Development and validation of the Tool for Pharmacists to Predict 30-day hospital readmission in patients with Heart Failure (ToPP-HF)

Abstract

Purpose

Pharmacists are well positioned to provide transitions of care (TOC) services to patients with heart failure (HF); however, hospitalizations for patients with HF likely exceed the capacity of a TOC pharmacist. We developed and validated a tool to help pharmacists efficiently identify high-risk patients with HF and maximize their potential impact by intervening on patients at the highest risk for 30-day all-cause readmission.

Methods

We conducted a retrospective cohort study including adults with HF admitted to a health system between October 1, 2016, and October 31, 2019. We randomly divided the cohort into development (n = 2,114) and validation (n = 1,089) subcohorts. Nine models were applied to select the most important predictors of 30-day readmission. The final tool, called the Tool for Pharmacists to Predict 30-day hospital readmission in patients with Heart Failure (ToPP-HF) relied upon multivariable logistic regression. We assessed discriminative ability using the C statistic and calibration using the Hosmer-Lemeshow goodness-of-fit test.

Results

The risk of 30-day all-cause readmission was 15.7% (n = 331) and 18.8% (n = 205) in the development and validation subcohorts, respectively. The ToPP-HF tool included 13 variables: number of hospital admissions in previous 6 months; admission diagnosis of HF; number of scheduled medications; chronic obstructive pulmonary disease diagnosis; number of comorbidities; estimated glomerular filtration rate; hospital length of stay; left ventricular ejection fraction; critical care requirement; renin-angiotensin-aldosterone system inhibitor use; antiarrhythmic use; hypokalemia; and serum sodium. Discriminatory performance (C statistic of 0.69; 95% confidence interval [CI], 0.65-0.73) and calibration (Hosmer-Lemeshow P = 0.28) were good.

Conclusions

The ToPP-HF performs well and can help pharmacists identify high-risk patients with HF most likely to benefit from TOC services.

Key Points.

• Pharmacists are well positioned to provide transitions of care interventions at hospital discharge, but the number of hospitalized patients at risk for rehospitalization typically exceeds pharmacist resources.

• User-friendly risk prediction tools could help pharmacists efficiently identify high-risk patients who are most likely to benefit from transitions of care services.

• The ToPP-HF tool was designed specifically for pharmacists to identify patients with heart failure who are at the greatest risk for 30-day all-cause hospital readmission and includes 13 variables that are easily collectable through the electronic medical record.

Despite nationwide efforts, rehospitalization remains a common occurrence for patients with heart failure (HF), with over 20% experiencing 30-day readmission.^1-3 Although nationwide programs and reimbursement policies have incentivized institutions to implement programs focused on reducing 30-day readmissions, hospitals have seldom been successful at significantly lowering readmission rates in patients with HF.⁴ As substantial resources are required to improve outcomes in this complex patient population, and institutional resources are finite, there is an urgent need for novel tools that can assist with prioritizing high-risk patients with the greatest need for intervention.

Pharmacists are well positioned to manage complex diseases like HF through medication reconciliation, optimizing guideline-directed pharmacotherapy, therapeutic drug monitoring, promoting medication adherence, mitigating adverse effects, and providing patient education.⁵ Although clinical pharmacy services may reduce the risk of hospitalization by 29% to 54% when delivered in both inpatient (prior to hospital discharge)^6-8 and outpatient (post–hospital discharge)^9-12 settings, pharmacists remain underutilized in clinical practice. Adding pharmacists to the multidisciplinary care team can reduce the burden on other providers and ensure safer and higher-quality care is expanded to a larger number of patients.

Due to finite institutional resources and the increasing prevalence of HF as the United States population ages,¹³ it is not feasible for pharmacists to provide targeted interventions to every patient with HF prior to and post hospital discharge. In the absence of a premade risk stratification tool that can be used to screen patients, pharmacists are responsible for manually reviewing an extensive number of medical records to decide if a patient would be considered at high risk for readmission, which is resource-intensive and prone to error. While several readmission models have been created for patients with HF, including multiple that used machine learning algorithms,^14-20 they have critical limitations restricting their uptake into clinical practice, especially pharmacy practice. These limitations include (1) reliance on a large number of variables that take too much time for a pharmacist or practicing clinician to collect, (2) the use of variables or data sources that are difficult for pharmacists or other clinicians to access, and (3) the lack of a clear end-user, resulting in tools that are difficult to embed into ongoing routine workflow.

The creation of a pharmacist user–friendly prediction tool for 30-day readmission, based on clinically relevant and readily available information, would equip pharmacists with a resource to efficiently prioritize patients at greatest risk for medication review, patient education, and early outpatient follow-up after hospital discharge, which is critical for those with HF.^21,22 Therefore, we developed and validated a pragmatic risk prediction tool for 30-day all-cause unplanned hospital readmission that could be applied to any hospitalized adult patient with HF, regardless of the primary admission diagnosis. We hypothesized that our prediction tool would have similar discriminatory performance and calibration to previously developed rehospitalization models for HF but would be generalizable to a larger group of patients with HF and could be easily integrated into routine clinical practice.

Methods

Study design and data source.

A retrospective cohort study was conducted to develop and validate a tool for predicting 30-day all-cause unplanned hospital readmissions among adult patients with HF. Data were ascertained from laboratory and vital sign flow sheets, procedure and surgery documentation, medical history diagnoses, medication administration records, and demographics documented in the electronic medical records (EMRs) of inpatients admitted to the hospital between October 1, 2016, and October 31, 2019. We collected data that was available in the EMR during the index hospital admission. The majority of data were measured from records generated during the index hospitalization; however, if information was not available during the index hospital admission (eg, left ventricular ejection fraction [LVEF]), we ascertained older EMR data.

Study population.

The source population included adult (≥18 years of age on the date of hospital admission) patients with a documented diagnosis of HF admitted to 2 hospitals within a health system during the study period. Patients with HF were included in the source population regardless of whether their index hospital admission was for HF (eg, HF exacerbation) or not (eg, fall, fracture). Although the Centers for Medicare and Medicaid Services (CMS) Hospital Readmissions Reduction Program measure and previous readmission prediction models for HF have limited index hospital admissions to those of patients with a principal diagnosis of HF,^20,23,24 our rationale was to ensure that the results and final tool were generalizable to a larger group of patients with HF. Prior studies have noted that HF diagnosis codes in the principal position have only modest sensitivity and may miss a large number of patients for whom HF was an important cause or the primary cause of the hospital admission.^25,26 The Healthcare Cost and Utilization Project Clinical Classifications Software was used to identify International Classification of Diseases, 10th Revision (ICD-10) codes associated with HF.²⁷ Patients were included if an ICD-10 code for HF was documented in the hospital admission problem list in any position (eTable 1). Since patients could have had multiple hospital admissions during the study period, we randomly sampled 1 per patient as the index hospitalization (from which readmissions were assessed) so that each patient had a single index hospitalization. This approach was taken to avoid implicitly weighting the results to be more generalizable to patients with higher healthcare utilization, which would have occurred if all hospitalizations had been included, or to be more generalizable to the earlier time periods (either closer to the October 2016 data collection start date or earlier in the patients’ disease course), which would have occurred if the first hospitalization had been sampled instead. Additionally, patients with Veterans Affairs (VA) insurance only, those discharged to the department of corrections or hospice care, who left against medical advice, or who were transferred to another hospital outside the health system were excluded due to small subgroup size or were considered a unique group that would likely require specialized transitions of care (TOC) interventions. Those who died within the 30-day follow-up period were excluded since death is a competing outcome that would complicate the interpretation of our results. Those with complete data were included in the study population, then randomly separated into the development and validation subcohorts (eFigure 1).

Outcome variable: 30-day all-cause unplanned readmission.

Patients were followed for 30 days after hospital discharge to determine their readmission status. Patients could have either of 2 values for the outcome variable: unplanned readmission within 30 days or remaining alive and without an unplanned readmission at 30 days. Patients who were readmitted to any hospital with a shared EMR within 30 days from the day of hospital discharge were counted as having the outcome of interest, regardless of the principle diagnosis for the readmission. We chose a follow-up period of 30 days and the outcome of all-cause hospital readmission because they are aligned with CMS reimbursement policies and programs.^23,24 Planned treatment follow-up or planned chemotherapy was not included in the outcome of interest, as our goal was to assess only unplanned 30-day readmissions. Additionally, patients who returned to the emergency department but were not subsequently admitted to an inpatient unit were not counted as having the outcome of interest. Study follow-up concluded on November 29, 2019, 30 days after the latest date of hospital discharge in our study population.

Potential predictors.

One hundred potential predictor variables were collected, including measures of demographics, vital signs, laboratory values, procedures, medications, comorbidities, and healthcare utilization (eTable 2). Variables relevant to explaining unplanned readmissions among individuals with HF were based on published literature and consensus between coinvestigators.^28,29 To ensure usability of the final tool, individual factors associated with unplanned readmission were, to the best of our team’s knowledge, readily collectable through all EMR systems. Variables were not considered for inclusion in the prediction models if we observed a low prevalence (<2% of patients) because these variables are less likely to be impactful. Additionally, some variables were excluded if they were highly colinear (ie, had a variance inflation factor of ≥5) with others. After these exclusions, we considered 52 potential predictors, from which we planned to identify a smaller core set that would comprise the final risk prediction tool.

Development and validation of the risk prediction tool.

The study population was randomly divided into development (66% of sample) and validation (33% of sample) subcohorts. Baseline characteristics and readmission status of the development versus validation subcohort were compared using proportions for categorical variables and medians with 25th and 75th percentiles for continuous variables. For the purpose of identifying which variables were most important to include in the final model on which the prediction tool would be based, we implemented 9 variable selection models in the development subcohort using machine learning algorithms and standard regression models. These included least absolute shrinkage and selection operator (LASSO)(with and without variables prespecified by the research team), random forest, elastic net (with and without variables prespecified by the research team), least angle regression (LARS), kernel regularized least squares (KRLS), boosted regression, and backward stepwise multivariable logistic regression (MLR) models.³⁰ The C statistic was used to evaluate the models’ ability to discriminate between readmission versus no readmission in the development subcohort, which helped to inform the optimal number of variables to include in the final prediction tool as well as which combinations of variables were most important. The variables selected were compared across models, and variables that appeared in multiple models were considered for inclusion in a final prediction model and tool. We ultimately aimed to include less than 15 variables in our final risk prediction tool because we felt more variables would be too cumbersome to collect during the course of routine clinical practice. However, we also aimed to include the lowest number of variables while maintaining good discriminatory performance.

The final set of variables was entered into a logistic regression model that was estimated in the development subcohort. The final prediction model was then validated in the validation subcohort. In the development subcohort, scores for the final tool (ToPP-HF) were assigned to each level of each variable in the final model by dividing all other beta coefficients from the regression model by the smallest beta coefficient and then rounding to the nearest integer. Both positive and negative values were permitted. An integer-based points system was also assigned to ensure the tool was easily calculable. To calculate the final ToPP-HF score, a pharmacist summed the integers for each variable included in the tool. The final score distribution was categorized to classify patients above the 90th percentile of the distribution as being at high-risk, those in the 81st to 90th percentile as being at moderate to high risk, those in the 51st to 80th percentile as being at moderate-risk, and those at or below the 50th percentile as being at low risk for readmission. In the validation subcohort, the C statistic was used to evaluate the final tool’s performance, and the Hosmer-Lemeshow test was used to determine calibration.

Software.

Data were analyzed using SAS, version 9.4 (SAS Institute, Inc., Cary, NC) and Stata, version 15.0 (StataCorp, College Station, TX) software.

Ethics approval.

The institutional review board of Lifespan Health System approved the study protocol. Informed consent was not required.

Results

Participants.

Of the 7,177 encounters included in the source population, 3,203 encounters were included in the final sample population after applying exclusion criteria and randomly sampling one index hospitalization per patient (eFigure1). This cohort was randomly split into the development (n = 2,114) and validation (n = 1,089) subcohorts. The overall mean (standard deviation [SD]) age was 75 (14) years, and the subcohorts included a total of 1,661 males (52%) and 2,617 patients of white race (82%). The overall median (interquartile range [IQR]) LVEF was 50% (30, 60). HF was the principal admission diagnosis for 1,771 patients (55%) overall, and the median (IQR) length of hospital stay was 5 (3, 8) days. The 30-day all-cause unplanned readmission outcome occurred in 16.7% of the overall study population. Patient characteristics were similar between the development and validation subcohorts (Table 1, eTable 3).

Table 1.

Selected Study Population Characteristics From Index Hospital Admission Stratified by Subcohort^a

	No. (%)b
Characteristics	Development (n = 2,114)	Validation (n = 1,089)
Demographics
Age, mean (SD), years	75 (14)	76 (14)
Male	1,091 (51.6)	570 (52.3)
White race	1,730 (81.8)	887 (81.5)
Hispanic ethnicity	189 (8.9)	98 (9.0)
Insurance type
Commercial only	315 (14.9)	156 (14.3)
Medicare only	1,339 (63.3)	706 (64.8)
Medicaid only	116 (5.5)	68 (6.2)
Medicare and Medicaid	285 (13.5)	131 (12.0)
No insurance	59 (2.8)	28 (2.6)
Medical history from index hospital admission
Atrial fibrillation or flutter	1,025 (48.5)	527 (48.4)
Chronic kidney disease	673 (31.8)	365 (33.5)
Chronic obstructive pulmonary disease	522 (24.7)	248 (22.8)
Clinical ASCVDc	1,275 (60.3)	685 (62.9)
Diabetes mellitus	937 (44.3)	483 (44.4)
Hyperlipidemia	1,225 (57.9)	665 (61.1)
Hypertension	1,782 (84.3)	928 (85.2)
No. of comorbidities, median (IQR)d	9 (6, 12)	9 (6, 12)
Index hospital admission data
Length of hospital stay, median (IQR), days	5 (3, 8)	5 (3, 8)
Any critical care during admission	175 (8.3)	74 (6.8)
Principal admission diagnosis of HF	1,175 (55.6)	596 (54.7)
Hospital admissions in previous 6 months
0	1,153 (54.5)	586 (53.8)
1	542 (25.6)	287 (26.4)
≥2	419 (19.8)	216 (19.8)
Discharge location
Home	1,498 (70.9)	743 (68.2)
Institutional postacute caree	616 (29.1)	346 (31.8)
Last recorded ejection fraction
≥50%	1,111 (52.6)	563 (51.7)
40%–49%	280 (13.2)	144 (13.2)
<40%	723 (34.2)	382 (35.1)
Admission BNP, median (IQR), pg/mL	551.7 (248.7, 1,145.4)	597.2 (283.7, 1,147.6)
Admission SBP, median (IQR), mm Hg	140 (121, 160)	140 (120, 161)
Admission DBP, median (IQR), mm Hg	77 (65, 90)	76 (64, 89)
Admission SCr, median (IQR), mg/dL	1.18 (0.9, 1.7)	1.17 (0.9, 1.7)
Discharge medications
ACEI, ARB, or ARNI	1,164 (55.1)	576 (52.9)
Beta-blocker	1,695 (80.2)	889 (81.6)
Aldosterone antagonist	419 (19.8)	202 (18.5)
Loop diuretic	1,827 (86.4)	944 (86.7)
Statin	1,487 (70.3)	760 (69.8)
Aspirin	1,433 (67.8)	741 (68.0)
Antiarrhythmic	188 (8.9)	110 (10.1)
30-day unplanned readmission	331 (15.7)	205 (18.8)

Abbreviations: ACEI, angiotensin-converting enzyme inhibitor; ARB, angiotensin II receptor blocker; ARNI, angiotensin receptor neprilysin inhibitor; ASCVD, atherosclerotic cardiovascular disease; BNP, B-type natriuretic peptide; DBP, diastolic blood pressure; HF, heart failure; IQR, interquartile range; SBP, systolic blood pressure; SCr, serum creatinine; SD, standard deviation.

^aSome percentages have been rounded and may not total 100. Additional patient characteristics were collected and are represented in eTable 3.

^bUnless otherwise indicated.

^cClinical ASCVD was defined as history of acute coronary syndrome, myocardial infarction, stable or unstable angina or coronary or other arterial revascularization, stroke, transient ischemic attack, or peripheral artery disease including aortic aneurysm, all of atherosclerotic origin.

^dIncludes sum of chronic conditions and major events (eg, myocardial infarction, stroke). Duplicate conditions were counted once (eg, chronic heart failure and heart failure with reduced ejection fraction).

^eInstitutions include skilled nursing, rehabilitation, or long-term acute care facility.

Variable selection for the ToPP-HF tool.

Across the 9 variable selection models, the number of variables selected ranged from 13 (in the LASSO model) to 52 variables (in the random forest, KRLS, and boosted regression models). Variables contained within a majority (≥5) of the models included number of hospital admissions in the 6 months prior to index hospitalization, primary admission diagnosis of HF for index hospitalization, number of scheduled medications on hospital discharge, diagnosis of chronic obstructive pulmonary disease (COPD), total number of comorbidities, estimated glomerular filtration rate (eGFR) on hospital admission, eGFR on hospital discharge, diagnosis of chronic kidney disease (CKD), most recent LVEF value, and any endoscopic intervention into the gastrointestinal tract during index hospital admission. All models performed moderately well, with C statistics ranging from 0.63 (95% confidence interval [CI], 0.59-0.67) for the random forest model to 0.68 (95% CI, 0.64-0.72) for the LASSO method (eTable 4).

Final set of variables included in the ToPP-HF.

By comparing variables common across multiple models and integrating that information with clinical knowledge, our research team selected the final predictors to be included in the ToPP-HF tool. These 13 variables included number of hospital admissions in the 6 months prior to index hospitalization; primary admission diagnosis of HF for index hospitalization; number of scheduled medications on hospital discharge; diagnosis of COPD; total number of comorbidities; eGFR on hospital admission; index hospitalization length of stay; most recent LVEF value; any critical care during hospital admission; angiotensin-converting enzyme inhibitor (ACEI), angiotensin II receptor blocker (ARB), or angiotensin receptor neprilysin inhibitor (ARNI) use on hospital discharge; antiarrhythmic use on hospital discharge; hypokalemia (serum potassium concentration of <3.5 mEq/L) on hospital discharge; and serum sodium level on hospital discharge. Individual medications included in the ACEI/ARB/ARNI and antiarrhythmic classes are listed in eTable 5.

ToPP-HF scores and performance.

Calculated by summing the integers for each variable in the tool (Table 2), the ToPP-HF score ranges from –30 to 49, with a score of –30 to –1 defined as low risk; 0 to 11, moderate risk; 12 to 16, moderate to high risk; and 17 to 49, high risk of unplanned readmission within 30 days of hospital discharge. The estimated risk of all-cause unplanned readmission within 30 days in the validation subcohort was 32.4% (vs an observed readmission rate of 39.1%) for high-risk patients, 24.2% (vs an observed rate of 29.5%) for moderate- to high-risk patients, 17.9% (vs an observed rate of 22.2%) for moderate-risk patients, and 8.7% (vs an observed rate of 10.0%) for low-risk patients (Table 3). Patients in the study population had ToPP-HF scores ranging from –28 to 41. In the validation subcohort, the final ToPP-HF tool performed well, with a C statistic of 0.69 (95% CI, 0.65-0.73) and good calibration (Hosmer-Lemeshow P = 0.28). Relative to patients in the low-risk category, those in the high-risk category had more than 5 times greater odds of unplanned readmission within 30 days (odds ratio, 5.78; 95% CI, 3.58-9.32) (eTable 6).

Table 2.

ToPP-HF Scoring in the Development Subcohort^a

Variable	Score in ToPP-HF Toolb	Beta Coefficient	Odds Ratio (95% CI)	P Value
Hospital admissions in previous 6 months
0	0	Reference
1	5	0.32	1.38 (1.02-1.86)	0.04
≥2	11	0.67	1.95 (1.43-2.65)	<0.001
Primary admission diagnosis of HF
No	0	Reference
Yes	–4	–0.23	0.79 (0.62-1.01)	0.06
No. of scheduled medications on discharge (excludes PRN medications)
≤8	0	Reference
9–13	4	0.23	1.26 (0.90-1.77)	0.18
≥14	7	0.43	1.54 (1.05-2.25)	0.03
No. of comorbiditiesc
≤7	0	Reference
8–10	2	0.12	1.13 (0.81-1.58)	0.47
≥11	4	0.22	1.25 (0.89-1.74)	0.20
Diagnosis of COPD
No	0	Reference
Yes	3	0.18	1.20 (0.90-1.58)	0.21
eGFR on hospital admission (mL/min/1.73 m2)
≤29	0	Reference
30–44	–2	–0.14	0.87 (0.61-1.24)	0.45
45–59	–5	–0.31	0.74 (0.52-1.04)	0.08
≥60	–10	–0.57	0.57 (0.39-0.82)	<0.01
Hospital admission length of stay, days
≤3	0	Reference
4–5	3	0.19	1.21 (0.87-1.69)	0.26
6–8	2	0.12	1.13 (0.79-1.61)	0.52
≥9	5	0.27	1.31 (0.94-1.82)	0.11
Recent ejection fraction closest to hospital discharge, %
≤39	0	Reference
40–49	8	0.49	1.63 (1.13-2.35)	<0.01
≥50	–1	–0.06	0.94 (0.71-1.26)	0.69
Any critical care during admission
No	0	Reference
Yes	5	0.31	1.36 (0.90-2.06)	0.14
Prescribed ACEI, ARB, or ARNI on hospital discharged
No	0	Reference
Yes	–2	–0.13	0.88 (0.68-1.15)	0.34
Prescribed antiarrhythmic on hospital discharged
No	0	Reference
Yes	–9	–0.52	0.60 (0.37-0.96)	0.03
Hypokalemia on hospital discharge (<3.5 mEq/L)
No	0	Reference
Yes	6	0.33	1.40 (0.97-2.01)	0.07
Serum sodium on hospital discharge, mEq/L
<135	0	Reference
≥135	–4	–0.24	0.79 (0.58-1.07)	0.13

Abbreviations: ACEI, angiotensin-converting enzyme inhibitor; ARB, angiotensin II receptor blocker; ARNI, angiotensin receptor neprilysin inhibitor; CI, confidence interval; COPD, chronic obstructive pulmonary disease; eGFR, estimated glomerular filtration rate; HF, heart failure; PRN, as-needed.

^aThe ToPP-HF tool performed well, with a C statistic of 0.69 (95% CI, 0.65-0.73) and good calibration (Hosmer-Lemeshow P = 0.28).

^bThe potential scores of the ToPP-HF tool range from –30 to 49, with a score of –30 to –1 defined as low risk; 0 to 11, moderate risk; 12 to 16, moderate to high risk; and ≥17, high risk of readmission within 30 days.

^cIncludes sum of chronic conditions and major events (eg, myocardial infarction, stroke). Duplicate conditions were counted only once (eg, chronic heart failure and heart failure with reduced ejection fraction).

^dIndividual medications included in each class are listed in eTable 5.

Table 3.

Calibration Information for ToPP-HF Tool in Validation Subcohort After Risk Categorization (n = 1,089)^a

ToPP-HF Score	Risk Category	No. Patients in Risk Category	No. 30-Day Readmissions	Observed Readmission Rate, %	Estimated Readmission Risk, %
–30 to –1	Low	520	52	10.0	8.7
0 to 11	Moderate	347	77	22.2	17.9
12 to 16	Moderate-high	112	33	29.5	24.2
17 to 49	High	110	43	39.1	32.4

^aThe ToPP-HF tool performed well, with a C statistic of 0.69 (95% CI, 0.65-0.73) and good calibration (Hosmer-Lemeshow P = 0.28).

Since it would be nearly impossible for a pharmacist to intervene on all patients with HF prior to hospital discharge due to time constraints and limited personnel resources, on the basis of the tool’s performance in the validation subcohort, we estimate that by applying a TOC intervention to all patients with scores of ≥17 (indicating high readmission risk), the intervention would be applied to 10.1% of patients with HF admitted to the hospital during the study period. These proportions would be 10.3%, 31.9%, and 47.8%, respectively, for cutoffs of 12 to 16 points (moderate to high risk), 0 to 11 points (moderate risk), and –30 to –1 points (low risk).

Discussion

In the retrospective cohort study described here, we developed and validated the ToPP-HF tool, which predicts 30-day all-cause unplanned readmissions for adult patients with HF. Our tool has good discriminatory performance (C statistic = 0.69) and calibration (Hosmer-Lemeshow P = 0.28). We designed our tool to be applied to any hospitalized patient with HF regardless of the primary admission diagnosis. It is intended for use by inpatient pharmacists to guide TOC interventions prior to hospital discharge. Targeting patients in the high-risk category based on the ToPP-HF score would account for an estimated 10.1% of patients with HF admitted to the hospital. Although the intent of the tool was to guide inpatient TOC pharmacist workflow, future research should examine the best point in the TOC continuum at which to use the ToPP-HF. An alternative strategy may be to calculate the ToPP-HF score to guide referrals for ambulatory care or outpatient TOC pharmacist visits after hospital discharge.

Although some prior models that predict readmission in patients with HF performed better than the ToPP-HF, critical limitations restrict their uptake into clinical practice. For example, studies that used historical data (eg, data obtained prior to 2017) would not account for more recent HF treatment guidelines that favor use of an ARNI in select patients with chronic symptomatic HF with reduced ejection fraction.²⁹ Additionally, some models included data focused on a specific subset of patients (eg, VA patients) and may not be generalizable to a broader population of patients with HF. Models based on registry data pose a major issue, as some variables may not be readily available to the practicing clinician in the EMR (eg, household income, high school degree, years of college). Finally, and most importantly, all published models included a large number of predictors, were not created into a user-friendly tool, or were not designed for a specific end-user (eg, a clinical pharmacist). Because it includes only 13 predictors, the ToPP-HF score can be easily calculated during routine clinical pharmacy practice.

We hypothesized that the ToPP-HF would have similar discriminatory performance to established readmission prediction models for patients with HF but would be easier to use in clinical practice. With these established models, which were developed using both machine learning algorithms and standard regression, reported C statistics ranging between 0.53 to 0.84.^14-20 Notably, reported C statistics for many models were less than 0.7, and nearly all models included patients with a primary admission diagnosis of HF. We considered an extensive list of variables for inclusion in the ToPP-HF that were aligned with previous readmission prediction models,²⁰ such as vital signs, comorbidities, medications, and healthcare utilization, and our tool performed similarly, with a C statistic of 0.69. Because we included patients with HF regardless of their primary admission diagnosis, our results are generalizable to a larger population. Since HF diagnosis codes in the principal position have modest sensitivity, restricting risk prediction to this group may exclude patients for whom HF was a contributing cause of hospitalization but was not listed as the primary reason for admission.^25,26 In addition, HF is one of the most common reasons for rehospitalization, regardless of the index admission diagnosis, so there is a significant opportunity for pharmacists to improve clinical outcomes for patients with HF through TOC services.³¹

When comparing predictors across our variable selection models, we found it interesting that endoscopic intervention was considered an important predictor in 5 models (eTable 4). We chose not to include endoscopic intervention in the ToPP-HF tool as we felt this variable may be potentially difficult to collect in practice and the clinical relevance was unclear. Another interesting finding was that primary admission diagnosis of HF had a negative score in the ToPP-HF tool, which would lower the predicted risk of readmission. It is important to note that the models used to develop the ToPP-HF were not designed to provide causal effect estimates. In the case of HF as a primary admission diagnosis, this factor may not necessarily lower risk of readmission, but rather it could be strongly associated with other factors that do lower risk of readmission. For example, providers who documented HF as the principal admission diagnosis may have focused their attention more on managing and escalating the intensity of care related to HF, whereas if HF was documented in a lower position, the provider may have spent less time optimizing care and providing patient education related to HF. Those with a primary admission diagnosis of HF may have received more frequent or higher-quality TOC services, which could also lower the risk of hospital readmission.

Limitations.

The findings of our study must be interpreted in light of several limitations. First, we were only able to ascertain data that was documented in our health system’s EMR. Therefore, if a patient was hospitalized within the follow-up period at an institution without a shared EMR, we could not account for that patient having had an unplanned readmission. Similarly, we may not have collected the most current LVEF value if a more recent echocardiogram was performed at an outside institution (unless that information had been provided to our health system and documented in the EMR). We also chose not to include some variables that we considered time-consuming to collect, not routinely collected, or likely to have unclear EMR documentation (eg, cognitive or physical functioning, social determinants of health). Although these factors could be important in predicting an unplanned readmission, we felt the time requirements for ascertaining the data would diminish the tool’s utility in clinical practice. There is also the possibility that a patient may have died during the 30-day follow-up period but was not excluded from the source population if the information was not accurately represented in the EMR.

Second, our study was designed to target adult patients with HF who were discharged to the community (eg, home, assisted living facility) or to an institutional postacute care setting (eg, skilled nursing facility). Therefore, the ToPP-HF tool may not generalize to certain groups who were excluded from our study population. Additionally, since the ToPP-HF was developed using data from a single health system, future validation studies should be conducted to examine the tool’s generalizability to other institutions.

Finally, as it includes 13 variables, the ToPP-HF score may not be easy enough to calculate by hand for some busy pharmacists. Although some institutions with significant information technology resources may have the ability to automatically calculate the ToPP-HF score in the EMR, for institutions where EMR integration is not possible we created a downloadable calculator (eAppendix A) and example of ToPP-HF scoring (eAppendixB) to assist with calculations. A Web-based version of the calculator is also being designed to allow for widespread use of the tool. Future research should examine the feasibility and sustainability of using the ToPP-HF tool in clinical practice.

Conclusion

The ToPP-HF tool, designed specifically for easy use by pharmacists, performs well and can be used to predict 30-day all-cause, unplanned readmissions in adult patients with HF. Future research should be conducted to evaluate whether using the ToPP-HF to guide patient selection for pharmacist TOC interventions reduces readmissions for those at high risk.

Disclosures

The research was supported by a grant from the ASHP Foundation (Pharmacy Resident Research Grant). Dr. Zullo is also supported, in part, by grants from the National Institute on Aging (grants R21AG061632, R01AG065722, RF1AG061221, and R01AG062492). The authors have declared no potential conflicts of interest.

Previous affiliations

Dr. Riester was affiliated with Department of Pharmacy, Rhode Island Hospital, Providence, RI, for a portion of the project.

Additional Information

Presented as a poster at the ASHP Virtual Conference for Pharmacy Leaders, October 19, 2020.