Geriatrics-focused indicators predict mortality more than age in older adults hospitalized with COVID-19

Background Age has been implicated as the main risk factor for COVID-19-related mortality. Our objective was to utilize administrative data to build an explanatory model accounting for geriatrics-focused indicators to predict mortality in hospitalized older adults with COVID-19. Methods Retrospective cohort study of adults age 65 and older (N = 4783) hospitalized with COVID-19 in the greater New York metropolitan area between 3/1/20-4/20/20. Data included patient demographics and clinical presentation. Stepwise logistic regression with Akaike Information Criterion minimization was used. Results The average age was 77.4 (SD = 8.4), 55.9% were male, 20.3% were African American, and 15.0% were Hispanic. In multivariable analysis, male sex (adjusted odds ration (adjOR) = 1.06, 95% CI:1.03-1.09); Asian race (adjOR = 1.08, CI:1.03-1.13); history of chronic kidney disease (adjOR = 1.05, CI:1.01-1.09) and interstitial lung disease (adjOR = 1.35, CI:1.28-1.42); low or normal body mass index (adjOR:1.03, CI:1.00-1.07); higher comorbidity index (adjOR = 1.01, CI:1.01-1.02); admission from a facility (adjOR = 1.14, CI:1.09-1.20); and mechanical ventilation (adjOR = 1.52, CI:1.43-1.62) were associated with mortality. While age was not an independent predictor of mortality, increasing age (centered at 65) interacted with hypertension (adjOR = 1.02, CI:0.98-1.07, reducing by a factor of 0.96 every 10 years); early Do-Not-Resuscitate (DNR, life-sustaining treatment preferences) (adjOR = 1.38, CI:1.22-1.57, reducing by a factor of 0.92 every 10 years); and severe illness on admission (at 65, adjOR = 1.47, CI:1.40-1.54, reducing by a factor of 0.96 every 10 years). Conclusion Our findings highlight that residence prior to admission, early DNR, and acute illness severity are important predictors of mortality in hospitalized older adults with COVID-19. Readily available administrative geriatrics-focused indicators that go beyond age can be utilized when considering prognosis. Supplementary Information The online version contains supplementary material available at 10.1186/s12877-021-02527-w.

implicated as one of the main risk factors for poor outcomes in COVID-19, including severe disease, hospitalization, and mortality [3][4][5][6]. While studies evaluating COVID-19 related mortality account for demographic and clinical characteristics, few have considered indicators that are highly relevant to the older adult population, namely functional and cognitive status as well as baseline advance directives regarding life-sustaining treatment preferences.
Although age in general contributes to models predicting mortality, there is robust literature demonstrating that function and frailty are strong predictors of poor outcomes among older adults [7][8][9]. This has also been shown in recent studies evaluating frailty in patients with COVID-19 [10,11]. Unfortunately, functional status and frailty are infrequently incorporated into models due to insufficient assessment and documentation by medical professionals [12]. Another critical, yet frequently unaddressed and undocumented factor in older adults is advance directives, including goals of care and life-sustaining treatment preferences [13]. Early Do-Not-Resuscitate (DNR), documented within 24 h of admission, has been associated with baseline care preferences and prognosis that may not be related to the patient's acute illness, and has been shown to be an independent risk factor for mortality [14][15][16]. Therefore, an early DNR can be used as a proxy for mortality risk. However, COVID-19 studies evaluating mortality have generally either excluded or failed to account for patients with DNR documentation on admission [3][4][5].
As COVID-19 continues to impact older adults globally, concerns about present and future shortages in hospital and critical care beds, staffing, and treatments (e.g. ventilators and vaccines), have focused attention on resource allocation. Decisions about resource allocation during COVID-19 first started in Northern Italy, where the number of sick patients grossly outnumbered available health care resources [17]. Since then, many countries and states have prepared or activated pre-existing crisis standards of care. Some of these strategies highlight older age as the main criteria, placing older adults at a disadvantage regardless of their baseline characteristics [18][19][20][21][22][23][24][25]. In a position statement regarding resource allocation during COVID-19, the American Geriatrics Society (AGS) indicated that age should never be used as the main indicator to exclude older adults from receiving care [26].
The objective of this study was to utilize readily available administrative data that incorporates geriatricsfocused indicators to build an exploratory model to predict mortality in hospitalized older adults during the initial peak of the COVID-19 pandemic.

Methods
The study was conducted at a large academic health system serving approximately 11 million people in the greater New York metropolitan area. The COVID-19 Research Consortium and institutional review board approved the study. Data was abstracted from inpatient electronic medical records with the following inclusion criteria: age greater than or equal to 65; admitted to one of eleven health system hospitals between March 1st, 2020 and April 20th, 2020; and confirmed diagnosis of COVID-19 infection by positive result on polymerase chain reaction of a nasopharyngeal sample. Clinical course and outcomes were monitored until June 10th, 2020. Transfers from one in-system hospital to another were merged and considered as a single visit. For patients with multiple admissions during the study period, only the first admission was included. Data were collected from enterprise electronic health record (EHR) (Sunrise Clinical Manager; Allscripts). De-identified data can be provided upon request.

Data elements
All data elements underwent a vigorous process (Sinvani, Qiu, Yan, Marziliano, Makhnevich) of data harmonization-ensuring that the data fields represent clinically relevant information. An iterative process was undertaken that consisted of a recurring loop of query, refinement, data validation by manually reviewing randomly selected charts, and inter-rater agreement, followed by further query refinement, until variable accuracy was achieved. In addition, the team hand-searched a random selection of charts to ensure data quality and integrity. Clinical members of the research team led quality control and data validation of clinical variables including comorbidity index, baseline functional status, preferences for life-sustaining treatments and discharge disposition, Modified Early Warning System (MEWS), and oxygen requirement (including mechanical ventilation).
Patient demographics included age, sex, race, ethnicity, insurance, and body mass index (BMI); all elements were captured from pre-specified categories in the EHR. BMI was categorized as underweight (< 18.5), normal (18.5 to 24.9), overweight (25 to 29.9), and obese (≥ 30). Comorbid conditions were collected using ICD-9/10 codes based on past medical history documented prior to or at index admission, including: diabetes mellitus, hypertension, chronic obstructive pulmonary disease, interstitial lung disease, asthma, cancer, coronary artery disease, atrial fibrillation, chronic kidney disease, and dementia. Comorbidity index was calculated based on the Charlson Comorbidity Index (CCI), excluding the age component [27]. Residing in a facility prior to admission was used as a surrogate indicator for functional status or the need for assistance in activities of daily living. Baseline functional status was coded as "residence prior to admission" home versus any facility (skilled nursing facility, SNF, assisted living, and group home). Further details on type of facility and level of assistance provided (e.g. long-term care vs. sub-acute rehabilitation) is not consistently documented in most EMR and was not available.
Patient preferences for life-sustaining treatments on admission were based on the presence of an "early" (within 24 h of admission) do-not-resucssitate (DNR) order. Early DNR has been described in the literature as a surrogate for worse baseline prognosis as well as preferences for life-sustaining treatment and has been shown to be an independent predictor of mortality [14][15][16].
Illness acuity on presentation consisted of the first documented MEWS, first documented method of oxygen delivery (no oxygen, nasal cannula, venturi mask, nonrebreather mask, high flow, noninvasive ventilation, and mechanical ventilation), and temperature on arrival (no fever vs. fever). MEWS is pre-calculated in the EHR and includes respiratory rate (breaths/min), oxygen saturation, temperature, systolic blood pressure (SBP), heart rate (HR), and level of consciousness [28]. MEWS was categorized as non-critically ill (0-4) and critically-ill (> 4) [29][30][31]. Certain methods of oxygen delivery (venturi masks, high flow oxygen, noninvasive mechanical ventilation) were discouraged within the health system to minimize risk of aerosolized spread and so were rarely used. They were categorized as follows: venturi mask (n = 30, 0.6%) was grouped with nasal cannula, and high flow oxygen (n = 16, 0.3%) and non-invasive ventilation (n = 26, 0.5%) with nonrebreather masks. Fever on arrival was defined as temperature of > 37.8 degrees Celsius.
The primary outcome of interest was hospital mortality. There were no missing values for the majority of the variables including the outcome variable. The only variables with a significant number of missing values used in the analysis were BMI (30% missing) and MEWS (9% missing).

Analysis
Means and standard deviation (SD) were used to summarize distributions of continuous variables and frequencies and percentages were used for categorical variables. Patients who were still hospitalized at time of data truncation (4/20/20) or transferred to a facility in which data for their complete hospital course could not be obtained, and those with a missing comorbidity index or residence prior to admission were excluded from the unadjusted and adusted analyses (Fig. 1).
The objective was to build an explanatory model with good predictive properties that helps to explain the impact of predictors of mortality. Initial analysis was conducted with a series of simple logistic regression analyses to evaluate patient factors related to hospital mortality. These factors included: age, sex, race, ethnicity, insurance, individual comorbidities, BMI, comorbidity index (total comorbidity index excluding age contribution), residence prior to admission -home vs. facility (as a surrogate for baseline functional status), DNR order within 24 h of admission (as a surrogate for preferences for declining life-sustaining treatment), and severity of acute illness (first documented MEWS and oxygen delivery method).
Stepwise logistic regression with Akaiike Information Criterion (AIC) minimizationwas used to choose the parcimonious model. Stepwise variable selection was then applied to the entire data set of N = 4783 (only two variables BMI and MEWS had missing data: 1438 missing BMI; 473 missing MEWS). Two multiple imputation models minimizing AIC were found: (1) a multinomial regression model for missing BMI (4 BMI categories: underweight, normal, overweight, obese) and (2) a logistic regression model for missing MEWS (2 categories: non-critically ill and critically ill). Using multiple imputations, 50 imputed data sets were created. Then, Rubin's approach was used to combine results from the 50 logistic regression analyses [32].
Modelling was expanded to test interactions with two linear predictors (age and comorbidity index). Multicollinearity was evaluated using variance inflation factor (VIF). Analyses were performed using SPSS v26, IBM Corp, Armonk, NY, and SAS 9.4, SAS Institute Inc., Cary, NC.; R 4.0.3 (www.p-proje ct. com) was used to apply machine learning methods.

Hospital outcomes
The median length of stay was 7 days (mean 9.9, SD = 9.5), 36.3% expired, 42.1.6% were discharged home, and 21.6% were discharged to a facility. See Table 1 for details on hospital presentation and outcomes.

Univariate analysis (N = 4783)
Each of the following patient characteristics were related to mortality in the univariate analysis (  all significantly related to mortality. In contrast to previously published series, race, ethnicity, insurance status, diabetes mellitus, and BMI were not significantly related to mortality [3,4]. Additional univariable analyses of age and the other patient characteristics can be found in Supplemental Table 1.

Sensitivity analyses
In logistic regression analysis, the direction and significance of the association with mortality did not change when missing data on BMI and MEWS were treated as missing data categories.

Discussion
Our study found that in hospitalized older adults with COVID-19 infection, male sex, Asian race, admission from a facility, presence of an early DNR, higher multimorbidity and illness severity were predictive of increased mortality. Our study found that sex, multimorbidity, residence prior to admission, early DNR, and illness severity were predictive of hospital mortality. While age was not an independent predictor of mortality, we observed several interactions of age with factors predicting mortality: increasing age was negatively associated with mortality in patients with hypertension and early DNR, and age was positively associated with mortality for  patients residing at home prior to admission and those that were not critically-ill. In other words, the impact of age on mortality is minimal in the setting of severely critical disease and other factors indicating highly impaired function. These findings highlight the importance of considering readily available geriatrics-focused indicators, other than age, when considering prognostication and management, especially in the context of a pandemic when medical resources are in danger of being overwhelmed. Therefore, finding predictors other than age would be beneficial in preventing ageism. There is robust literature demonstrating that baseline functional status is strongly associated with poor outcomes among older adults [7,8]. However, baseline functional status is infrequently assessed at hospital admission. Medical professionals often rely on subjective impressions of function, which have been shown to be inaccurate [12]. Furthermore, even when assessed, functional status is rarely documented in a meaningful or consistent fashion [12]. While there is increasing evidence demonstrating an association between frailty and poor outcomes in those with COVID-19 [10,11,33], these measures may not be readily available on hospital presentation nor captured in a systematic fashion in most EHRs (especially if no previous admissions exist) [34], as is the case in our health system. During the first surge of the COVID-19 crisis, health systems may utilize shortened, "emergency documentation" protocols, which do not include functional measures such as activities of daily living. Truncated protocols are often used in emergency situations when large volumes of patients are admitted within a short period of time. Administrative data based frailty scores may also be used, but have shown only fair to moderate agreement with clinical frailty scales [35].
Our study utilized the 'residence prior to admission' variable as a surrogate marker of the need for skilled care-baseline functional status (arriving from a facility prior to admission). Although we were unable to distinguish between those who were "life-stay" residents in long-term care versus limited-stay for sub-acute rehabilitation, all of these locations house populations who generally have functional impairments and/or lack adequate support in the home setting. Our findings differ from a previous study from Spain, which demonstrated that retirement facilities were not associated with mortality during the first wave in adjusted models. However, the total number of COVID cases reported in this series was small, 239, just over 60% were over age 60 (~ 145), and only 26 resided in a retirement facility, suggesting this cohort would be underpowered to detect such differences. While it is important for future studies to evaluate baseline functional status using more precise, objective measures, this is often difficult in the real-world setting. Therefore with further validation, 'residence prior to admission' may be a good surrogate marker that can be easily obtained. Assessing and incorporating functional status is essential when evaluating clinical outcomes in older adults.
Another critical factor that must be accounted for is advance directives, including discussion and documentation of life-sustaining treatment preferences and DNR orders. Provider predictions about life expectancy typically do not match the patient's treatment preferences [13,36]. Although age has traditionally been used as a "proxy" for life expectancy, older adults of the same age can have very different trajectories. While advance directives should be discussed openly, this discussion and decision should not be imposed or pressured, which is most likely to happen in an acute, emergent setting. As such, timely, patient-centered goals of care discussions that focus on what matters most to the patient are essential [37]. A recent study by Alhatem et al., demonstrated that DNR status prior to admission determines mortality in patients with COVID-19 [38]. There was a negative interaction between age and early DNR, in that early DNR was less predictive of mortality in older age. This suggests that early DNR in younger age indicates overall poor health, whereas, early DNR in older age indicates wishes that may be independent of overall health. In response to treatment decisions and resource allocation during COVID-19, several strategies, guidelines, and frameworks have been proposed, most of which highlight age as the main factor in making resource allocation determinations [18][19][20][21][22][23][24][25]39]. While a few of these guidelines recommend advanced age as a sole exclusion, the majority include age along with other criteria such as comorbid conditions, functional status, advance directives (specifically preferences for life-sustaining interventions), and severity of acute illness [21][22][23][24]39]. However, due to the limited ability to account for functional status and advance directives, age often becomes the de facto factor used for resource allocation [40]. Our findings support the AGS position statement indicating age alone should never be used to make decisions regarding resource allocation under conditions of resource scarcity [26]. Although age is still an important factor in the overall risk of COVID-19 mortality through the risk of having increased baseline vulnerability and severe symptoms that require one to be hospitalized in the first place, a comprehensive approach that accounts for the above factors is essential in preventing ageism. Consideration of geriatrics-focused indicators will also be essential for future studies evaluating post-COVID sequelae, including post-discharge outcomes within and beyond 30 days [41].
This study has several limitations. First, due to the overwhelming number of cases admitted at the peak of COVID-19, our large integrated health system functioned as "one hospital. " Patients were transferred within the health system as well as to nationally-operated army facilities. While transferred patients were included in the descriptive data, they were excluded from the univariate and multivariate analyses if their discharge disposition was unknown. In addition, patients who were still admitted at the time of the last data update were also excluded. While the number of transferred and admitted patients (N = 161 and N = 29, respectively) was small and unlikely to make a difference statistically, there is still a small risk of bias. A second limitation is that given the retrospective nature of the study and large numbers of patients, we were limited to elements in the EHR and thus used surrogate markers for functional status and life-sustaining treatment preferences as well as the MEWS, which has not been evaluated in COVID-19. The advantage of utilizing administrative hospital data is that it does not rely on manual scores that must be entered by clinicians into the medical record or manual data extraction. We have previously published the validity of these data elements as surrogate markers [15].

Conclusion
COVID-19 has disproportionally affected older adults. This study confirms findings reported by other investigators that male sex, higher comorbidity index and higher illness severity are associated with hospital mortality. The confirmation of early DNR and 'residence prior to admission' are important additions to the literature and should be considered in future studies when considering patient-centered outcomes research. Lastly, while age remains an important factor in the overall risk of COVID-19 mortality, through the greater likelihood of physiological vulnerability and severe symptoms leading to hospitalization, a comprehensive approach that accounts for geriatrics-focused indicators is essential when evaluating prognosis and treatment decisions for hospitalized older patients with COVID-19.