Frailty Assessed by Administrative Tools and Mortality in Patients With Pneumonia Admitted to the Hospital and ICU: a Population-based Data-linkage Study

Background: The ideal method of identifying frailty is uncertain, and data on long-term outcomes is relatively limited. We examined frailty indices derived from population-scale linked data on Intensive Care Unit (ICU) and hospitalised non-ICU patients with pneumonia to elucidate the inuence of frailty on mortality. Methods: Longitudinal cohort study between 2010-2018 using population-scale anonymised data linkage of healthcare records for adults admitted to hospital with pneumonia in Wales. Primary outcome was inpatient mortality. Age, hospital frailty risk score (HFRS), electronic frailty index (eFI), Charlson comorbidity index (CCI), and social deprivation index were entered in the multivariate regression models. Results: Of the 107,188 patients, mean (SD) age was 72.6 (16.6) years, 50% were men. The two frailty indices and the comorbidity index had an increased risk of mortality for individuals with an ICU admission. Advancing age, increased frailty and comorbidity affected short- and long-term mortality. For predicting inpatient deaths, the CCI and HFRS based models were similar, however for longer term outcomes the CCI based model was superior. Discussion: Frailty and comorbidity are signicant risk factors for patients admitted to hospital with pneumonia. Frailty and comorbidity scores based on administrative data have only moderate ability to predict outcome. using with the binary dependent separate models for eFI, HFRS and CCI. In each of models we included following variables as xed effects: ICU admission, eFI Mild, Moderate, Severe), HFRS Intermediate, High), CCI >10), age (18-49,50-59,60-69,70-79,80+), and sex (Male/Female). We included changes Statistical analysis was performed using R version and R2MLwiN 22 . The best tting logistic regression models were determined using the Area Under the (AUC). Chi-squared tests were used for categorical variables and a t-test for the mean age to calculate p-values to determine statistically signicant differences between demographic variables in the ICU and non-ICU admissions. Data presented as absolute counts, mean (SD) or percentages; ICU: Intensive Care Unit; WIMD: Welsh Index of Multiple Deprivation 2019. 52-item ICU Recent study of 18 very severe frailty scores benet We found and similar in 70 who high frailty ICU. Our data their ageing


Introduction
Outcomes from critical illness among older patients and those with poorer health status are of increasing signi cance as global populations age, and have been the subject of intense interest during the SARS-CoV-2 pandemic 12 . Frailty is more common with increasing age and is often de ned as a syndrome of physiological decline, characterised by marked vulnerability to adverse health outcomes 3 4 . Frailty has been shown to modify the treatment effect of multiple high-risk interventions and independently predicts adverse outcomes beyond traditional comorbidities in several populations, including pneumonia 5 .
The Clinical Frailty Scale 6 has been widely adopted in the acute setting as an easily applied judgementbased tool, and is speci cally referenced within national COVID-19 decision-making guidance 7 . It is often considered to be a composite assessment, which explicitly incorporates disability and the consequences of comorbidity 8 . However, there is inherent subjectivity in such assessment, and we have previously described the limits to inter-rater reliability, and the need for training to enable consistent application 9 .
Overall, there remains some uncertainty about how best to identify frailty in relation to acute and critical illness 1011 .
Alternative approaches to the identi cation of frailty which adopt a cumulative de cit model have recently been applied to UK National Health Service (NHS) electronic primary care data, the electronic frailty index (eFI) 12 , and to hospital records, the Hospital Frailty Risk Score (HFRS) 13 , but these measures have not been fully evaluated in a critical care setting. Pneumonia has historically been a major cause of mortality 14 15 , a leading cause of sepsis, and one of the most common indications for critical care admission 16 . A very recent US study indicated signi cant association of HFRS with 30-day mortality in older patients hospitalised with pneumonia 5 , however data is currently limited regarding the predictive validity of frailty indices in relation to longer-term outcomes from critical illness.
In this study we aim to answer the following question: how is patient frailty and comorbidity, as identi ed by administrative tools, associated with inpatient, 6-month and 1-year mortality following hospitalisation with pneumonia?

Data sources
We used the Secure Anonymised Information Linkage (SAIL) Databank (www.saildatabank.com) to investigate short and long-term mortality in patients admitted with pneumonia to Welsh National Health Service (NHS) hospitals. The development of the SAIL Databank as a secure privacy protecting Trusted Research Environment (TRE) of anonymised person-based records has been described previously 17

Study cohort
We used the PEDW inpatient data to identify patients admitted to hospital for the rst time with a diagnosis of pneumonia or u between 1 st January 2010 and 31 st December 2018. A pneumonia or u diagnosis was de ned using the International Coding of Disease version 10 (ICD-10) codes J09-J18. This data was linked with the WDSD to obtain the week of birth, sex and Lower Super Output Area (LSOA) of patient residence version 2011 (each LSOA contains approximately 1,500 people) and onward linked to Welsh Index of Multiple Deprivation (WIMD) 2019 of each patient. We used week of birth to generate age on the admission date and included all patients aged 18-years or more. We identi ed patients with an ICU admission as those with an in-patient spell from PEDW that also contained a spell recorded in CCDS. Furthermore, we limited the cohort to those with high quality matching from the linkage and anonymisation process as previously described 18 .

Assessment of comorbidity and frailty
Hospital Frailty Risk Score The Hospital Frailty Risk Score (HFRS) was developed using Hospital Episode Statistics (a database containing details of all admissions, Accident and Emergency (A&E) attendances and outpatient appointments at NHS hospitals in England), and validated on over 1-million older people admitted to hospitals in 2014/15 13 . The HFRS uses ICD-10 codes to search for speci c conditions from secondary care. A weight is then applied to the conditions and a cumulative sum is used to determine a frailty status of: Low, Intermediate or High 13 . We calculated the HFRS using PEDW on the date of admission, with a 2- year look back of all hospital admissions occurring in Wales.

electronic Frailty Index (eFI)
The eFI is based on an internationally established cumulative de cit model 19 , and assigns a frailty score to an individual calculated using 36 variables from primary care data including symptoms, signs, diseases, disabilities and abnormal laboratory values, referred to as de cits 12 . The eFI score is the number of de cits present, expressed as an equally weighted proportion of the total. An individual with a single de cit would be assigned an eFI of 1/36 (0.03); another with nine de cits would be assigned an eFI of 9/36 (0.25). The eFI score is used to categorise individuals as: t (eFI value of 0-0.12), mild (>0.12 -0.24), moderate (>0.24 -0.36), or severely frail (>0.36) 12 . We calculated the eFI retrospectively on the date of admission to hospital using 10-years of previous WLGP data for each individual. As the eFI requires primary care data to be calculated, we restricted the analyses to only patients with a recorded WLGP registration during the hospital spell.

Charlson Comorbidity Index (CCI)
The modi ed CCI was generated using the ICD-10 codes and weights detailed in the NHS Summary Hospital-level Mortality Indicator speci cation 20 . ICD-10 codes relating to human immunode ciency virus (HIV) are redacted from the data available within SAIL in accordance with SAIL policies, all other hospital admissions recorded in PEDW in the year preceding the discharge date were included in the CCI calculation 21 . Scores were categorised into three groups: low (-1-0), medium (1-10) and high (>10).

Outcomes of interest
The primary outcome was in-patient mortality. The secondary outcomes were 6-month and 1-year mortality. We compared outcomes between those who had been admitted to ICU during the index hospitalisation and those who had not.

Statistical analysis
Basic demographic data are presented as counts, percentages and gures. To determine if the eFI, HFRS and CCI identify the same groups of individuals, we calculated the categorical scores of both tools for each patient on the day before their index admission date. We then used Chi-squared tests to determine if the tools were associated, and Cramer's V to determine the strength of association. To be consistent with our methodology at speci ed time-points and determine associations we calculated multilevel logistic regression models for inpatient, 6-month, and 1-year mortality using a logit link function with mortality as the binary dependent variable. We created separate models for the eFI, HFRS and CCI. In each of the models we included the following variables as xed effects: ICU admission, eFI (Fit, Mild, Moderate, Severe), HFRS (Low, Intermediate, High), CCI (<1, 1-10, >10), age (18-49,50-59,60-69,70-79,80+), and sex (Male/Female). We included a random intercept for each admission year to account for potential changes in recording. Statistical analysis was performed using R version 4.0.0 and R2MLwiN 22 . The best tting logistic regression models were determined using the Area Under the Curve (AUC).

Participants
There were 133,604 index admissions for patients with pneumonia or u to Welsh NHS hospitals over the study period. As the eFI is calculated using primary care records the cohort was restricted to only include those registered with a general practice contributing data to SAIL in the study period (80%), this limited the cohort to 107,188 patients. The following owchart demonstrates how the data for the ICU and hospital cohort was obtained ( Fig. 1).

Demographics and distribution of frailty and comorbidity according to eFI, HFRS and CCI
The demographic details of the patients are presented in Table 1. The non-ICU hospital admissions were dominated by those aged 80 years and over, whereas 50% of critical care patients were in the 60-79 year group. Higher levels of frailty de ned by the eFI were seen in the non-ICU group, and only 23% of the ICU group were considered moderately or severely frail. Whereas the HFRS categorised a larger proportion of individuals as low risk in the non-ICU cohort, most ICU patients were considered "intermediate" risk. There were statistically signi cant, but clinically irrelevant differences in chronic illness prevalence, as quanti ed by the CCI, between ICU and non-ICU patients. Interestingly, patients with high comorbidity burden were represented identically between the ICU and non-ICU groups. ICU admissions were more likely to be from deprived areas compared with non-ICU admissions. We also found an increase in frailty (eFI and HFRS) and comorbidities (CCI) with age (Fig. 2).  Figure S1 and S2). In patients between 50 and 70 years of age the gap only closed in the high HFRS and CCI groups (Supplementary Figure S1 and S2). We did not observe similar patterns using the eFI score (Supplementary Figure S3).

Association between eFI, HFRS and CCI
The cross tabulations for eFI, HFRS and CCI are included in the supplementary material (Supplementary  Table S1-S3). Chi-squared tests for independence indicated that the eFI, HFRS and CCI were all associated (p < 0.001). Cramer's V showed the strength of association was similar amongst all the tools, but strongest between the eFI and HFRS (0.255) and weaker between the CCI and frailty tools (eFI: 0.206, HFRS: 0.217) (Supplementary Table S4).

Risk factors of mortality at different time-points
Multivariate logistic regression analysis based on the two frailty indices and the comorbidity index consistently had an increased risk of inpatient, 6-month and 1-year mortality for individuals with an ICU admission. Similarly, in all models, advancing age, increased frailty and comorbidity burden together with residence in the most a uent areas affected short-and long-term mortality. The estimated AUC revealed that each model had only moderate discriminatory power for predicting mortality. For predicting inpatient deaths, the CCI and HFRS based models were similar, however for longer term outcomes the CCI based model was superior to the frailty scores (Tables 2-4).

Discussion
In this population-scale, data linkage study, we examined in-patient and long-term mortality among patients hospitalised with pneumonia (including those admitted to ICU), to investigate the predictive validity of frailty assessment tools derived from administrative data. Our data indicates: that an in-patient hospital admission with pneumonia carries a high risk of death, with two out of ve patients dying at 1- year; that eFI, HFRS and CCI identify different cohorts of patients within the ICU and non-ICU groups and has only moderate agreement between each other and that each investigated model has moderate predictive power for short-and long-term mortality, with the CCI based model showing best predictive ability at 1-year.
To the best of our knowledge, this is the largest dataset evaluating long-term outcomes for in-patient hospitalisations with pneumonia in relation to frailty. We examined mortality up until 1-year following hospitalisation. Unsurprisingly, patients admitted to ICU had signi cantly worse outcomes in the shortterm. The biggest difference was observed for in-patient mortality, likely re ecting the impact of acute illness severity and need for invasive organ support. This is consistent with previous reports for pneumonia and sepsis 2324 . The overall in-patient mortality rates observed in our cohort (23.9%) are higher than a recent report from the US (12%), but similar to mortality rates in a German study from 2010-11 (22%) 2526 . Direct comparisons are made challenging by differences in study design, discharge destination and care administrative processes.
Given such an apparent susceptibility to poorer long-term outcomes, patients hospitalised with pneumonia are a relevant cohort in which to investigate characterisation and predictive validity of frailty scores. Since frailty measures which derive from administrative data may avoid the implementation burden of bedside assessment, reliance upon proxy information, and potential for recall bias, we utilised two measures we have previously validated in other NHS cohorts, the electronic frailty index (eFI) and Hospital Frailty Risk Score (HFRS) 1012132728 .
Interestingly, we found that the eFI, which is probably closer to the theoretical description of an accumulated de cit frailty model, had only moderate agreement with HFRS categorisation in ICU and non-ICU groups. Most ICU patients were considered at intermediate risk according to HFRS but t or only mildly frail according to eFI, whereas most non-ICU patients were low risk according to HFRS but had a higher proportion who would be considered moderate or severely frail according to eFI. Comorbidity was distributed evenly between ICU and non-ICU patients.
Differences in characterisation within ICU and non-ICU groups according to frailty and comorbidity are not unexpected. A fair to moderate level of agreement between frailty assessment methods is not uncommon and our very recent direct comparison of hospitalised patients 65 years and over, also identi ed low correlation between eFI and HFRS 1328 . Consistent with previous studies, we observed signi cantly poorer outcomes with higher categories of frailty for eFI and HFRS among hospitalised patients -including those admitted to ICU -with the HFRS discriminating to a slightly greater degree than eFI 529 .
Conceptualisation and measurement of frailty using retrospective healthcare records presents challenges for HFRS and eFI, each method is dependent upon the quality of coding; for eFI, the coding of de cits (from clinical features, speci c illnesses, disabilities, and laboratory values) in primary care, and for HFRS the identi cation of speci c conditions according to ICD-10 during hospitalisation. Maintaining consistency with a cumulative de cit model (with an absolute predicted risk), the eFI assigns equal weight to each variable, whereas HFRS produces a weighted score 12 . The HFRS may miss elements which are important to consideration of frailty, such as polypharmacy or care requirements 13 . The look back periods studied also differ by design: for eFI this was 10-years, and for HFRS 2-years 28 . This may have led to the exclusion of important conditions in HFRS pre-dating hospital coding; alternatively, HFRS, with a shorter look back period, may have served as a more dynamic measure of immediately relevant conditions. Furthermore, important to note that although both the CCI and HFRS use ICD-10 codes, they express different risks prevalent in the same population. Whilst CCI is derived using ICD-10 codes of 17 comorbidities, the HFRS was developed giving weighting to 109 frailty-related ICD-10 codes depending on their association with frailty 1321 . In a recent study in heart failure patients only moderate agreement was found between CCI and HFRS, although both scores were predictive of outcome 29 .
A novel aspect of our study is the application of frailty assessment tools derived from administrative data in relation to ICU and non-ICU outcomes in pneumonia. Most previous studies reported intensive care outcomes according to the CFS and only two utilised a frailty index as predictor of long-term outcome: a 43-item tool derived from comprehensive geriatric assessment, and a novel 52-item instrument based upon data from patients admitted to a specialist geriatric ICU 303132 . Recent study of adult patients aged 18 years and over with pneumonia admitted to Australia / New Zealand intensive care units found that only severe or very severe frailty scores (according to CFS) predicted in-hospital mortality; however, such patients will have been selected and ICU admission agreed on the basis of perceived likelihood of bene t 33 . We found that short and long-term mortality was similar in the patients aged 70 years and over who had high frailty scores, whether they were admitted or not to the ICU. Our data further highlights the challenges faced by critical care teams in their management of an ageing and frail population.
An important further consideration is that eFI and HFRS were developed and validated in patient populations aged 65 years and over. However, the potential to assess physiological reserve and vulnerability to poorer outcomes for all hospitalised adults (and all age cohorts) is hugely appealing, and our study presents novel data with regards to the application of these frailty scores in younger patients. In our cohort, approximately one third of the patients were aged less than 65 and it is in this younger group that the frailty scores and the comorbidity evaluation showed the greatest divergence. Clearly, further work is required to establish how it may be possible to incorporate assessment of comorbidity and "frailty" determined on the basis of primary and/ or secondary care administrative data, among all age cohorts who may be admitted to critical care -and this seems likely to require the development of a new predictive model.

Limitations
Our study has limitations. The administrative in-patient data lacks detailed information about disease severity and medical treatments; comorbidities are dependent upon reliable ICD-10 coding, and variation in documentation and coding of diagnoses could contribute to measurement error 34 . However, in Wales, routinely collected population-scale EHR data are su ciently robust and accurate to be used in research or for health services planning 35 . We could not analyse the direct impact of possible differences in treatment intensity during ICU admission, as comparative data are not available currently in the SAIL Databank in the hospital cohort. We were unable to examine long-term outcomes other than mortality, and it would be useful to explore how these frailty indices predict subsequent healthcare utilisation, physical function, disability and quality of life 31363738 . The observed higher mortality in the low frailty and comorbidity cohorts indicates that acute physiological disturbance leading to ICU admission will have signi cant effect on short-term outcomes. On the other hand, recent data, including ours, indicates that following ICU and hospital discharge, acute illness characteristics have diminishing impact on longterm outcomes 36394041 . Future research should focus on exploring the utility of combining frailty instruments with age, comorbidity and illness severity data, and explore patient-centred outcomes other than mortality.

Conclusion
Patients admitted to hospital with pneumonia face a signi cant mortality risk in all age groups with less than three out of ve alive at 1-year. When evaluating frailty assessment tools, HFRS and eFI performed differently in the ICU and non-ICU population, HFRS indicating higher risk amongst the ICU patients. Older patients with high frailty scores appear to have similarly poor outcomes with or without ICU admissions.
Our results illustrate that there is value in the identi cation of frailty using indices derived from administrative data, but further work is needed to investigate how best to integrate such data, and to incorporate further with assessment of severity of acute illness and chronic comorbidity. Declarations