Vital signs and common blood tests improve the predictive power of the Hospital Frailty Risk Score to predict poor outcomes across all adult ages Authors: Huda Kutrani, Jim Briggs, David Prytherch, Claire Spice Published: 2026 ABSTRACT BackgroundFrailty is associated with poor health outcomes and is a public health challenge worldwide. The Hospital Frailty Risk Score (HFRS) has been widely used to identify patients at risk of frailty and predict poor outcomes including long length of stay (LOS) and in-hospital mortality for older patients. This study aimed to explore and determine variables that might influence the ability of the Hospital Frailty Risk Score to predict LOS and in-hospital mortality across all adult ages.MethodsThis is a retrospective cohort study using data from Queen Alexandra Hospital in Portsmouth, UK of consecutive patient admissions over 10 years between 01/01/2010 to 31/12/2019. The study included patients aged 16 years and older. The HFRS was calculated for each patient based on ICD-10 diagnostic codes with a 2-year look-back. The National Early Warning Score (NEWS) and the Laboratory Decision Tree Early Warning Score (LDT-EWS) were calculated for each patient. Vital signs and blood tests were the first available routine data from patients after admission. We developed logistic regression models (alone and adjusted) for 9 prediction periods of length of stay and 8 prediction periods of in-hospital mortality and assessed the model performance using AUROC.ResultsCombining HFRS with the LDT-EWS had the highest discrimination (AUROC ranging from 0.764 to 0.810) compared to adjusted models (AUROC ranging from 0.716 to 0.796) or HFRS alone (AUROC ranging from 0.723 to 0.798) for 9 periods of length of stay. For in-hospital mortality, combining HFRS with NEWS had the highest discrimination (AUROC ranging from 0.786 to 0.829) compared to HFRS alone or HFRS combined with other variables for 3, 7, 10 and 14-day mortality across all adult ages. And combining HFRS with LDT-EWS had the highest discrimination (AUROC ranging from 0.789 to 0.794) for mortality after more than 14 days across all adult ages.ConclusionsCombining HFRS with additional routinely available variables significantly improves the predictive power for length of stay and mortality. This is the first paper to show that LDT-EWS significantly improves the predictive power of Hospital Frailty Risk Scores to predict longer length of stay in hospital and later in-patient mortality across all adult ages. The predictive power of the HFRS was improved by NEWS for early in-patient mortality. Introduction Frailty is a medical syndrome characterized by reduced physiological function and increased vulnerability to stressors such as hospital admission. It is associated with morbidity and increased risk of poor health outcomes and is a reliable measure to predict health outcomes such as length of stay and mortality [1–3]. Identifying patients at high risk of frailty assists in enabling timely interventions that reduce the risk of poor health outcomes, reduce healthcare resource use, and improve patient care [1–4]. However, identification of those with frailty is not routine in many systems and where it is mandated, such as in the English National Health System, take up is not universal [5]. Although frailty increases with age, it is also prevalent in younger adults [6–8] and has been identified as a strong predictor of length of stay (LOS) and mortality in adults of all ages [8–10]. Tools to measure frailty are becoming increasingly important in identifying people at a high risk of poor outcomes but many require face-to-face assessment, which can be difficult in an acute care setting [11–13]. The Hospital Frailty Risk Score (HFRS) was developed initially in older people in hospital and has the advantage of being based on ICD 10 diagnosis codes [11,14]. Subsequently it has since been validated in various countries and in specific populations to identify older people at risk of frailty and predict longer LOS and in-hospital mortality [12–16]. The recent studies found that HFRS can predict hospital LOS [17] and in-hospital mortality [18] across all adult ages, not just for older people. It can be implemented in most hospital information systems at low cost and without additional burden on clinical staff [11,12,14,19]. Whilst frailty scoring and scales can identify those at risk of longer hospital stays and mortality there is potential to improve prediction through combination with other relevant variables. Further evaluation of the HFRS in combination with variables such as physiological scores and in more differentiated cohorts has been suggested as a focus for research [20].The majority of studies have included patients’ characteristics such as age, sex, and Charlson Comorbidity Index (CCI) with HFRS to predict long hospital stay and in-hospital mortality and the predictive power of the HFRS models improved [11,16,21] or remained significant [12–14]. However, common blood tests and vital signs, along with age and gender are also predictors of poor health outcomes [22–26]. Vital signs represented by NEWS [23] and 7 common blood tests (Haemoglobin, Urea, Potassium, Sodium, White Cell Count, Creatinine, and Albumin) represented by LDT EWS (Laboratory Decision Tree Early Warning Score) [22] can assist in the detection of clinical deterioration in patients. Previous studies showed that NEWS [23,24,27,28], LDT-EWS [22] and separate blood tests (a part of LDT-EWS) [22,26,29,30] are predictive of poor outcomes. Some studies reported that adding blood tests results [31,32] and vital signs [27,28] to prediction models improved the predictive power of the models. Adding a clinical frailty scale to a variety of physiological scores has been shown to improve the prediction of admission to hospital from the Emergency Department and in-hospital mortality in older people [33]. The analysis of factors that may influence the prediction of HFRS for poor outcomes, including readily available clinical data, has the potential to improve the ability to identify those who are admitted to the hospital who are at risk of a prolonged length of stay or in-hospital mortality. Therefore, this study aimed to explore whether other variables such as laboratory results, physiological scoring and comorbidity influence the predictive power of HFRS to predict length of stay and in-hospital mortality across all adult ages. Methods Study design and participants This study follows similar methods to those of Kutrani et al. (2025) and summarizes them here for any convenience [17,18,34]. This is a retrospective cohort study of non-elective patients who were admitted to a large acute hospital (Queen Alexandra Hospital in Portsmouth, UK) from 1st January 2010–31st December 2019. The study included patients aged 16 years and older who had not registered for the national data opt-out to stop their data being used for research. The full dataset was 575,045 non-elective admissions. Since the calculation of HFRS relies on having 2 previous years’ data for optimal construction of HFRS [14], the analysis was based on patients admitted from 1st January 2012–31st December 2019. We excluded maternity cases and certain types of admission from the study (eligible admissions were 378,916 admissions) as described in Fig 1. Flowchart showing the inclusion and exclusion criteria for the study. Hospital frailty risk score calculation The HFRS was calculated for each patient, based on 109 International Classification of Diseases 10th revision (ICD-10) diagnoses documented in their hospital admission records. The final HFRS for each patient is calculated by adding the weighted points together for each code present from the index admission with any diagnosis recorded from previous admissions during the previous two years [11,14]. In addition, HFRS were categorised into low frailty risk (score less than 5), intermediate frailty risk (5–15), and high frailty risk (>15) as per the original study [11]. Outcomes and variables We determined the predictive ability of the HFRS on nine periods of length of stay and eight periods of in-hospital mortality (death discharge) as outcomes of the study. These prediction periods were shown in Table 1. 10.1371/journal.pone.0348669.t001Table 1Prediction periods of length of stay and in-hospital mortality.9 periods of LOS8 periods of in-hospital mortalityLOS > 3 days3-day mortalityLOS > 7 days7-day mortalityLOS > 10 days10-day mortalityLOS > 14 days14-day mortalityLOS > 21 days30-day mortalityLOS > 30 days60-day mortalityLOS > 45 days90-day mortalityLOS > 60 days6-month mortalityLOS > 90 days Prediction periods of length of stay and in-hospital mortality. Variables were HFRS, age, gender, Charlson Comorbidity Index (CCI), aggregate NEWS, c-reactive protein test (CRP), and aggregate LDT-EWS. Laboratory tests and vital signs studied were the first available results gathered from patients at the point of the hospital stay. Most of the blood tests and vital signs were collected on the first day of admission. The seven vital signs were Respiration Rate (RR), body temperature (Temp), Systolic Blood Pressure (SBP), Saturation of Peripheral Oxygen (SPO2), Heart Rate (HR), ACVPU, and Oxygen state. The NEWS is an aggregate score that represents these vital signs, with a higher score indicating increased risk of poor outcomes [23]. The LDT-EWS is based on seven blood tests: Haemoglobin (HB), Urine (U), Potassium (K), Sodium (Na), White Cell Count (WCC), Creatinine (CR), and Albumin (Alb). For each test, LDT-EWS assigns a score from 0 to 3, which are then summed to give an aggregate score under one variable named LDT_EWS. A higher value of LDT-EWS represents a higher relative risk of in-hospital death [22]. We included C-reactive protein (CRP) separately because it is not included in LDT-EWS. An abnormal level of CRP is typically associated with increased risk of longer LOS and in-hospital mortality [35]. An elevated CRP signifies infection or inflammation. The CRP ranges from <1 mg/L to over 500 mg/L. A value below 3.0 mg/L is normal; a value from 3.0–10.0 mg/L is slightly elevated. A value from 10.0 to 100.0 mg/L is moderately elevated, 100.0 to 500.0 mg/L is elevated, and above 500 mg/L is severely elevated [36]. Statistical analysis Patient characteristics were described using mean, standard deviation, median, and interquartile range (IQR) for continuous variables, and number and percentage for categorical variables. Logistic regression models were developed with nine LOS and eight in-hospital mortality periods as outcomes. In the first group, models used each variable alone, including the HFRS score to predict each period of LOS and in-hospital mortality. In the second group, models used the HFRS score (a continuous variable) in combination with one other variable to predict each outcome of LOS and in-hospital mortality. In the third group a multivariate model (including all variables in our study) was used. For LOS analysis, we excluded patients who died in the hospital, because patients who died in the hospital had a shorter length of stay than if they had survived, which may have a negative effect on predicting longer LOS [37]. Further LOS analysis was performed for all patients, including those who died in the hospital. The study cohort was randomly split into 70% training dataset and 30% validation dataset. The predictive power of models was assessed by discrimination using the Area Under Receiver Operating Characteristic (AUROC or c-statistic) with 95% confidence intervals (95% CI). A value of AUROC below 0.6 indicates no discrimination; values ranging from 0.60–0.7 indicate poor discrimination, values of 0.7–0.8 indicate fair discrimination, values from 0.8–0.9 indicate good discrimination, and values above 0.9 indicate excellent discrimination [38]. We compared the AUROC value of the HFRS alone model with other models for each period of LOS and in-hospital mortality. A higher AUROC indicated improved predictive power of the model. P values <0.05 were considered statistically significant. Data manipulation and logistic regression modelling were performed using R-Studio version 4.2.1. Cross validation We conducted validation experiments to evaluate the stability of the models to see whether the discrimination (AUROC) of the best model is the best predictor for the nine periods of LOS and eight periods of in-hospital mortality. We evaluated: What is the best model for each LOS and in-hospital mortality period? What is the best model for each LOS and in-hospital mortality period? Do different samples of training and testing datasets have different results? Do different samples of training and testing datasets have different results? Does the best model always consistently give the best results for any sample data? Does the best model always consistently give the best results for any sample data? In addition, we generated several samples of data for validation experiments which included the following: 8 samples of data according to admission year – each calendar year from 2012 to 2019. 8 samples of data according to admission year – each calendar year from 2012 to 2019. 4 samples of data according to age (<45 years, 45–64 years, 65–84 years, and ≥85 years). 4 samples of data according to age (<45 years, 45–64 years, 65–84 years, and ≥85 years). 2 samples of data according to gender. 2 samples of data according to gender. Full details of these samples are reported in S1 Table. Ethics The study included patients aged 16 years and older who had not registered for the national data opt-out for their data being used for research. We accessed data from the Portsmouth CORE-D routine care data repository on 1st November 2022. The research team did not have access to any information that could identify individual patients during or after data collection. The dataset used in this study was covered by existing ethical approval granted by an NHS Research Ethics Committee in April 2021. REC reference is 21/SC/0080. IRAS project ID is 281193. The data was made available by Portsmouth Hospital University NHS Trust under a data sharing agreement. Results A total of 378,916 non-elective admissions for patients aged 16 years and older were included. Patients had HFRS ranging from 0 to 77 with a mean of 7.1. Regarding the HFRS category, 223414 (59.0%) of 378916 admissions were categorised as low risk of frailty, 96623 (25.5%) of 378916 admissions were categorised as intermediate risk of frailty, and 58879 (15.5%) of 378916 admissions were categorised as high risk of frailty. The mean age was 61.8 years and 54.9% were females. The mean and median aggregate LDT-EWS was 2.0 and 1, NEWS was 1.9 and 1, CCI was 4.8 and 0, and CRP was 49.2 and 13.0. For about 40% of admissions, the speciality of discharge was General Medicine, followed by 12.2% for Accident and Emergency (S2 Table). HFRS and length of stay prediction Patients’ characteristics in the dataset and each of the length of stay period subsets are presented in Table 2. Mean HFRS increased with a longer length of stay. The mean and median aggregate LDT-EWS was around 3.2 in all periods of length of stay. The mean and median aggregate NEWS were low risk in all length of stay periods. In general, CCI increased with the length of stay but then decreased after (LOS > 21 days). The mean and median CRP level was “moderately elevated” in all length of stay periods. 10.1371/journal.pone.0348669.t002Table 2Characteristics of patients by predicted length of stay periods.prediction Length of Stay (LOS)LOS > 3daysN = 128,606LOS > 7daysN = 74,677LOS > 10daysN = 54,948LOS > 14daysN = 38,955LOS > 21daysN = 23,497LOS > 30daysN = 14,020LOS > 45daysN = 6,783LOS > 60daysN = 3,484LOS > 90daysN = 1,059 HFRS Mean (SD)10.8 (10.7)12.8 (11.2)13.7 (11.3)14.5 (11.4)15.4 (11.4)15.9 (11.4)16.2 (11.3)16.2 (11.2)16.2 (10.2) Median (IQR)7.4 (2.6–15.7)10.0 (4.2–18.6)11.0 (5.1–19.6)12.0 (5.8–20.4)13.0 (6.6–21.6)13.6 (7.1–22.2)14.0 (7.6–22.4)14.0 (7.9–22.1)14.1 (8.6–21.6) Age in years Mean (SD)71.6 (17.9)75.0 (16.0)76.3 (15.5)77.0 (14.9)77.5 (14.8)77.2 (14.9)76.6 (14.8)75.2 (15.3)72.4 (16.2) Median (IQR)76.0 (62.0–85.0)79.0 (67.0–87.0)80.0 (69.0–87.0)81.0 (70.0–88.0)81.0 (71.0–88.0)81.0 (70.0–88.0)81.0 (69.0–87.0)79.0 (67.0–86.0)76.0 (63.0–85.0) LDT-EWS Mean (SD)2.9 (2.4)3.2 (2.4)3.3 (2.4)3.3 (2.4)3.3 (2.4)3.2 (2.3)3.2 (2.3)3.2 (2.3)3.2 (2.3) Median (IQR)2.0 (1.0–4.0)3.0 (1.0–4.0)3.0 (2.0–4.0)3.0 (2.0–4.0)3.0 (2.0–4.0)3.0 (1.0–4.0)3.0 (1.0–4.0)3.0 (1.0–4.0)3.0 (1.0–4.0) NEWS Mean (SD)2.3 (2.4)2.4 (2.4)2.4 (2.4)2.5 (2.4)2.5 (2.4)2.5 (2.4)2.4 (2.3)2.4 (2.3)2.4 (2.3) Median (IQR)2.0 (0.0–3.0)2.0 (0.0–4.0)2.0 (1.0–4.0)2.0 (1.0–4.0)2.0 (1.0–4.0)2.0 (1.0–4.0)2.0 (0.0–4.0)2.0 (0.0–4.0)2.0 (0.0–4.0) CCI Mean (SD)7.1 (10.5)7.9 (11.1)8.2 (11.3)8.3 (11.4)8.3 (11.3)7.8 (11.1)7.2 (10.7)6.6 (10.2)5.7 (9.6) Median (IQR)0.0 (0.0–12.0)3.0 (0.0–14.0)3.0 (0.0–14.0)3.0 (0.0–14.0)3.0 (0.0–14.0)0.0 (0.0–14.0)0.0 (0.0–13.0)0.0 (0.0–11.0)0.0 (0.0–10.0) CRP Mean (SD)68.2 (90.0)71.4 (91.7)71.5 (92.3)71.7 (92.1)71.9 (92.9)72.6 (93.9)70.3 (92.4)69.6 (92.7)73.9 (98.5) Median (IQR)27.0 (7.0–98.0)30.0 (7.0–104.0)31.0 (8.0–104.0)30.0 (8.0–104.0)31.0 (8.0–104.0)31.0 (8.0–105.0)29.0 (7.0–102.0)27.0 (7.0–98.0)29.0 (7.0–108.0) Gender Female No. (%)38009 (52.9%)40350 (54.0%)29842 (54.3%)21115 (54.2%)12639 (53.8%)7405 (52.8%)3407 (50.2%)1720 (49.4%)494 (46.6%) Male No. (%)60597 (47.1%)34327 (46.0%)25106 (45.7%)17840 (45.8%)10858 (46.2%)6615 (47.2%)3376 (49.8%)1764 (50.6%)565 (53.4%)HFRS: Hospital frailty risk score; NEWS2: aggregate National Early Warning Score2; LDT-EWS: aggregate Laboratory Decision Tree Early Warning Score; CCI: Charlson Comorbidity Index; CRP: c-reactive protein test. Characteristics of patients by predicted length of stay periods. HFRS: Hospital frailty risk score; NEWS2: aggregate National Early Warning Score2; LDT-EWS: aggregate Laboratory Decision Tree Early Warning Score; CCI: Charlson Comorbidity Index; CRP: c-reactive protein test. Evaluation of the predictive power of each variable alone and HFRS combined with other variables is summarized in S3 Table and Fig 2. In general, the predictive power of crude HFRS (HFRS alone) increased with length of stay (AUROC = 0.723 for LOS > 3 days, rising to 0.798 for LOS > 90 days). Combining HFRS with age or CCI improved predictive ability of HFRS for LOS in excess of 3, 7, 10 and 14-days. However, in each LOS period, the AUROC value of the HFRS combined with LDT-EWS model was superior to each variable alone or models where HFRS was combined with any other variable. Combining HFRS with LDT-EWS gave AUROCs ranging from 0.764 to 0.810 (with all AUROCs being statistically significant, P < 0.001) showing that LDT-EWS significantly improves the predictive power of HFRS across all LOS periods. Area under ROC for 9 periods of prediction long length of stay. To confirm the effectiveness of LDT-EWS, we investigated whether there were specific individual laboratory tests that influenced the result. However, the results showed that no single laboratory tests influenced HFRS as effectively as the LDT-EWS aggregate score. The results are detailed in S4 Table. Validation results for LOS. Applying our models to the validation dataset, we confirmed that HFRS + LDT-EWS models delivered the best results in terms of AUROC for all nine prediction periods. Table 3 shows that for all eight validation samples according to admission year (2012–2019) there was fair or good discrimination. Also, for all four samples data according to age groups, discrimination was poor, fair, good or excellent. The two subsets according to gender had fair and good discrimination. Full details of the results are reported in S5-S7 Tables. 10.1371/journal.pone.0348669.t003Table 3Results of AUROC from Combining HFRS with LDT-EWS model for validation data.Samples dataRange of AUROC Samples of data according to admission year Sample data 2012 (n = 46642)0.756–0.782 Sample data 2013 (n = 45922)0.761–0.826 Sample data 2014 (n = 45668)0.758–0.803 Sample data 2015 (n = 45745)0.779–0.811 Sample data 2016 (n = 45973)0.765–0.810 Sample data 2017 (n = 47431)0.771–0.858 Sample data 2018 (n = 50375)0.762–0.801Sample data 2019 (n = 51160)0.710–0.817 Samples of data according to age Age groups <45 years (n = 96963)0.741–0.902 Age groups 45–64 years (n = 83952)0.731–0.843 Age groups 65–84 years (n = 131042)0.693–0.751 Age groups ≥85 years (n = 66959)0.636–0.653 Samples of data according to gender Male data (n = 171008)0.778–0.809 Female data (n = 207908)0.745–0.782 Results of AUROC from Combining HFRS with LDT-EWS model for validation data. Additional analyses were performed for multivariate models in S8 Table. The multivariate models offered lower results in terms of AUROC (ranging from 0.636 to 0.713) than the HFRS alone models. These results show that combining all variables in our study with the HFRS does not improve the predictive power of HFRS to predict a long length of stay. Further analyses were performed on the dataset, after including patients who died in the hospital (S9 Table). The models offered results that were very similar to those above. These results demonstrate that HFRS+LDT-EWS significantly improves models’ performance to predict length of stay regardless of whether they died in hospital or not. HFRS and in-hospital mortality prediction Patients’ characteristics according to in-hospital mortality periods are presented in Table 4. The mean HFRS was above 10, and most patients were categorised as intermediate or high risk of frailty, in all in-hospital mortality periods. The mean and median aggregate LDT-EWS was 4 and above. The mean and median aggregate NEWS were medium risk in all periods of in-hospital mortality. The mean CCI was about 11 and median was about 8. The mean CRP level was “elevated” and median was “moderately elevated” in all periods of in-hospital mortality. 10.1371/journal.pone.0348669.t004Table 4Characteristics of patients by predicted in-hospital mortality periods (n = 378,916).Prediction in-hospital mortality3 days- mortality(n = 5,315)7 days- mortality(n = 8,542)10 days-mortality(n = 10,175)14 days-mortality(n = 11,770)30 days-mortality(n = 14,847)60 days-mortality(n = 16,123)90 days-mortality(n = 16,376)6 month-mortality(n = 16,460) HFRS score Mean (SD)10.7 (9.9)11.3 (10.1)11.5 (10.2)11.8 (10.3)12.5 (10.5)12.9 (10.6)12.9 (10.7)13.0 (10.7) Median (IQR)7.9 (3.3–15.4)8.5 (3.8–16.1)8.8 (4.0–16.3)9.1 (4.2–16.7)9.8 (4.6–17.7)10.2 (4.8–18.2)10.2 ((4.9–18.3)10.3 (4.9–18.4) HFRS Category Low1834 (34.5)2680 (31.4)3063 (30.1)3419 (29.0)3949 (26.6)4085 (25.3)4119 (25.0)4119 (25.0) Intermediate2089 (39.3)3475 (40.7)4189 (41.2)4858 (41.3)6098 (41.1)6602 (41.0)6735 (40.9)6738 (40.9) High1392 (26.2)2387 (27.9)2923 (28.7)3493 (29.7)4800 (32.3)5553 (33.9)5606 (34.1)5607 (34.1) Age in years Mean (SD)79.0 (13.0)79.2 (12.7)79.3 (12.5)79.4 (12.5)79.7 (12.3)79.8 (12.2)79.8 (12.2)79.8 (12.2) Median (IQR)82.0 (72.0–89.0)82.0 (72.0–88.0)82.0 (72.0–88.0)82.0 (72.0–88.0)82.0 (73.0–89.0)82.0 (73.0–89.0)82.0 (73.0–89.0)82.0 (73.0–89.0) LDT-EWS Mean (SD)4.2 (2.1)4.2 (2.7)4.2 (2.7)4.2 (2.7)4.2 (2.7)4.2 (2.7)4.2 (2.6)4.2 (2.6) Median (IQR)4.0 (2.0–6.0)4.0 (2.0–6.0)4.0 (2.0–6.0)4.0 (2.0–6.0)4.0 (2.0–6.0)4.0 (2.0–6.0)4.0 (2.0–6.0)4.0 (2.0–6.0) NEWS Mean (SD)5.4 (3.4)4.9 (3.2)4.7 (3.2)4.5 (3.1)4.2 (3.1)4.1 (3.0)4.1 (3.0)4.1 (3.0) Median (IQR)5.0 (3.0–8.0)5.0 (2.0–7.0)4.0 (2.0–7.0)4.0 (2.0–7.0)4.0 (2.0–6.0)4.0 (2.0–6.0)4.0 (2.0–6.0)4.0 (2.0–6.0) CCI Mean (SD)11.0 (13.3)11.4 (13.6)11.5 (13.7)11.7 (13.9)11.8 (13.9)11.8 (13.9)11.7 (13.9)11.7 (13.9) Median (IQR)6.0 (0.0–18.0)7.0 (0.0–18.0)8.0 (0.0–19.0)8.0 (0.0–19.0)8.0 (0.0–20.0)8.0 (0.0–20.0)8.0 (0.0–20.0)8.0 (0.0–20.0) CRP Mean (SD)111.7 (117.0)110.1 (113.0)108.4 (111.2)107.1 (109.7)103.5 (106.6)101.4 (105.4)101.0 (105.1)100.9 (105.1) Median (IQR)72.0 (18–170.0)75.0 (19–166.0)75.0 (20–162.0)73.0 (20–159.0)71.0 (19–153.0)68.0 (19–150.0)68.0 (19–149.0)68.0 (18–149·0) Gender Female No· (%)2662 (50.1%)4236 (49.6%)5005 (49.2%)5758 (48.9%)7183 (48.4%)7810 (48.4%)7925 (48.4%)7963 (48.4%) Male No· (%)2653 (49.9%)4306 (50.4%)5170 (50.8%)6012 (51.1%)7664 (51.6%)8313 (51.6%)8451 (51.6%)8497 (51.6%)HFRS: Hospital frailty risk score; Low frailty risk is score <5; Intermediate frailty risk is score 5–15; High frailty risk is score >15; LDT-EWS: aggregate Laboratory Decision Tree Early Warning Score; NEWS2: aggregate National Early Warning Score2; CCI: Charlson Comorbidity Index; CRP: c-reactive protein test. Characteristics of patients by predicted in-hospital mortality periods (n = 378,916). HFRS: Hospital frailty risk score; Low frailty risk is score <5; Intermediate frailty risk is score 5–15; High frailty risk is score >15; LDT-EWS: aggregate Laboratory Decision Tree Early Warning Score; NEWS2: aggregate National Early Warning Score2; CCI: Charlson Comorbidity Index; CRP: c-reactive protein test. Evaluation of the predictive power of each variable alone and HFRS combined with one other variable is summarized in S10 Table and Fig 3. Although combining HFRS with age or LDT-EWS or CCI or CRP improves the predictive power of HFRS for predicting 3, 7, 10 and 14-day in-hospital mortality, the performance of HFRS combined with NEWS is superior to models where HFRS was combined with any other variable or any variable alone. For all periods longer than 14 days in-hospital mortality, combining HFRS with age or NEWS or CCI or CRP improves the ability of HFRS, but the performance of HFRS combined with LDT-EWS is superior to models where HFRS was combined with any other variable. All AUROCs for all periods of in-hospital mortality were statistically significant (P < 0.001). These results show that NEWS and LDT-EWS improve the predictive power of HFRS in predicting in-hospital mortality across all adult ages. AUROC for 8 periods of prediction in-hospital mortality. Validation results for in-hospital mortality. Applying our models to the validation dataset, we confirmed that HFRS + NEWS models (for 3, 7, 10 and 14-day mortality) and HFRS + LDT-EWS models (for more than 14-day mortality) delivered the best results in terms of AUROC for samples according to admission year, age groups and gender. Table 5 shows that for all eight validation samples according to admission year (2012–2019) there was fair and good discrimination. The four subsets according to age and the two subsets according to gender also had fair and good discrimination. Full details of the results are reported in S11-S13 Tables. 10.1371/journal.pone.0348669.t005Table 5Results of AUROC from Combining HFRS with NEWS or HFRS with LDT-EWS models for validation data.Samples dataRange of AUROC for HFRS+NEWS for 3, 7, 10 and 14-day mortalityRange of AUROC for HFRS+LDT-EWS for more than 14 days mortality Samples of data according to admission year Sample data 2012 (n = 46642)0.802–0.8370.803–0.805 Sample data 2013 (n = 45922)0.798–0.8480.800–0.806 Sample data 2014 (n = 45668)0.793–0.8470.793–0.800 Sample data 2015 (n = 45745)0.789–0.8060.797–0.799 Sample data 2016 (n = 45973)0.781–0.8450.790–0.796 Sample data 2017 (n = 47431)0.779–0.8320.790–0.801 Sample data 2018 (n = 50375)0.790–0.8230.784–0.786 Sample data 2019 (n = 51160)0.786–0.8000.798–0.801 Samples of data according to age Age groups <45 years (n = 96963)0.888–0.8990.877–0.895 Age groups 45–64 years (n = 83952)0.821–0.8540.824–0.828 Age groups 65–84 years (n = 131042)0.749–0.7970.735–0.738 Age groups ≥85 years (n = 66959)0.724–0.7860.700–0.710 Samples of data according to gender Male data (n = 171008)0.791–0.8330.782–0.784 Female data (n = 207908)0.786–0.8250.795–0.797 Results of AUROC from Combining HFRS with NEWS or HFRS with LDT-EWS models for validation data. Additional analyses were performed for multivariate models (S14 Table). The multivariate models offered results that were close to those of simpler (i.e., HFRS+LDT-EWS) models. These results show that HFRS+NEWS significantly improves models’ performance to predict short term in-hospital mortality (up to 14 days) and HFRS+LDT-EWS significantly improves models’ performance to predict longer term in-hospital mortality. Discussion Our study, involving hundreds of thousands of patients aged 16 years and above admitted to urgent care at a large secondary care hospital in the UK, found a significant association between HFRS and important clinical outcomes of LOS and in-hospital mortality. We also found that a laboratory test score (LDT-EWS) established shortly after admission combined with frailty risk score (HFRS) improved the prediction of LOS and in-patient mortality after 14 days more than other additional variables. Earlier mortality (less than 14 days) in hospital predictive power was most improved by the first physiological score (NEWS) at admission. The HFRS has previously been associated with adverse outcomes for older people [13,14,16,19,21,39] including longer length of hospital stay. Our findings are similar for LOS in a broader group of non-electively admitted adults of all ages with the crude HFRS (HFRS alone) model having fair discrimination to predict LOS (between 0.723 to 0.798); as shown in Fig 2 and S3 Table. The predictive power of the HFRS for length of stay might partially be attributed to it including some significant healthcare events during hospitalisation (for example nosocomial infection). Although identifying those at high risk of long LOS may be helpful to enable earlier interventions to reduce the risks associated with longer hospital stays such as deconditioning, hospital acquired infections and increased healthcare resource utilisation [3,40,41], the utility of the HFRS as an identification tool in clinical practice is limited by the inclusion of coding for the current admission, which is typically not available until after discharge. Previous studies have demonstrated limitations to utility in individuals [10]. The finding that AUROCs were improved for LOS of >3 days by age and CCI is in keeping with results in the original HFRS study that showed that discrimination of the model improved slightly by including patients’ characteristics [11]. Others have found HFRS to predict a prolonged LOS and that predictive value remains after adjustment for common variables such as age, sex, and the CCI or other comorbidity indices [11–15]. We found that age and CCI increased the predictive power of the HFRS, although still only with fair discrimination. Most importantly combining the aggregate LDT-EWS with the HFRS was consistently a more powerful predictor of length of stay than any other variables explored for all periods of length of stay, increasing the discrimination compared to HFRS alone, although this remained fair until LOS > 45 days when the performance increased to good. Common laboratory test results can predict a higher risk of complication in hospital and community settings [26,31,32]. We found that separate specific laboratory tests did not influence the predictive power of HFRS, in contrast to laboratory tests together which were represented by LDT-EWS as one variable (S4 Table). There have been few studies of the HFRS, or other frailty risk indices, combined with laboratory tests for hospital inpatients. One study, Redfern et al, used the combination of LDT-EWS, NEWS and HFRS, along with the patient’s age, gender, and CCI in a multivariate model for the study outcomes of unplanned admission to the ICU and in-hospital mortality but LOS was not an outcome measured [42]. We looked at the potential for multivariate models to be more effective in predicting LOS but found these delivered lower results in terms of AUROC when compared to HFRS alone models or those combing HFRS with LDT-EWS (S8 Table). In another study, Sharma et al found that the model to predict hospital LOS remained significant after adding covariates (age, gender, CCI, Haemoglobin, and Creatinine) to the HFRS model but did not improve it [12]. HFRS alone has been found to be better than a warning score which incorporates the urea blood test result, at predicting LOS although it was not superior for mortality prediction [41]. So, whilst it is unlikely that individual blood test results improve the identification of those with frailty at risk of poor outcomes it may be that the larger number of blood tests that make up the LDT-EWS better reflect the spectrum and severity of illness that is represented in an undifferentiated cohort of hospital inpatients, including infection, that are not captured by one or two blood tests. The cumulative effect of abnormal laboratory tests with values that individually may not be predictive of outcomes, such as LOS or mortality, but when combined do so is recognised in a variety of settings, particularly in older people [43]. In this study we found that the HFRS combined with LDT-EWS remained a more effective predictor of length of stay than any other models whether those who died were included or not (S9 Table). The majority of patients who died in-hospital, across all eight time periods, were categorized as having a high or intermediate risk of frailty. This is in keeping with other studies that found HFRS was an independent and good predictor of in-hospital mortality [39,44,45]. Our study extends this finding to a much broader group of patients – non-electively admitted adult patients of all ages although we found that HFRS had poor discrimination for predicting early mortality within 3 days when used alone. Although HFRS has been widely used and validated in many countries for older patients [12,14,19,21,39,44,45], few studies have combined HFRS with other variables – common variables such as age, sex, and the CCI or other comorbidity indices – to assess improvement in its abilities to predict poor outcomes. A study found that HFRS with CCI offered only slightly better predictive power for 30-day in-hospital mortality for older people [46]. We have confirmed this finding in adults of all ages where combining HFRS with CCI did not improve (or only marginally improved) inpatient mortality prediction. Early mortality prediction using HFRS alone was poor possibly because it does not capture illness severity or physiological parameters which might reflect the intensity of the acute stressor. Vital signs (NEWS) [23], LDT-EWS [22] and CRP [47] along with frailty measured using HFRS [11] are all associated with in-hospital mortality. We found that HFRS combined with NEWS was superior to other models, with good predictive ability for 3- and 7-day mortality and better, but still fair, for 10- and 14-day mortality (S10 Table and Fig 3). Küçükceran et al. and Alshibani et al. found that the NEWS2 was a significant predictor of in-hospital mortality [13,48] in older people who were admitted from Emergency Departments. A systematic review reported that NEWS2 has good predictive power in predicting early mortality in patients, but it has poor performance in predicting long-term mortality [49]. The additive information from combining a frailty index or Clinical Frailty Scale with an illness acuity score has previously been described in a study of older (average of about 80 years) general internal medicine referred patients where 30-day mortality likelihood was much higher in those with severe frailty who had higher acuity scores [50]. Others have found that a single admission physiological scoring system (NEWS2) was not helpful in predicting mortality in a group of older people with frailty and Covid 19, perhaps because it does not capture some of the risks associated with mortality such as delirium [51]. The HFRS does include delirium and common presenting complaints such as falls which are more common in those with frailty. LOS is complex, and whilst the HFRS is broad in component codes, it does not reflect all the factors that may affect LOS, including patient characteristics, in-hospital processes and the availability of non-hospital services such as social care. For longer-term mortality, we found that models that combined HFRS with LDT-EWS were superior to other models, with fair discrimination for more than 14-day mortality across all adult ages (S10 Table and Fig 3). Others have reported that LDT-EWS was important in predicting mortality with good discrimination across all ages [22,43]. Adults with low early NEWS scores do subsequently die during a hospital admission and how to predict who is at higher risk of this is not clear [52]. Indices that combine laboratory tests and frailty can indicate a higher risk of mortality in a variety of settings and patient cohorts [43,53,54]. Whilst the LDT-EWS includes far fewer laboratory tests than typical frailty indices it may be that in an acute setting it contributes enough cumulative information to risk of mortality than individual laboratory tests. This study has several strengths. It is the first study reporting factors that influence the power of HFRS to predict LOS and in-hospital mortality across all adult ages in an unselected non-elective cohort. The study used an optimal construction of the HFRS (from index admission through previous admissions over 2 years) [14]. Previous studies have used a cut-off of 10 or more days to indicate prolonged LOS and 30-day in-hospital mortality, which may not be clinically relevant enough to capture very long hospital stays or in-hospital mortality associated with patient complexity. We have used nine different periods of time for LOS and eight different periods for in-hospital mortality. We used a laboratory warning score (comprising a small number of laboratory tests) and vital signs typically available for patients admitted overnight. We also utilised multiple sampling methods and validation checks of the results. While this study demonstrated that a higher HFRS is associated with an increasing risk of longer LOS and in-hospital mortality across all adult ages, it does have some limitations. This study uses a large amount of data, but it is restricted to one hospital, so further research is warranted to see whether the results are replicated in different settings. Also, death was based on the patient’s clinical status at the time of discharge; we could not link records to the General Register Office (death records) to know patients who died (perhaps, shortly) after discharge. We selected a broad non-elective presentation cohort of patients. This most likely reflects the situation of patients living with frailty who often present acutely to hospital with a number of simultaneous complaints and have multiple diagnoses. Future work could focus on whether there is enhanced predictive ability in specific cohorts, for example those with a hip fracture. HFRS was calculated using the original methodology which includes index admission coding, which would not be known at the point of admission, limiting clinical applicability of the HFRS in an acute admission. Previous studies have demonstrated limitations to utility in individuals [10]. A modified HFRS (by excluding index admission from HFRS calculation) has been reported to be correlated with the HFRS and the Clinical Frailty Scale in older patients and had similar associations with outcomes of longer length of stay and mortality [55]. We did not undertake similar analysis to know whether similar findings would apply in our cohort of all ages patients. The calculation of HFRS and the other indices used is likely to be technically easy to apply in hospital systems, however this is reliant on electronic coding of diagnoses, observations and blood test results. Conclusions We conclude that the LDT-EWS can significantly improve the predictive power of HFRS to predict LOS in the hospital whilst the NEWS and LDT-EWS can improve the predictive power of HFRS to predict in-hospital mortality. HFRS combined with NEWS was a good predictor in short-term in-hospital mortality across all adult ages, and HFRS with LDT-EWS for longer-term mortality in all adults admitted urgently to hospital. Therefore, the applications of these results are particularly useful for retrospectively characterising at-risk hospital populations on a macro scale or for use in epidemiological research. Further research is needed and could investigate a number of areas including: whether these findings are replicated in other hospitals, if the findings are replicated in different groups of patients (such as those admitted electively), the performance of an HFRS based only on previous admissions (by excluding current admission), whether targeted intervention in those at higher risk of longer LOS or inpatient mortality using the HFRS+NEWS and HFRS+LDT-EWS results in improved outcomes, and the utility of the combined score in a clinical setting to support identification and guide service development approaches. Supporting information Cross validation. (DOCX) Speciality of discharge. (DOCX) Results of AUROC for 9 period of longer length of stay for each variable alone and HFRS combined with one other variable. (DOCX) Results of AUROC for HFRS with separate laboratory tests models, and HFRS with LDT-EWS models for 9 periods of LOS. (DOCX) S5a-S5h Tables. Results of AUROC for 9 period of LOS according to admission year. (DOCX) S6a-S6d Tables. Results of AUROC for 9 period of LOS according to age groups. (DOCX) S7a-S7b Tables. Results of AUROC for 9 period of LOS according to gender. (DOCX) Results of AUROC for 9 periods of longer length of stay for multivariate models. (DOCX) Results of AUROC for 9 periods of length of stay for data after included patients who died in hospital for each variable alone and HFRS combined with one other variable. (DOCX) AUROC for 8 prediction in-hospital mortality for each variable alone and HFRS combined with one other variable. (DOCX) AUROC for HFRS combined with one other variable according to admission year for 8 periods of in-hospital mortality. (DOCX) AUROC for HFRS alone and HFRS combined with one other variable according to age groups for 8 periods of in-hospital mortality. (DOCX) AUROC for HFRS alone and HFRS combined with one other variable according to gender for 8 periods of in-hospital mortality. (DOCX) AUROC for multivariate models for 8 prediction in-hospital mortality. (DOCX)