Developing a Parsimonious Frailty Index for Older, Multimorbid Adults with Heart Failure Using Machine Learning

Abstract

Frailty is associated with adverse outcomes in heart failure (HF). A parsimonious frailty index (pFI) that predicts outcomes of older, multimorbid patients with HF could be a useful resource for clinicians. Retrospective study of veterans hospitalized from 10/2015 to 10/2018 with HF, age ≥ 50 years, and discharged home developed a 10-item pFI using machine learning from diagnostic codes, labs, vital signs, and ejection fraction (EF) from outpatient encounters. An unsupervised clustering technique identified 5 FI strata: severely frail, moderately frail, mildly frail, prefrail, and robust. We report hazard ratios (HR) of mortality adjusting for age, sex, race, and EF, and odds ratios (OR) for 30-day and 1-year emergency department visits and all-cause hospitalizations post-discharge. We identified 37,431 veterans (age, 73±10; Comorbidity index, 5±3; 43.5% with EF ≤40%). All frailty groups had higher mortality compared with robust group: severely frail (HR, 2.63, [95%CI]:2.42, 2.86), moderately frail (HR, 2.04, [95%CI]:1.87, 2.22), mildly frail (HR, 1.60, [95%CI]:1.47, 1.74), and prefrail (HR, 1.18[95%CI]:1.07, 1.29). Associations between frailty and mortality remained unchanged in stratified analysis by age or EF. The combined (severely, moderately, and mildly) frail group had higher odds of 30-day emergency visits (OR, 1.62, 95%CI: 1.43,1.83) and all-cause readmission (OR,1.75, 95%CI:1.52,2.02), and 1-year emergency visits (OR,1.70,95%CI:1.53,1.89), and rehospitalization (OR,2.18, 95%CI:1.97, 2.41), than the robust group. In conclusion, a 10-item frailty index is associated with post discharge outcomes among patients discharged home following a hospitalization for HF. A parsimonious FI may aid clinical prediction at the point of care.

INTRODUCTION

Heart failure (HF) is a progressively debilitating disease and 80% of patients with HF are 65 years or older.^1-5 Frailty is important in HF (prevalence > 70% ⁵) because older adults with HF and frailty are at increased risk for death, longer hospital stays, and reductions in activities of daily living.⁶ Assessing frailty is challenging in clinical care for a few reasons. First, there are varying definitions of frailty that engenders some confusion.⁷ Second, the cumulative deficits definition of frailty can embed within electronic medical records (EMR). However, this model remains impractical given an expectation that clinicians input values for a large number of heterogeneous variables.⁷ To address some of these challenges, we sought to develop and validate a parsimonious FI (pFI). The propose a pFI model using machine learning techniques over the full breadth of the Veterans Health Administration (VHA) electronic health record to identify deficits via an unbiased approach. We trained the machine learning algorithm to propose a limited set of predictors to develop and validate an efficient pFI.

METHOD

Study Design and Setting.

The design was a retrospective cohort study of veterans using Veterans Health Administration (VHA) to assemble the cohort and extract variables. The methods supplement provides further details on design, methods, and analysis.

Participants / Cohort.

The cohort included all patients admitted to any VHA medical center with a principal diagnosis of HF from October 2015 to October 2018. The International Classification of Diseases (ICD10) hospital discharge codes were used to identify patients with a principal diagnosis of heart failure (I09.9,I11.0, I13.0, I13.2, I25.5, I42.0, I42.5-I42.9, I43.x, I50.x, P29.0). ⁸ Index admissions were excluded if they resulted in a discharge against medical advice, resulted in a transfer to another hospital, the index admission was for observation status, or the patient was discharged from the index admission to a location other than home (e.g., nursing home, skilled nursing facility). For patients with more than 1 eligible heart failure admission, only the first admission was included in this analysis (Supplement Figure 1).

Study Outcomes.

The primary outcome of interest was post-discharge all-cause mortality over the study period. The time to death was measured by subtracting the date of death from the date of hospital discharge. The follow-up period was through March 22, 2021.

The secondary outcomes were 30-day all-cause readmission, 1-year all-cause hospitalization, and 30-day and 1-year emergency department (ED) or urgent care (UC) visits post discharge within VA facilities.

Study Variables and Preparation.

We used all available outpatient data available in EMR data from one year prior to index of admission. (1) We mapped the ICD-10 codes and the Current Procedural Terminology (CPT) codes to 31 geriatric risk factors (deficits) developed previously to build an FI in the VHA (VA-FI).⁹ If the patient had any diagnosis, the related field for geriatric risk score was noted as 1 (a deficit), (Methods-Supplement). (2) We used the ejection fraction (EF) from a VA table that extracts EF from an echocardiogram text document. ¹⁰ If the EF value was less than 40%, then it mapped to 1 or a deficit. (3) Each lab value has a Logical Observation Identifiers, Names and Codes, (LOINC) code. We grouped the LOINC codes together, based on the component, time aspect and system, see methods-supplement. We included lab groups in which ≥ 70% of patients in the cohort had a value. Each LOINC group was mapped as a deficit if the patient had any abnormality. (4) We collected demographic information, such as race (White, Black, and others), age (categorized as 50-65, 65-75, 75-85, and ≥85 years), ethnicity, marital status, and zip code of residence, through the patient domain table.¹¹ (5) We extracted the following vital signs: blood pressure (systolic normal range 90-140, diastolic normal range, 60-90, mean arterial pressure, 70-105), body mass index (BMI, normal range 18.5-30), pulse (normal range 51-99), pulse oximetry, SpO2, (normal range ≥ 95%), respiration (normal range 12-20) and temperature (normal range 96.98-100.4°F).^{12, 13}

FI Models Development.

Most Important Variable Selection:

We developed a cascade of most important variable selection techniques to identify salient features ^{14, 15}. The process used a development training-validation dataset with performance reported using a test dataset. We curated 76 variables (deficits) that have been recommended by previous published FI models (36 labs, 8 vital signs, 31 geriatric risk factors, and EF). First step, univariant filter: We used univariate analysis and any deficits that were significant with p-value < 0.05 remained. Second step, a machine learning algorithm, the least absolute shrinkage and selection operator (LASSO): We used LASSO with 10-fold cross validation setting. The LASSO-Cox is a regression analysis method that performs both variable selection (filter method) and regularization (wrapper method) to identify a smaller list of variables with acceptable accuracy and interpretability¹⁶. Step 3, Survival Random Forest (Surv-RF), We identified the top 10 deficits by applying an additional machine learning algorithm using Surv-RF survival random forests.

Developing the Parsimonious Frail Index (pFI) Model:

We used the accumulation of deficits approach to frailty measurement¹⁷. In this method, we identify the total number of existing deficits divided to the total number of potential deficits. For instance, a patient with four deficits was reported as 0.4 (four exiting deficits divided by 10 potential deficits).

Model Performance.

To measure the performance of models in predicting death and time to death, we used an area under the curve (AUC) for Cox regression analysis variable: timed AUC (Timed-AUC). ¹⁸ We randomly split the database to 70% (number of observations, 26,202) for training-validation and 30% (number of observations, 11,229) for test. The performance was reported for the validation dataset and test dataset.

Comparing to other Frailty Index Model.

The VA Frailty Index (VA-FI) was developed for the general veteran population, based on the Rockwood accumulation of deficits theory and the comprehensive geriatric assessment ^{9, 19}. We compared the AUC of 1-year mortality of VA-FI to the FI models, using a logistic regression. Additionally we compared the performance of pFI to other comorbidity indices (i.e. Charlson Comorbidity index²⁰, Elixhauser comorbidity index ²¹).

Identifying FI Strata.

We used the k-means clustering approach to identify FI strata using R-package stats v3.6.2. In this approach, we identified patterns in patients’ FI that could be summarized into a finite number of groups or clusters. To estimate the optimal number of stable clusters, we used a heuristic elbow method.²² (Methods-supplement)

Statistical Analysis.

We used Cox regression analysis to represent the association between FI strata and mortality. We reported the hazard ratios (HRs) both unadjusted and adjusted using sex, race, and age. To better understand the effect of EF and age, we reported stratified HRs. We also visualized the association using Kaplan-Meier curves with 95% CIs (R-package survminer v0.4.9). The HRs were estimated using SPSS (IBM, V27.0.0).

The association between FI categories and secondary outcomes was analyzed by a binary logistic regression. We reported the odds ratios (ORs) and 95%CIs by considering the robust group as reference. We further combined the severely frail, moderately frail and mildly frail groups as a single frail group for the secondary analyses.

RESULTS

Characteristics of Study Population

Our development cohort included 37,431 participants (Supplemental Figure 1). The average age of the participants was 73.4±10.3, most were men (98%), and 26,046 were White (69.6%). About 47.6% of population were obese with BMIs>30, and 43.5% had EF≤40% (Table 1). The mortality rate during the follow-up interval was 63.1% (n= 23,614), with an average time to death of 3.0±2.1 (Median, 2.6; inter quartile, 0.9-5.5) years post-discharge.

Table 1.

Demographics and characteristics of the cohort.

N	37,431
Sex-Male, N (%)	36676(98.0)
age, M (SD)	73.4(10.3)
age 50-65, N (%)	7016(18.7)
age 65-75, N (%)	15363(41.0)
age 75-85, N (%)	8410(22.5)
age >=85, N (%)	6642(17.7)
BMI, M (SD)	31.4(7.4)
BMI >=30, N (%)	17805(47.6)
Race
White, N (%)	26046(69.6)
Black, N (%)	8966(24.0)
Others, N (%)	2419(6.5)
EF<=40%, N (%)	16289(43.5)
LOS, M days (SD)	4.8(3.8)
LOS >= 6 days, N (%)	10881(29.1)
Comorbidities
Charlson Comorbidity Index, M(SD)	5.0(2.8)
Exhauster Comorbidity Index VW, M(SD)	17.7(9.0)
Myocardial Infarction	7597(20.3)
Congestive Heart Failure	34045(91.0)
Peripheral Vascular Disease	12125(32.4)
Cerebrovascular Disease	6585(17.6)
Dementia	2476(6.6)
Chronic Pulmonary Disease	17688(47.3)
Diabetes without Complications	21748(37.8)
Diabetes with Complications	14163(37.8)
Moderate-Severe Renal Disease	16295(43.5)
Malignancy	5916(15.8)
In Hospital Labs
Natriuretic Peptide Tests (BNP)
N	10,741
Median (IQR)	1,314 (496.0, 4098.0)
Abnormal Values, N(%)	8,212 (76.5)
Troponin
N	21,700
Median (IQR)	0.058 (0.030, 0.110)
Abnormal Values, N(%)	9,629 (44.4)

M(SD) = mean and standard deviation, LOS = length of stay, EF = Ejection Fraction

^†= data gathered till 01/04/2021.

Development of FI Models

We started with 76 variables (labs, 36; vitals, 8; ICD/CPT, 31; EF, 1), the LASSO nominated 42 variables (labs, 15; vitals, 8; ICD/CPT, 18; EF, 1), and the Surv-RF selected the top 10 variables (labs, 0; vitals, 2; ICD/CPT, 7; EF, 1). The final list of 10 variables were SpO2<95%, BMI≥30, EF<40%, atrial fibrillation, anemia, coronary artery disease, cancer, dementia, fatigue, and failure to thrive. The 10 variables used to construct the pFI were based on accumulation of deficits, Table 2. The pFI had the highest timed-AUC in training-validation and test phases. The cross tabulation between VA-Fi and pFI presented in the Supplement table 2. The pFI classified 84% of patients as frail while the VA-FI classified 68.3% as frail. Therefore, more patients fell into prefrail (25.3%) and robust (6.5%) by VA-FI compared to pFI (prefrail, 11.7% and robust, 4.3%).

Table 2.

The performance model as measured by timed-AUC and the number of variables and the type of variables in each model.

Parsimonious Frailty Index Model Performance, Timed-AUC (95%CI)
	Y1	Y2	Y3	Y4	Y5
Validation	0.64 (0.63,0.64)	0.64 (0.63,0.64)	0.64 (0.63,0.65)	0.65 (0.64,0.66)	0.65 (0.64,0.66)
Test	0.64	0.64	0.64	0.65	0.65
Deficits Selected in Each step of Most Variable Selection
	All Variables†	LASSO	Survival Random Forest‡
Labs	36	15	0
Vitals	8	8	2
Geriatric Risk	31	18	7
EF	1	1	1
Total	76	42	10

Timed-AUC (95%CI) = timed area under the curve 95 percentage confidence intervals.

^†= Full listing of variables for standard FI and comprehensive FI are included in supplement table 1.

^‡= The list of 10 variables in the pFI are: 1) blood oxygenation saturation (SpO2) <95%, 2) Body Mass Index≥30, 3) Ejection Fraction<40%, 4) atrial fibrillation, 5) anemia, 6) coronary artery disease, 7) cancer, 8) dementia, 9) fatigue, and 10) failure to thrive.

We compared the AUC for 1-year mortality for VA-FI (AUC, 0.60), the PFI (AUC, 0.64), EF alone (AUC, 0.54), Charlson Comorbidity index (AUC, 0.62), and Elixhauser comorbidity index (AUC, 0.62), Supplemental Table 3.

FI Strata

We observed that the optimal number of strata (or clusters) was 5 (Supplemental Figure 2): robust, prefrail, mildly frail, moderately frail, and severely frail.. The number of deficits for pFI for each stratum was robust, 0, prefrail, 1, frail, 2 or 3, moderately frail, 4, and severely frail, ≥5.

The shortest time to death occurred in the severely frail, and the longest time to death was observed in the robust group (Supplemental Figure 3).. Figure 1 shows results of the Cox regression models for frailty status as a function of time to death. Table 3 displays the HRs and 95%CIs for mortality, adjusted by age, sex, and race for pFI. The highest unadjusted HR in the pFI model was observed in younger patients with age 50-65 (HR, 2.44, 95%CI: 2.89, 4.08). Consistently among the FI models, the HRs remained similar in each frailty strata independent of stratification by age or EF (Table 3).

Figure 1:

Comparing the overall survival by FI strata in parsimonious FI (pFI) by hazard ration and 95 percentage confidence intervals. For pFI, the range of each stratum are presented by the number of deficits.

Table 3:

Number and percentage who died with hazard ratio and 95% confidence intervals for association between frailty strata and survival.

	Time to Death, M(IQR), Year	Death, N (%)	HR (95%CI)
			Unadjusted	Adjusted
Robust	5.5(2.4, 5.5)	618(38.8)	Reference	Reference
Prefrail	5.5(1.8, 5.5)	1985(45.2)	1.24(1.13, 1.35)	1.18(1.07, 1.29)
Mildly Frail	3.2(1.1, 5.5)	8425(58.0)	1.78(1.64, 1.93)	1.60(1.47, 1.74)
Moderately Frail	2.3(0.8, 5.5)	4678(68.6)	2.38(2.19, 2.59)	2.04(1.87, 2.22)
Severely Frail	1.5(0.4, 3.9)	7908(78.3)	3.27(3.01, 3.55)	2.63(2.42, 2.86)
EF > 40%
Robust	5.5(2.4, 5.5)	618(38.8)	Reference	Reference
Prefrail	5.5(1.8, 5.5)	1770(46.2)	1.27(1.16, 1.39)	1.18(1.07, 1.29)
Mildly Frail	3.1(1.1, 5.5)	5214(58.7)	1.82(1.67, 1.98)	1.53(1.40, 1.66)
Moderately Frail	2.3(0.8, 5.5)	2241(69.8)	2.44(2.23, 2.67)	1.90(1.73, 2.08)
Severely Frail	1.5(0.5, 3.7)	2846(78.4)	3.29(3.02, 3.59)	2.40(2.19, 2.62)
EF ≤ 40%
Robust			Reference	Reference
Prefrail	5.5(2.2, 5.5)	215(38.5)	X	X
Mildly Frail	3.4(1.1, 5.5)	3211(56.8)	X	X
Moderately Frail	2.3(0.8, 5.5)	2437(67.6)	X	X
Severely Frail	1.4(0.4, 3.9)	5062(78.2)	X	X
Age 50-65
Robust	5.5(3.9, 5.5)	159(28.1)	Reference	Reference
Prefrail	5.5(3.0, 5.5)	399(32.0)	1.18(0.98, 1.42)	1.18(0.98, 1.42)
Mildly Frail	5.5(2.0, 5.5)	1357(44.3)	1.79(1.51, 2.10)	1.78(1.51, 2.09)
Moderately Frail	3.6(1.2, 5.5)	621(56.9)	2.57(2.16, 3.06)	2.56(2.15, 3.04)
Severely Frail	2.4(0.8, 5.5)	704(66.9)	3.44(2.89, 4.08)	3.40(2.86, 4.04)
Age 65-75
Robust	5.5(2.4, 5.5)	272(38.3)	Reference	Reference
Prefrail	5.5(2.1, 5.5)	828(43.0)	1.17(1.02, 1.34)	1.16(1.01, 1.33)
Mildly Frail	3.8(1.3, 5.5)	3375(54.0)	1.62(1.43, 1.83)	1.61(1.42, 1.82)
Moderately Frail	2.6(0.9, 5.5)	1841(64.5)	2.15(1.89, 2.44)	2.13(1.88, 2.42)
Severely Frail	1.8(0.5, 5.5)	2644(73.1)	2.84(2.50, 3.22)	2.82(2.49, 3.19)
Age 75-85
Robust	4.7(1.7, 5.5)	108(50.9)	Reference	Reference
Prefrail	3.8(1.3, 5.5)	426(54.4)	1.12(0.91, 1.39)	1.12(0.91, 1.39)
Mildly Frail	2.7(1.0, 5.5)	1960(64.0)	1.44(1.18, 1.74)	1.44(1.18, 1.75)
Moderately Frail	2.1(0.7, 5.5)	1149(72.7)	1.84(1.51, 2.24)	1.83(1.51, 2.24)
Severely Frail	1.4(0.4, 3.7)	2185(78.9)	2.33(1.92, 2.82)	2.33(1.92, 2.82)
Age ≥ 85
Robust	2.3(1.1, 5.1)	79(75.2)	Reference	Reference
Prefrail	1.9(0.8, 4.7)	332(76.1)	1.07(0.84, 1.37)	1.07(0.84, 1.37)
Mildly Frail	1.6(0.5, 3.5)	1733(80.5)	1.27(1.01, 1.59)	1.26(1.01, 1.58)
Moderately Frail	1.4(0.4, 3.1)	1067(82.9)	1.39(1.10, 1.74)	1.38(1.10, 1.74)
Severely Frail	0.9(0.3, 2.3)	2375(89.3)	1.82(1.46, 2.28)	1.81(1.45, 2.27)

^†= the Hazard Ratio adjusted by age, sex, race, and ejection fraction.

Healthcare Utilization

Table 4 displays healthcare utilization post-discharge among the frail, prefrail, and robust groups. We observed 62% (OR,1.62, 95%CI:1.43,1.83) higher odds of 30-day ED/UC visit in frail group and 11% (OR,1.11, 95%CI:0.97,1.28) in prefrail group compared to robust group. We observed 70% (OR,1.70,95%CI:1.53,1.89) higher odds of 1-year ED/UC visit in frail group and 15% (OR,1.15, 95%CI:1.02,1.29) in prefrail group compared to robust group. The odds of 30-day readmission among the frail group was higher than in the prefrail and robust groups. We observed 75% (OR,1.75, 95%CI:1.52,2.02) higher odds of 30-day readmission in frail group and 11% (OR,1.11, 95%CI:0.95,1.31) in prefrail group compared to robust group. We observed 2.18-fold (OR,2.18, 95%CI:1.97,2.41) higher odds of 1-year hospital admission in frail group and 33% (OR,1.33,95%CI:1.19,1.50) in prefrail group compared to robust group.

Table 4.

Comparing the healthcare utilization after discharge between robust, pre-frail and frail group where frail group is composed of severely frail, moderately frail, and Mildly Frail.

	Robust N (%)	Prefrail N (%)	Frail N (%)	OR(95%CI)†
				Prefrail vs Robust	Frail vs Robust
30-Day ED/UC visits	331(20.8)	990(22.6)	9374(29.8)	1.11(0.97, 1.28)	1.62(1.43, 1.83)
1-Year ED/UC visits	1018(63.9)	2941(67.0)	23611(75.1)	1.15(1.02, 1.29)	1.70(1.53, 1.89)
30-Day Readmission	224(14.1)	676(15.4)	7006(22.3)	1.11(0.95, 1.31)	1.75(1.52, 2.02)
1-Year Readmission	730(45.8)	2327(53.0)	20387(64.8)	1.33(1.19, 1.50)	2.18(1.97, 2.41)

ED: emergency department visit, UC: urgent care.

^†OR (95%CI): odds ratio (95 percentage confidence interval).

DISCUSSION

The current study provides evidence for the predictive validity of a novel parsimonious FI for older patients admitted to the hospital with HF. Notably, the pFI is exclusively developed for HF patients. The pFI is robust, as it uses a national sample that integrates baseline outpatient data from the prior 12 months and follows patients posthospitalization for multiple years; and it uses an innovative cascade of machine learning techniques. The pFI only requires 10 variables and can thus improve clinical interpretation and integration into clinical care. Even with just 10 variables, the pFI effectively discriminates strata of frailty severity, with accurate prediction of mortality, even after adjustment for age and EF; and strong associations with 30-day and 1-year outcomes for ED and urgent care (UC) use and hospital readmission.

Awareness of baseline frailty status at the time of hospital admission is important for HF patients, as it predicts adverse health outcomes after discharge ^{23, 24}. This valuable prognostic information helps patients and their family members better understand their overall health trajectory and helps clinicians with treatment planning and care transitions. For example, there is data showing that outcomes in patients undergoing LVADs are worse among those who are frail compared to those who are not frail ²⁵. Moreover, frailty is used in decision-making related to TAVRs and cardiac surgery, and some call for incorporation of frailty into decision-making related to pharmacotherapy²⁶. This supports the need for a parsimonious frailty tool that permits clinicians to more easily identify frailty in routine care. A parsimonious frailty index facilitates communication and comprehension of risk between clinicians and patients. Future studies constructing easy-to-use digital applications or EHR dashboards are needed for additional pragmatic validation.

Additionally, identifying HF patients at pre-frail stage is important. Prefrail stage, an intermediate stage that precedes frailty, may be potentially reversable with effective multimodality interventions²⁷. For example, the evidence-based Rehabilitation Therapy in Older Acute Heart Failure Patients (REHAB-HF) intervention targeting four physical-function domains (strength, balance, mobility, and endurance) is well-suited for patients at a prefrail stage²⁸. A systematic review showed that interventions that incorporate physical exercise and nutrition support (caloric, protein support, and vitamin D supplementation), and avoid polypharmacy can reverse patients’ frailty status²⁹.

There are several reasons why this parsimonious pFI may be superior to other frailty indices. First, like previous FI models, we used the Searle et al (2008) approach to construct FI using full breadth of a national EHR on various sets of diagnostics codes, services codes, or labs (Supplemental Table 1). ^{9, 30-32} Second, similar to Kim et al., ³⁰ we used a machine learning technique, LASSO, to identify deficits and tune the FI. We further used machine learning to develop a pFI model. Third, the innovation of the current study is the use of survival analysis to model time to death as the outcome variable for the machine learning algorithm. FI models typically consider death as a constant factor at baseline, which does not allow differentiation of death at 3 months versus 11 months in the same year. Fourth, We used unsupervised machine learning algorithm to identify clinical interpretable strata. Clegg et al. developed an electronic FI using population quartiles to define robustness, mild frailty, moderate frailty, and severe frailty.³³ The VA-FI model by Orkaby et al. used cut-points to categorize frailty into nonfrail, prefrail, mildly frail, moderately frail, and severely frail strata ⁹. The current study used a heuristic approach, k-means clustering, to estimate the optimum number of strata (Supplemental Figure 1). This approach yielded 5 strata, similar to Orkaby et al., ⁹ with distinctly different odds at predicting mortality and other clinical outcomes in this HF focused population. For the pFI model, these strata translate into pragmatically measured simple counts of 10 clinically relevant variables [0, 1-2, 3, 4, 5+]. Fifth, the proposed pFI allocated patients into the higher frailty strata compared to VA-FI. This finding is in agreement with previous report that HF patient experiences higher frailty prevalence (60% to 90%)³⁴ compared to general older adult population age ≥ 65 years (10% to 25%)³⁵. This finding highlights the importance of using an FI designed specifically for the HF population. Sixth, the proposed pFI met predictive validity standards by predicting death and subsequent adverse healthcare outcomes. Prior studies described the association of frailty with higher ED and hospital use among HF patients over a two year follow-up. ³⁶ We observed higher odds of ED visits and hospitalization at 1 month and 1 year follow-up among frail and prefrail compared with robust clusters. In some instances, frailty may be modifiable with appropriate interventions. We observed BMI ≥ 30 as one of the final candidate variables in the parsimonious model. This finding is in-line with previous report that the prevalence of frailty has a u-shape association with BMI ³⁷. In the other words, both end of BMI spectrum suffers from higher rate of frailty.

In this study, we reported that a parsimonious frailty index with 10 specific deficits curated from the full breadth of electronic medical records has better predictive validity compared to previous approaches independent of age and ejection fraction. Parsimonious frailty index is a practical clinical decision support tool for day-to-day use.

STRENGTHS AND WEAKNESSES

This study has several strengths. We used data from the largest integrated national healthcare system in the United States. The VHA is a good representation of race and ethnicity groups in the US, although women are underrepresented. Evidence suggests that users of the VA are comparable to patients using Medicare.³⁸ However, the limitations of a largely male, Veteran sample limit the generalizability of this model to other healthcare systems. Additional validation is needed to establish external validity. At the level of cohort definition, we excluded patients who discharged to institutional settings such as skilled nursing homes as those patients are often understood as frail without the need for computational models. Additionally, we used recommended ICD10 codes to define patients with HF which is prone to error, although there is report that showed these codes produced 88% positive predictive value compared to manual chart review⁸. Another limitation is the incomplete capture of veteran users who received care and were hospitalized in non-VHA facilities. This can be addressed in future studies by including Medicare data. In this study, we did not search for new deficits that might explain the frailty status. We instead used the deficits recommended in the literature to build an FI. We used machine learning algorithms to identify relevant deficits in HF patients and to identify the frailty strata. We did not couple the machine-learning algorithm with natural language processing techniques to extract data from unstructured free-text patient notes, such as the New York Heart Association (NYHA).

CONCLUSIONS

A pFI with 10 variables has similar predictive validity as the VA-FI. The pFI is thus a feasible and practical tool that can be incorporated into routine practice, and provides clinically useful insights that are distinct from exiting mortality prediction models ³⁹ especially for older adults with multiple comorbidities. The pFI index applied at the point of care could guide clinical decisions regarding care transitions, urgency of interventions to reduce readmissions, and appropriateness of earlier referral to palliative care. Future research to test and evaluate clinical applications of the pFI are needed.

PERSPECTIVES

COMPETENCY IN MEDICAL KNOWLEDGE:

We propose a parsimonious frailty index with 10 specific deficits derived from full breadth of electronic medical records from outpatient settings using machine learning algorithm. The parsimonious frailty index has better predictive validity compared to previous approaches independent of age and ejection fraction in predicting mortality and healthcare utilization following hospital discharge for congestive heart failure

TRANSLATIONAL OUTLOOK:

Parsimonious frailty index is a practical clinical decision support tool for day-to-day inpatient use.