Remote monitoring for heart failure: Assessing the risks of readmission and mortality

Abstract

Study objective

Remote monitoring (RM) can help patients with heart failure (HF) remain free of hospitalization. Our objective was to implement a patient-centered RM program that ensured timely clinical response, which would be associated with reduced mortality.

Design

This was a retrospective, observational, propensity-matched study.

Setting

A large regional health system between 9/1/2016–1/31/2018.

Participants

Patients admitted with acute HF exacerbation were matched on key variables. Up to two comparison patients were selected for each RM user.

Interventions

We used an algorithmic approach to assess daily physiologic data, assess symptoms, provide patient education, encourage patient self-management, and triage medical problems.

Main outcome measures

We assessed all-cause mortality using Kaplan-Meier and log rank analysis. We used Cox proportional hazards to compare risk of death.

Results

Our cohort of 680 RM users and 1198 comparisons were similar across baseline characteristics except age (74.7 years versus 76.6 years, p < 0.001, respectively). Having one or more admissions in the preceding 120 days was more prevalent in the RM group (35.9% versus 29.8%, p = 0.013). The 30- and 90-day all-cause readmission rates were each higher among the RM users compared with the comparison patients (p = 0.013 and p < 0.001 for 30 and 90 days, respectively). Mortality was lower in the RM group at 30 and 90 days post-discharge (p < 0.001).

Conclusions

RM that responds to biometric data and encourages patient self-management can be used in a large hospital system and is associated with decreased all-cause mortality. Our findings underscore RM technology as a method to improve HF care.

Highlights

• •Propensity-matching allows study of remote monitoring's effect on mortality.
• •Patients discharged after heart failure admission were selected to use the program.
• •Rates of readmission were higher among the remote monitoring users.
• •Mortality was lower in the remote monitoring group at 30 and 90 days post-discharge.

1. Introduction

Patients admitted to the hospital with acute decompensated heart failure (HF) are at high risk for hospital readmissions and mortality in the post-discharge period [1]. Remote monitoring (RM) and mobile health technologies have been employed as strategies to help community-dwelling patients with HF remain free of hospitalization [2], [3], [4], [5]. RM tools often incorporate physiologic measures like patient weight, blood pressure, and heart rate, which have been shown to be associated with worsening HF [1], [6], [7]. Symptom monitoring is also an important aspect often collected by RM interventions. RM is a useful complement to usual HF care because it allows for early detection of worsened disease while the patient remains outside of the hospital.

Despite these conceptual benefits, RM for HF has shown mixed results in clinical trials [8], [9], [10]. Two previously documented reasons for RM's limited success have involved both patient and provider factors [2], [11]. Regarding patient factors, selection of higher risk patients suggests trends toward significant improvements in clinical endpoints in randomized controlled trials [2], [12]. On the other hand, a focus on low-risk, ambulatory patients has been associated with neutral results [13]. Patients who are admitted in the hospital at the time of study enrollment are considered more unstable and higher risk than patients whose HF is effectively managed in an ambulatory setting. Regarding provider factors, clinician interaction has been critical to the success of studies that have shown a decrease in readmission rates [3], [5], [13]. RM programs that generate alerts to which no actionable response is achieved have generally been less successful at reducing readmissions [2].

Aside from readmission, mortality is another important clinical outcome that could be affected by RM programs. When assessing risk of mortality, complimentary clinical factors must be adequately assessed and adjusted for, such as comorbid diseases. Similarly, patient gender and race have been associated with HF outcomes including mortality, thus accurate reporting of such baseline characteristics should be acknowledged when assessing risk for poor HF outcomes. Notably, predictors of mortality are different from those for readmission [1]. For example, systolic blood pressure predicts mortality whereas weight gain more accurately predicts readmission [1]. To date, attempts to show that RM improves outcomes have been met with disappointing results due to difficulty in selecting the appropriate patient population and measuring the most relevant clinical variables [8], [11], [14]. Propensity-matching methodology can be applied to retrospective data to help overcome logistical barriers to studying a high-risk patient population while adjusting for multiple baseline variables and comorbid conditions.

Therefore, to understand how RM could benefit a cohort of patients hospitalized with HF, we designed a program to be implemented upon discharge after acute HF exacerbation. The objective was to implement a pragmatic RM program that could be easily and widely deployed at our large, multi-hospital, academic medical system. We aimed to ensure patient adherence and timely clinical response. We hypothesized that use of our RM program would be associated with reduced all-cause readmission and mortality.

2. Materials and methods

2.1. Study design

This was a retrospective, observational study wherein the association of RM with cardiac outcomes was determined by propensity matching on key variables using available data within the hospital system's electronic medical record.

2.2. Setting and data collection

The UPMC (University of Pittsburgh Medical Center) Health System is a large not-for-profit academic system located in the Pittsburgh, Pennsylvania region. Over 1 million unique patients are seen annually within UPMC, leading to approximately 5.6 million outpatient encounters and 382,000 hospital admissions per year. [15] We analyzed data from 929 patients who enrolled in the RM program between 9/1/2016 and 1/31/2018 and who were identified as having had inpatient HF admissions within the 14 days prior to enrollment. Patients with a primary admission diagnosis of HF were identified by referral from their providers or their insurance company. To be enrolled, a remote monitoring order was required from either the primary care physician or primary cardiologist. RM program recruiters used the following criteria to screen patients, and if present, determined them to be poor candidates for the program: psychosis, dialysis, physical inability to provide biometric data, lack of permanent home address, end-stage HF (palliative inotropes, heart transplant candidate, ventricular assist device), life expectancy less than six months, weight greater than 400 pounds, or lack of provider to respond to alerts. The UPMC Quality Review Committee approved this observational study as quality improvement and informed consent for research was not required. The data that support the findings of this study are available from the corresponding author upon reasonable request.

2.3. Remote monitoring program

The intervention consisted of a kit-based program including a tablet, external blood pressure monitor, pulse oximeter, and scale. Using the tablet, the patient would respond to a short series of multiple-choice questions and input their home vital sign readings daily Saturday through Sunday. In response to the patient's input, the RM protocol featured an algorithmic, pragmatic approach. If the patient's data fell outside the normal range, an alert would be triggered to the UPMC Call Center. The Call Center is a group of nurse-level patient educators who are trained to encourage patient self-management, assess symptoms, and triage medical problems over the phone. Prior to the current HF RM initiative, the Call Center helped to remotely manage other clinical conditions, and was repurposed for HF as the RM program was developed. If an alert required non-emergent medical attention, the Call Center nurse would notify the ordering physician via the outpatient electronic health record. The ordering physician would then manage the condition at the physician's clinical discretion.

2.4. Outcomes of interest

Patients were enrolled in the program for a 90-day period. The primary outcome was all-cause readmission (including observation visits) at 30 and 90 days following program enrollment to hospitals within our health system. We abstracted electronic medical record data for number of emergency department visits. Secondary outcomes included 90-day mortality and death, which was measured by the social security death index data available in the electronic health record.

2.5. Statistical analyses

We used the following assumptions in our matching approach. The extent of covariate overlap and balance is shown in Table 1. Exact matches were required within diagnosis-related group (DRG) code, readmission risk score category, [16] and the activity outcome measurement for post-acute care (AMPAC) mobility category. [17] Patients were also matched on mortality risk using the All Patient Refined Diagnosis Related Groups risk of mortality. [18] This is a product applied to the medical record that determines one of four possible risk categories (minor, moderate, major, extreme) for each patient based on available clinical data. The closest mathematical matches were acceptable for matched variables (nearest neighbor). [19] Up to two comparison patients were selected for each RM user to decrease the risk of unobserved confounding variables. [19], [20] These outcomes were compared using two-tailed t-tests. We also assessed mortality using Kaplan-Meier survival and log rank analysis. We used Cox proportional hazards to compare risk of death in the RM versus comparison group. Statistical analyses were performed with Stata version 15.0.

Table 1.

Baseline characteristics for remote monitoring (RM) users and propensity-matched comparison patients.

Characteristics Mean (SD) or n (%)	RM group n = 680	Comparison group n = 1198	p-Value
Age	74.7 (11.7)	76.6 (11.2)	<0.001
Elixhauser comorbidity index	6.7 (1.9)	6.5 (1.7)	0.040
Female	354 (52%)	620 (52%)	0.899
Married	330 (49%)	555 (46%)	0.358
Race
White	586 (86%)	1049 (88%)	0.392
Black	87 (13%)	123 (10%)	0.098
Other	7 (1%)	26 (2%)	0.060
Commercial insurance	22 (3%)	25 (2%)	0.132
Admissions in the preceding 120 days			0.013
0	436 (64%)	841 (70%)
1+	244 (36%)	357 (30%)
Risk of mortality [18]			0.957
Minor	16 (2%)	26 (2%)
Moderate	259 (38%)	457 (38%)
Major	334 (49%)	598 (50%)
Extreme	71 (10%)	117 (10%)
Severity of illness			0.955
Minor	22 (3%)	35 (3%)
Moderate	212 (31%)	374 (31%)
Major	377 (55%)	674 (56%)
Extreme	69 (10%)	115 (10%)
AMPAC [17]			0.945
Severe limitation (<11)	6 (1%)	9 (1%)
Moderate limitation (12 to 17)	67 (10%)	111 (9%)
Some limitation (18 to 23)	249 (37%)	451 (38%)
No limitation (24)	349 (52%)	612 (52%)
Length of stay	5.8 (3.9)	5.6 (3.7)	0.400
Discharge disposition
Home	93 (14%)	141 (12%)	0.232
Home with home health	573 (84%)	1001 (84%)	0.688
Skilled Nursing Facility	12 (2%)	43 (4%)	0.019
Other	2 (0.3%)	13 (1%)	0.046
Readmit score [16]	4.5 (1.8)	4.3 (1.8)	0.048
Readmit score risk category			0.390
Low	345 (51%)	634 (53%)
Medium	276 (41%)	480 (40%)
High	58 (9%)	83 (7%)

AMPAC = activity outcome measurement for postacute care.

3. Results

3.1. Baseline characteristics

Using the nearest neighbor method, suitable matches could not be found for 249 RM users. Our cohort therefore consisted of 680 RM users and 1198 comparison patients. Demographic data and baseline characteristics for the analyzed cohorts are shown in Table 1. The baseline characteristics for the entire cohort of RM users are shown in the Supplementary Table. Those excluded from analysis on average younger, more likely to identify as Black or Other, have commercial insurance, and have an extreme risk of death and more mobility limitations. Of those included in the propensity-matched analysis, RM users were on average two years younger (74.7 years versus 76.6 years, p < 0.001). The average Elixhauser morbidity index score was 0.2 higher in the RM group (p = 0.04). BMI was 1.4 units higher among RM users (p < 0.001). Having one or more admissions in the preceding 120 days was slightly more prevalent in the RM group (35.9% versus 29.8%, p = 0.013). Of those enrolled in the RM program during the study period, the median duration of participation was 82 days (mean = 71.0, SD = 54.5). Top reasons for discontinuing the program included no longer meeting eligibility criteria, patient expired, or patient no longer wished to participate. The RM program generated 129,450 alerts over an 18-month period. This was an average of 139 alerts per RM user. Nearly all patients had Medicare insurance with the remaining 3% of RM users and 2% of comparison patients having commercial insurance (p = 0.132). The majority (67%) of patients had a documented ejection fraction listed in the available electronic medical record during index admission. Of the 453 RM users and 797 comparison patients the mean (SD) ejection fractions were 45.3 (15.9) and 42.6 (16.4), respectively (p = 0.005).

3.2. Readmissions

At baseline, the overall a priori all-cause readmission prediction score [16] was 0.2 higher in the RM group. As displayed in Table 2, our analysis showed that 30- and 90-day all-cause readmission rates were each significantly higher among the RM users compared with the comparison patients. Those with the highest risk for readmission as determined by the a priori readmission risk score had 59% higher odds of readmission within 30 days (p = 0.02) and 82% higher odds of readmission within 90 days (p = 0.001). Women consistently had higher odds of readmission at 30- and 90-days (OR 1.54, p = 0.006 and OR 1.72, p = 0.01, respectively). Fig. 1 shows 30- and 90-day all-cause readmission rates for categories of patients in the RM and comparison groups.

Table 2.

All-cause readmissions at 30 and 90 days among the remote monitoring program (RM) users compared with the comparison patients.

	n	RM group n (%)	Comparison group n (%)	p-Value	OR [95% CI]	p-Value
30-Day readmission outcomesa
Overall	1878	163 (24.0%)	229 (19.1%)	0.013	1.33 [1.06,1.67]	0.013
Low/med risk	1388	94 (19.1%)	149 (16.6%)	0.237	1.19 [0.89,1.58]	0.235
High risk	490	69 (36.5%)	80 (26.6%)	0.021	1.59 [1.07,2.35]	0.020
Black	210	17 (19.5%)	18 (14.6%)	0.350	1.42 [0.68,2.94]	0.349
White	1635	144 (24.6%)	210 (20.0%)	0.033	1.30 [1.02,1.66]	0.032
Female	974	94 (26.6%)	118 (19.0%)	0.007	1.54 [1.13,2.10]	0.006
Male	904	69 (21.2%)	111 (19.2%)	0.480	1.13 [0.81,1.58]	0.478
90-Day readmission outcomesa
Overall	1878	309 (45.4%)	419 (35.0%)	<0.001	1.55 [1.28,1.88]	<0.001
Low/med risk	1388	195 (39.7%)	282 (31.4%)	0.002	1.44 [1.14,1.81]	0.002
High risk	490	114 (60.3%)	137 (45.5%)	0.001	1.82 [1.26,2.63]	0.001
Black	210	40 (46.0%)	43 (35.0%)	0.108	1.58 [0.90,2.78]	0.109
White	1635	267 (45.6%)	373 (35.6%)	<0.001	1.52 [1.23,1.86]	<0.001
Female	974	169 (47.7%)	215 (34.7%)	<0.001	1.72 [1.32,2.25]	<0.001
Male	904	140 (42.9%)	204 (35.3%)	0.023	1.38 [1.05,1.82]	0.023

^aInpatient and observation stay combined.

Fig. 1.

Bar graphs with 30- and 90-day all-cause readmission rates for categories of patients in the remote monitoring and comparison groups.

3.3. Mortality

Mortality was significantly lower in the RM group at 30 and 90 days post-discharge. The mortality rates for RM users were significantly lower for the overall cohort as well as each of the race and gender subgroups analyzed (see Table 3). Our survival analysis showed that 90-day survival was higher in the RM group (94.4%) compared with propensity-matched comparisons (89.2%). Cox model hazard ratio ([95% CI] = 0.49 [0.34,0.71], p < 0.001). Fig. 2 shows plotted covariate-adjusted survival curves based on the Cox models.

Table 3.

Mortality outcomes at 30 and 90 days among the remote monitoring program (RM) users compared with the comparison patients.

	n	RM group n (%)	Comparison group n (%)	p value
30-day mortality	1878	6 (0.9%)	57 (4.8%)	<0.001
Low/med readmission risk	1388	3 (0.6%)	37 (4.1%)	<0.001
High readmission risk	490	3 (1.6%)	20 (6.6%)	0.006
Black	210	0 (0.0%)	3 (2.4%)	0.072
White	1635	6 (1.0%)	54 (5.1%)	<0.0001
Female	974	2 (0.6%)	19 (3.1%)	0.004
Male	904	4 (1.2%)	38 (6.6%)	<0.001
90-day mortality	1878	38 (5.6%)	131 (10.9%)	<0.001
Low/med readmission risk	1388	25 (5.1%)	96 (10.7%)	<0.001
High readmission risk	490	13 (6.9%)	35 (11.6%)	0.078
Black	210	1 (1.1%)	9 (7.3%)	0.024
White	1635	37 (6.3%)	121 (11.5%)	<0.001
Female	974	18 (5.1%)	65 (10.5%)	0.003
Male	904	20 (6.1%)	66 (11.4%)	0.007

Fig. 2.

Kaplan-Meier curves for mortality at 90 days among remote monitoring users and patients in comparison groups.

3.4. Sensitivity analysis

Exact matching on age would have decreased the sample size significantly. However, readmission and mortality results were similar with this slight age difference compared with exact matching on age, as determined by our sensitivity analysis.

4. Discussion

In this observational propensity-matched study, we showed that RM was associated with increased readmissions and decreased mortality in a high-risk cohort of patients with HF. Furthermore, though we matched on a priori mortality risk, we found that patients receiving the RM intervention were less likely to die from any cause. Having used the RM tool was associated with improved survival in a group of patients with HF.

This was a pragmatic approach to HF care given the involvement of the patients' primary care physicians and primary cardiologists. Our intervention did not solely rely on HF specialists to manage patients' conditions, but instead facilitated patient self-management and as-needed communication with staff from the Call Center. In repurposing the existing Call Center's workflow to the HF RM program, we were able to focus existing resources on the high-yield care of an at-risk patient population. There was a clear escalation protocol for management of program alerts in which the Call Center nurses provided feedback to patients and involved the ordering physician as needed. The program also emphasized the important elements of patient self-care and education with online content using the provided tablet. Importantly, the hospital-issued tablet ensured that patients were not required to use their own technology, thus removing a potential barrier from participation in telehealth. [21]

We used a previously validated readmission risk score for this analysis. [16] At baseline, the mean a priori all-cause readmission prediction score was 0.2 higher in the RM group and there was no difference in the categorical version of the score between the two groups. Our analysis showed that the 30- and 90-day all-cause readmission rates were significantly higher in the RM group, both overall and among each of the subgroups analyzed. It is therefore doubtful that a negligible baseline difference in risk score would drive the large difference in readmissions found in our results. Notably, our overall all-cause readmission rate was comparable to the recent national average of about 30% at 60 days. [1]

The current findings build on earlier work analyzing RM for HF but differ in important ways. A large systematic review conducted in 2015 found that RM (including telephone support and non-invasive telemonitoring) reduced all-cause mortality but was less effective at reducing readmissions. [22] A 2015 Cochrane review of telemedicine for HF care found no significant mortality benefit at six months. Readmissions ranged from decreased by 64% to increased by 60%. [14] Our work did not confirm the findings of prior work suggesting higher rates of readmission among Black patients. [23], [24]

Regarding mortality, the latest trends suggest that HF deaths are on the rise and are disproportionately higher among Black men and women. [25], [26] For our analysis, the APR-DRG risk of mortality score [18] was added to the matching algorithm to correct for any imbalance. Despite correcting for mortality risk, our analysis showed a reduction in 30-, 60-, and 90-day mortality. In addition, we showed that the RM intervention improved survival at one year. Our findings have important implications for HF care among Black patients. In our study, Black RM users had higher odds of all-cause readmission at each time interval than did White RM users, but these findings did not meet statistical significance. We have shown that RM use among Black patients is associated with a survival benefit. This will be important moving forward as policy makers seek ways to address the growing racial differences in HF outcomes. [27]

Readmissions may present an opportunity to provide additional care needed to keep patients alive. [28] RM enhances the ability for clinicians to make therapeutic decisions about their patients and to consider novel interventions such as same-day infusion centers to administer intravenous diuretics for patients with volume overload. [29], [30] In our current practice environment, we have become accustomed to using HF readmissions as a marker of poor hospital performance. Our findings suggest a need for improved measures of hospital quality. [31]

Policies to reduce HF readmissions have shown mixed results including variable associations with mortality. After the Hospital Readmission Reduction Program (HRRP) was implemented in 2010, there was an observed increase in mortality in an analysis of the Get With the Guidelines Heart Failure registry linked with Medicare data [32] An interrupted time series analysis showed a reduction in the all-cause readmission rate and a concomitant increase in all-cause mortality. [32] Similarly, a propensity weighted analysis of Medicare recipients determined an increase in mortality for patients with comparable case mix after HRRP. [33] A separate analysis of Medicare data found that the trends of increasing HF mortality began prior to the implementation of the HRRP and did not increase relative to the rate of decreased readmissions. [34] Overall, given limitations in retrospective analyses, any observed increase in mortality among HF patients in the post HRRP period cannot be definitively linked to a reduction in readmissions. Nevertheless, our findings support the need for closer monitoring of high-risk HF patients during the post-discharge period.

Some have proposed that all-cause readmissions are a fixed entity reflective of patient aspects that are not easily modifiable, such as socioeconomic factors. [1], [35] During the COVID-19 pandemic, the threshold for HF admissions has likely increased due to patients' desire to stay at home and the shift of inpatient hospital care to focus on COVID-related illness. [36] Healthcare systems have therefore incorporated RM as an important adjunct to HF care. Importantly, the acceptance of telehealth among older and chronically ill patients is on the rise. According to the National Poll on Healthy Aging, older adults are using telehealth technology at unprecedented rates. [37] RM and telehealth technologies may better allow clinicians to keep patients with HF safely in their homes and minimize the need for in-person hospital or clinic visits.

Our analysis had several strengths. Beyond simple observational analyses, we performed robust propensity matching to better determine the effect of RM on the outcome. Our assessment of a regional healthcare system includes patients living in different states and in urban and exurban locations. This work shows the benefit of a system-wide application of an RM tool, an example of how hospital systems can leverage the size of their networks to provide far-reaching patient care.

Our findings should be reviewed in light of study limitations. First, our large hospital system includes hospitals that vary in size and case-mix indices. We did not match based upon hospital of initial treatment and readmission rates might vary in tertiary care hospitals compared with others. Second, in an attempt to include patients who were being discharged to home, we limited the study sample to those who were likely to die within 6 months based upon the available clinical data. As a result, we may have introduced selection bias in our screening process. However, the matched sample did not differ from RM users regarding post-acute care settings. Furthermore, the sample included patients going home with home health, which could provide additional resources to this HF cohort, a factor which would have biased our findings toward the null. Third, this observational analysis was potentially affected by unobserved confounding. Sensitivity analyses aim to assess how strong an unobserved confounder would have to be to change the study conclusions. Sensitivity to unobserved confounders can be reduced by utilizing a large set of observed confounders in the analysis and by including more than one match per patient, as we have done. Fourth, use of data from the medical record such as the social security death index, while convenient, may be imprecise. Lastly, our program was dependent on hospital resources including a nurse-based call center, which may limit generalizability to settings without such resources. However, others have done similar work for HF management and found the programs to be cost-effective. [38], [39]

In conclusion we have shown that a kit-based RM program that responds to biometric data and encourages patient self-care can be used in a large hospital system to monitor high risk patients and is associated with decreased all-cause mortality. Our findings suggest that, irrespective of race or gender subgroups, investments in RM technology may help to improve HF care and save lives. Future research directions should include the association of RM and other telemedicine tools on HF outcomes, the implementation of these strategies, and their cost effectiveness.

The following is the supplementary data related to this article.

Supplementary Table

Baseline characteristics for all remote monitoring (RM) users including those matched and excluded from the analysis.

Disclaimers

None.

Source of funding

This work was supported by the National Institutes of Health [T32 T32HL129964-1A1].

Disclosures

All authors have nothing relevant to disclose.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.