Predictors of Mortality Among Long-term Care Residents With SARS-CoV-2 Infection

Douglas S. Lee MD, PhD; Shihao Ma BASc; Anna Chu MHSc; Chloe X. Wang BSc; Xuesong Wang MSc; Peter C. Austin PhD; Finlay A. McAlister MD, MSc; Sunil V. Kalmady PhD; Moira K. Kapral MD, MSc; Padma Kaul PhD; Dennis T. Ko MD, MSc; Paula A. Rochon MD, MPH; Michael J. Schull MD, MSc; Barry B. Rubin MD, PhD; Bo Wang PhD


J Am Geriatr Soc. 2021;69(12):3377-3388. 

In This Article


Study Cohort

Among 84,142 residents of LTC homes, 64,731 (median age 86 [78,91] years, 31.8% men) were tested for SARS-CoV-2 with 5029 (7.8%) testing positive prior to death, of whom 1442 (28.7%) died within 30 days of the initial SARS-CoV-2 positive PCR test. Of the 59,702 residents who tested negative, 2652 (4.4%) died within 30 days of testing. There was complete follow-up on all of the LTC study cohort. A study flow diagram is shown in Figure S1. In total, 304 variables were considered in our models (Table S3). Baseline characteristics of the study cohort, comprised of 5029 test positive persons, are shown in Table S4.

Mortality Predictors

The top 50 (or 47) characteristics associated with 30-day mortality after testing positive for SARS-CoV-2 are shown in Table 1, and importance scores are shown in Figure 1. Among these (excluding 3 dummy predictors, i.e., multiple subcategories for the same variable), 19 (40.4%) were laboratory features, 11 (23.4%) were functional characteristics, 11 (23.4%) were comorbidities, 3 (6.4%) were demographic, 2 (4.3%) were LTC-related features, and 1 (2.1%) a community characteristic. Variables with an indeterminate pattern of association with 30-day mortality included triiodothyronine, thyroxine, c-reactive protein, ionized Ca2+, transferrin, ferritin, serum iron, and the interRAI domain of limitation with communication (i.e., inability to understand or express).

Figure 1.

Variable importance scores. Using the VIMP function, the top 47 variables are shown in descending order of predictive ability

The top predictors of mortality included the Changes in Health, End-Stage Disease and Symptoms and Signs (CHESS) Scale (a measure of health stability with higher scores representing greater instability), functional measures of activities of daily living (ADLs), lower cognitive performance, lower levels of social engagement, and either the presence of or risk of developing pressure ulcers (Table 1). Functional variables comprised 6 of the top 10 predictors of mortality. Requiring or benefitting from clinical interventions to prevent a decline in ADLs or cognitive status in interRAI was associated with increased risk of mortality. Comorbidities associated with mortality included time since coronary revascularization and duration of heart failure, dementia, and chronic obstructive pulmonary disease (COPD). Older age (see Figure 2) and male sex were predictors of mortality to 30 days. As duration of these conditions increased, there was lower survival (see Figures S3 to S6). Other comorbid factors included prior emergency department visits or hospitalizations for respiratory disease, urinary incontinence, undernutrition risk, and dehydration. Larger LTC facility size and longer duration of residence were both associated with lower mortality while larger community size was associated with higher mortality. Laboratory features associated with higher mortality included lower GFR, lower cholesterol, lower hemoglobin, lower lymphocyte count, lower platelet count, abnormal iron indices, and lower albumin.

Figure 2.

AUC curves for 30-day mortality models with varying covariate categories

Model Performance

The final full model included 304 variables including all demographic, interRAI functional, LTC-related, and community characteristics, along with comorbidities and laboratory tests (AUC in test set 0.701, 95% CI: 0.666, 0.736). In Figure 2, AUC curves are shown for: (a) base model only including age, sex, and other demographic variables, (b) base model + community characteristics, (c) base model + comorbidities, (d) base model + laboratory tests, (e) base model + functional status measures, and (f) final full model. Inclusion of comorbidities, laboratory tests, and functional status information improved the model AUCs for prediction of 30-day mortality (Figure 2). Excluding community characteristics from the full model resulted in an AUC of 0.692 (95% CI: 0.658, 0.727). Including the time-dependent comorbidities (i.e., heart failure, dementia, COPD, and time since coronary revascularization, shown in Figures S3 to S6) as binary variables (i.e., present vs. absent) decreased the AUC slightly to 0.697 (95% CI: 0.663, 0.734).

Using the final machine learning model to stratify LTC residents into risk quartiles, we found that those in the highest quartile had a greater than a 12-fold risk of 30-day death (48.3%) compared with those in the lowest quartile (Table 2). The top 10 predictors of mortality in a parsimonious model (Table S5) exhibited only slightly inferior discrimination than the full model with an AUC of 0.695 (95% CI: 0.659, 0.730). While the highest risk quartile in the parsimonious model exhibited 48.1% mortality, this simpler model was suboptimal for identifying a low risk quartile who had 12% mortality, compared with 7% in the original full model (Table 2). Cumulative incidence curves for 30-day mortality stratified by mortality risk quartile (full cohort) showed early separation of the curves with accelerated mortality rates up to 15 days, and then a leveling-off from 16 to 30 days after a positive SARS-CoV-2 test (Figure 3).

Figure 3.

Cumulative incidence of death by predicted risk quartile. Variables allowing small cell size determinations are suppressed (supp)