | 1.19 | 0.4410 |
| Vaccination status (ref = No) | ||||
| Yes | 0.98 | 0.87 | 1.10 | 0.1048 |
| Pandemic wave (ref = Wave 1) | ||||
| Wave 2 | 0.89 | 0.87 | 0.97 | 0.0086 |
| Wave 3 | 0.82 | 0.75 | 0.97 | < 0.001 |
| Wave 4 | 0.82 | 0.71 | 0.95 | < 0.001 |
| Wave 5 | 0.63 | 0.54 | 0.72 | < 0.001 |
| Wave 6 | 0.25 | 0.21 | 0.28 | < 0.001 |
Survival analysis
In the Cox proportional hazards model, 42,118 patients were included and a total of 8,832 deaths occurred in this group. Among hospitalized patients, those in the moderately likely healthcare-acquired group had a hazard ratio of 1.04 (95% CI: 0.96–1.14, p = 0.2871), and those in the likely healthcare-acquired group had a hazard ratio of 1.08 (95% CI: 1.00-1.16, p = 0.0426) compared to the unlikely healthcare-acquired group after adjusting for covariates. The proportional hazards assumption was tested, and testing revealed that the assumption was violated, indicating that the hazard ratio changes over time (For LCA class: Chi square = 175.80, p < 0.0001, For global test: Chi square = 598.86, p < 0.0001). Full results of the Cox proportional hazards model are detailed in supplementary Table 3. A second Cox proportional hazards model was fit to allow for a time-varying effect of LCA class on survival. In the time-varying model, patients in the moderately likely healthcare-acquired group had a hazard ratio of 3.25 (95% CI: 2.99–3.52, p < 0.0001), and patients in the likely healthcare-acquired group had a hazard ratio of 4.88 (95% CI: 4.53–5.25, p < 0.0001). Hazard ratio values less than one for both interaction terms of log-transformed time since positive test with moderately likely healthcare-acquired (0.71, 95% CI: 0.68–0.73) and likely healthcare-acquired (0.58, 95% CI: 0.56–0.60) indicate that the relative hazard decreases over time (Fig. 4). Full results of the time-varying Cox proportional hazards model are detailed in Table 4. The time-varying Cox model treating LCA class as an ordinal variable resulted in a hazard ratio of 2.26 (95% CI: 2.19–2.34, p < 0.0001), indicating that each step increase in likelihood of nosocomial infection was associated with a 2.26-fold increase in hazard of death, suggesting a graded association, though not strictly linear across categories. Full results of the time-varying Cox proportional hazards model using LCA class as an ordinal exposure are detailed in supplementary Table 4.
Fig. 4.
Time-varying hazard ratios for in-hospital mortality comparing patients with moderate and high likelihood of healthcare-acquired SARS-CoV-2 infection to those with unlikely healthcare-acquired infection. Hazard ratios are plotted as a function of days since first positive test, with shaded areas representing 95% confidence intervals. Estimates are derived from a Cox proportional hazards model with time-varying effects (interaction between LCA class and log-transformed time), adjusted for age, sex, comorbidity status, immunocompromised status, long-term care residence, vaccination status, and pandemic wave
Table 4.
Cox proportional hazards regression results for survival of hospitalized COVID-19 patients in Ontario
| Variable | Hazard Ratio | 95% Confidence Interval | p-value | |
|---|---|---|---|---|
| LCA Group (ref = unlikely healthcare-acquired) | ||||
| Moderately likely healthcare-acquired | 3.25 | 2.99 | 3.52 | < 0.0001 |
| Likely healthcare-acquired | 4.88 | 4.53 | 5.25 | < 0.0001 |
| Age | 1.04 | 1.04 | 1.04 | < 0.0001 |
| Comorbidities reported (ref = No) | ||||
| Yes | 1.36 | 1.30 | 1.42 | < 0.0001 |
| Immunocompromised status (ref = No) | ||||
| Yes | 1.15 | 1.05 | 1.26 | < 0.0001 |
| Long-term care resident (ref = No) | ||||
| Yes | 1.70 | 1.59 | 1.81 | < 0.0001 |
| Unknown | 0.66 | 0.03 | 4.27 | 0.9383 |
| Sex (ref = Female) | ||||
| Male | 0.99 | 0.95 | 1.03 | 0.5714 |
| Unknown | 1.06 | 0.96 | 1.17 | 0.2588 |
| Vaccination status (ref = No) | ||||
| Yes | 0.96 | 0.92 | 1.01 | 0.0896 |
| Pandemic wave (ref = Wave 1) | ||||
| Wave 2 | 1.00 | 0.95 | 1.05 | 0.9802 |
| Wave 3 | 0.93 | 0.88 | 0.98 | 0.0068 |
| Wave 4 | 1.00 | 0.93 | 1.09 | 0.8936 |
| Wave 5 | 0.80 | 0.76 | 0.84 | < 0.0001 |
| Wave 6 | 0.37 | 0.35 | 0.40 | < 0.0001 |
| LCA Group * log(time) (ref = unlikely healthcare-acquired) | ||||
| Moderately likely healthcare-acquired * log(time) | 0.71 | 0.68 | 0.73 | < 0.0001 |
| Likely healthcare-acquired * log(time) | 0.58 | 0.56 | 0.60 | < 0.0001 |
Likelihood of healthcare-acquisition of SARS-CoV-2 was determined through latent class analysis (LCA) and included with time-varying effects in this model. Interaction terms represent the change in hazard ratio associated with log-transformed time since positive test. Interaction term values < 1 indicate that the relative hazard decreases over time
Discussion
In this population-based study of hospitalized COVID-19 cases in Ontario, we found that nosocomial SARS-CoV-2 infection is not only challenging to evaluate using single-indicator measures but also has important implications for patient outcomes. Patients moderately likely to have acquired COVID-19 in-hospital had significantly higher odds of in-hospital death than likely community-acquired cases among individuals sick enough to require hospitalization. This finding runs counter to the expectation that patients hospitalized for COVID-19 would represent those most at risk of mortality from the virus. Instead, our results suggest that acquiring COVID-19 during hospitalization confers even greater risk, presumably due to multiplicative risk of the infection in combination with the health issue that initiated the hospitalization. Patients classified as likely to have acquired SARS-CoV-2 did not have significantly elevated odds of in-hospital death compared to unlikely healthcare-acquired cases. This counterintuitive pattern is consistent with survivorship bias: patients classified as “likely healthcare-acquired” must, by definition, have survived hospitalization long enough (typically at least two weeks) to acquire and be diagnosed with SARS-CoV-2 before discharge or death. Patients who are admitted with severe underlying conditions and deteriorate rapidly within the first 1–2 weeks are systematically excluded from this class, regardless of whether they acquire nosocomial COVID-19. This conditioning on prolonged survival before diagnosis creates a selected cohort enriched for patients with better prognosis for their index admission diagnosis, offsetting the additional mortality risk associated with superimposed COVID-19. The moderately likely group, which includes fewer restrictive indicators, avoids this survivorship bias. These patients have not been selected for the ability to survive weeks of hospitalization, and therefore this group reveals the true mortality burden of hospital-acquired infection among a more representative sample of exposed patients.
In contrast to the logistic regression results, which reflect cumulative in-hospital mortality risk, the time-to-event analyses revealed important temporal dynamics in the relationship between nosocomial infection likelihood and mortality. The standard Cox proportional hazards model showed minimal differences in survival between LCA classes, but this model violated the proportional hazards assumption, indicating that the effect of nosocomial infection likelihood on mortality was not constant over time. When allowing for time-varying effects, substantially higher hazard ratios were observed for the moderately likely and likely healthcare-acquired groups, with evidence that these effects were strongest soon after the patient’s first positive test. This pattern suggests that the impact of healthcare-associated infection status on mortality is most pronounced in the early period following diagnosis and diminishes with longer survival. This may reflect differences in patient characteristics, disease severity, or other vulnerabilities that are most pronounced in the acute phase of the infection. The apparent discrepancy between the logistic and time-varying Cox results reflects differences in what each model estimates. Logistic regression captures overall risk across the entire hospitalization and is therefore more sensitive to survivorship bias embedded in the LCA classification, particularly for the “likely” group, which requires prolonged pre-diagnosis survival. In contrast, the time-varying Cox model isolates the instantaneous hazard of death following diagnosis, revealing an elevated early risk in both the moderately likely and likely healthcare-acquired groups which is obscured when mortality outcomes are aggregated over time. Importantly, these hazard ratios should not be interpreted purely as causal effects of nosocomial acquisition, as they also reflect differences in underlying patient trajectories, including severity of illness, frailty, functional status, timing of infection, and selection processes inherent in the construction of the LCA classes. These findings illustrate how different analytic approaches can capture different aspects of the same processes, including cumulative risk, average time-to-event effects, and dynamic changes in hazards over time.
The reported relative severity of nosocomial COVID-19 has varied across studies [4–6]. The COPE-Nosocomial study, for example, reported lower adjusted mortality among hospital-onset cases using time-to-event analysis. However, that study applied a highly restrictive definition of hospital-acquired infection, requiring that diagnosis occur more than 14 days after admission. Our own data show that this definition yields perfect specificity but a sensitivity of only 34%, meaning it misses many true nosocomial cases. In addition to low sensitivity, this approach introduces survivorship bias by conditioning case classification on remaining in hospital for at least 15 days without diagnosis or death. Patients who acquire COVID-19 during hospitalization but deteriorate rapidly and die before day 15 are systematically excluded from the nosocomial group, resulting in an artificially low-risk cohort and downwardly biased estimates of mortality.
Other frequently applied surveillance cutoffs, such as three- or eight-days post-admission, present different trade-offs: the three-day rule has moderate sensitivity, while the eight-day rule improves specificity at the cost of lower sensitivity. These findings underscore the limitations of single-indicator definitions and the risk of misclassifying cases in a way that would underestimate the relative severity of hospital-acquired infection by biasing effects toward the null.
In contrast, our study uses latent class modeling to probabilistically classify nosocomial infections based on multiple indicators. This approach avoids the sensitivity and specificity pitfalls of fixed cutoffs and captures a fuller spectrum of hospital-acquired infections, while highlighting the complexities of evaluating mortality risk using specific surveillance cutoffs. The resulting classification reveals the potential for elevated mortality risk corresponding to the likelihood of nosocomial acquisition, especially in the immediate period following infection, suggesting that hospital-acquired COVID-19 may be more deadly but also more common than strict definitions imply. Our findings are consistent with observations by Melançon and Ponsford [4, 5]. However, the Melançon analysis was limited to a single center and stratified results by age without extensive adjustment for other confounders, while the Ponsford study relied on hospital-level aggregate data and time-based definitions without individual-level multivariable control. By comparison, our study integrates individual-level data with robust adjustment and refined classification, providing stronger evidence that nosocomial infection contributes directly to poor outcomes, rather than simply reflecting underlying vulnerability.
These findings point to the need for a fundamental rethinking of how in-hospital transmission is measured, interpreted, and prevented. Case definitions that emphasize specificity over sensitivity (e.g., 14-day cut-off) may simplify reporting but fail to reflect the true scope and consequences of nosocomial infection. When most hospital-acquired cases go undetected, both surveillance systems and institutional responses are weakened. Underestimating this burden not only distorts our understanding of disease dynamics but also reduces accountability for preventable harm within healthcare settings.
This study has several limitations. First, as with all observational research, we cannot exclude the possibility of residual confounding from unmeasured patient- and hospital-level factors, despite adjusting for a wide range of clinical and demographic covariates informed by their causal relationships with both the exposure and the outcome. Second, while our use of latent class modeling offers a more nuanced and probabilistic approach to nosocomial classification, it depends on the quality and completeness of input indicators, which may vary across institutions or over time. Third, our findings are based on acute care hospitals in Ontario and may not generalize to settings with different infection control practices, healthcare infrastructure, or patient characteristics. Variability in testing availability and surveillance systems, particularly in lower-resource environments, may influence how healthcare-acquired infections are identified and measured. Finally, it is important to recognize that some hospital-acquired COVID-19 cases may not be tested during their hospitalization or may only be tested after discharge from the hospital. Despite these constraints, we believe that carefully constructed analyses using the best available data remain essential for guiding policy, particularly in rapidly evolving or resource-limited contexts.
Defining nosocomial SARS-CoV-2 infection using single indicators remains challenging and can lead to misclassification, particularly when relying on rigid time-based cutoff thresholds that trade sensitivity for specificity. Our findings highlight how these approaches may underestimate the true burden and severity of hospital-acquired infections. By integrating multiple indicators, latent class modeling provides a more comprehensive approach for identifying nosocomial cases and reveals elevated mortality risk in some groups. These results challenge common assumptions about the role and risk of hospital-acquired SARS-CoV-2 infection. Consistent with prior work, we found that these infections may be more deadly for patients than community-acquired infections among hospitalized patients [4, 5]. Addressing this challenge requires both conceptual and practical shifts: moving beyond static case definitions and embracing nosocomial respiratory viral infection prevention as a core pillar of healthcare safety. Recognizing and responding to the full impact of hospital-acquired infections, both today and in future pandemic scenarios, is essential for protecting individuals, strengthening health systems, and reducing population-level risk.
Electronic Supplementary Material
Below is the link to the electronic supplementary material.
Acknowledgements
Not applicable.
Author contributions
NJW and DNF designed the study. DNF conceived the work and acquired the data. NJW, AAG, CEL, and DNF contributed to analysis and interpretation of data and drafted and/or substantively revised the manuscript. All authors read and approved the final manuscript and take responsibility for the submitted work.
Funding
This project was supported by a Canadian Institutes for Health Research (CIHR) project grant (#518192). The funder had no role in study design, data collection, data analysis, data interpretation, or writing of the manuscript.
Data availability
The datasets used and/or analysed during the current study are not publicly available due to privacy and/or data-sharing restrictions, but may be available from the corresponding author on reasonable request and with any required institutional approvals and permissions.
Declarations
Ethics approval and consent to participate
This study was approved by the University of Toronto Health Sciences Research Ethics Board (Protocol #46160). The requirement for informed consent was waived by the Research Ethics Board because the study used pre-collected/secondary data, and neither patients nor the public were involved in the conduct of this research. This study was carried out in accordance with institutional guidelines and the principles of the Declaration of Helsinki. No tissue samples were analyzed, and no experiments were performed.
Consent for publication
Not applicable.
Competing interests
DNF has served on advisory boards related to influenza and SARS-CoV-2 vaccines for Seqirus, Pfizer, Astrazeneca and Sanofi-Pasteur Vaccines, and has served as a legal expert on issues related to COVID-19 epidemiology for the Elementary Teachers Federation of Ontario and the Registered Nurses Association of Ontario. NJW, AAG and CEL: no competing interests.
Footnotes
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
References
- 1.Lumley SF, Constantinides B, Sanderson N, Rodger G, Street TL, Swann J, et al. Epidemiological data and genome sequencing reveals that nosocomial transmission of SARS-CoV-2 is underestimated and mostly mediated by a small number of highly infectious individuals. J Infect. 2021;83:473–82. 10.1016/J.JINF.2021.07.034. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Mo Y, Eyre DW, Lumley SF, Walker TM, Shaw RH, O’Donnell D, et al. Transmission of community- and hospital-acquired SARS-CoV-2 in hospital settings in the UK: a cohort study. PLoS Med. 2021;18. 10.1371/JOURNAL.PMED.1003816. [DOI] [PMC free article] [PubMed]
- 3.Cooper BS, Evans S, Jafari Y, Pham TM, Mo Y, Lim C, et al. The burden and dynamics of hospital-acquired SARS-CoV-2 in England. Nat 2023. 2023;623:7985. 10.1038/s41586-023-06634-z. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Melançon E, Brosseau M, Bartoli A, Labbé AC, Lavallée C, Marchand-Senécal X, et al. Outcomes of hospital-acquired SARS-CoV-2 infection in the Canadian first wave epicentre: a retrospective cohort study. Can Med Association Open Access J. 2022;10:E74–81. 10.9778/CMAJO.20210055. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Ponsford MJ, Jefferies R, Davies C, Farewell D, Humphreys IR, Jolles S, et al. Burden of nosocomial COVID-19 in Wales: results from a multicentre retrospective observational study of 2508 hospitalised adults. Thorax. 2021;76:1246–9. 10.1136/THORAXJNL-2021-216964. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Carter B, Collins JT, Barlow-Pay F, Rickard F, Bruce E, Verduri A, et al. Nosocomial COVID-19 infection: examining the risk of mortality. The COPE-Nosocomial Study (COVID in Older PEople). J Hosp Infect. 2020;106:376–84. 10.1016/J.JHIN.2020.07.013. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Fisman DN, Greer AL, Brankston G, Hillmer M, O’Brien SF, Drews SJ, et al. COVID-19 Case Age Distribution: Correction for Differential Testing by Age. Ann Intern Med. 2021;174:1430–8. 10.7326/M20-7003/SUPPL_FILE/M20-7003_SUPPLEMENT.PDF. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Fisman DN, Tuite AR. Evaluation of the relative virulence of novel SARS-CoV-2 variants: a retrospective cohort study in Ontario, Canada. CMAJ. 2021;193:E1619–25. 10.1503/CMAJ.211248. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Aflaki K, Vigod S, Ray JG, Part I. A friendly introduction to latent class analysis. J Clin Epidemiol. 2022;147:168–70. 10.1016/j.jclinepi.2022.05.008. [DOI] [PubMed] [Google Scholar]
- 10.Kitsios GD, Blacka S, Jacobs JJ, Mirza T, Naqvi A, Gentry H, et al. Subphenotypes of self-reported symptoms and outcomes in long COVID: a prospective cohort study with latent class analysis. BMJ Open. 2024;14. 10.1136/BMJOPEN-2023-077869. [DOI] [PMC free article] [PubMed]
- 11.Sigwadhi LN, Tamuzi JL, Zemlin AE, Chapanduka ZC, Allwood BW, Koegelenberg CF, et al. Latent class analysis: an innovative approach for identification of clinical and laboratory markers of disease severity among COVID-19 patients admitted to the intensive care unit. IJID Reg. 2022;5:154–62. 10.1016/J.IJREGI.2022.10.004. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Ontario Ministry of Health. Ontario public health standards: requirements for programs, services and accountability infectious disease protocol severe acute respiratory syndrome (SARS) and Middle East Respiratory Syndrome (MERS). 2023.
- 13.Fisman DN, Greer AL, Hillmer M, Tuite R. Derivation and validation of clinical prediction rules for COVID-19 mortality in Ontario, Canada. Open Forum Infect Dis. 2020;7. 10.1093/OFID/OFAA463. [DOI] [PMC free article] [PubMed]
- 14.Mitchell R, Cayen J, Thampi N, Frenette C, Bartoskzo J, Choi KB, et al. Trends in Severe Outcomes Among Adult and Pediatric Patients Hospitalized With COVID-19 in the Canadian Nosocomial Infection Surveillance Program, March 2020 to May 2022. JAMA Netw Open. 2023;6:e239050–239050. 10.1001/JAMANETWORKOPEN.2023.9050. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Bosco S, Peng A, Tuite AR, Simmons A, Fisman DN. Impact of adjustment for differential testing by age and sex on apparent epidemiology of SARS-CoV- 2 infection in Ontario, Canada. BMC Infect Dis 2025. 2025;25:1. 10.1186/S12879-025-10968-6. [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
The datasets used and/or analysed during the current study are not publicly available due to privacy and/or data-sharing restrictions, but may be available from the corresponding author on reasonable request and with any required institutional approvals and permissions.