Abstract
Background:
To improve value in the care of patients with acute myocardial infarction (AMI), payment models increasingly hold providers accountable for costs. As such, providers need tools to predict length of stay (LOS) during hospitalization and the likelihood of needing post-acute care facilities after discharge for AMI patients. We developed models to estimate risk for prolonged LOS and post-acute care for AMI patients at time of hospital admission to facilitate coordinated care planning.
Methods and Results:
We identified patients in the NCDR ACTION registry who were discharged alive after hospitalization for AMI between July 1, 2008 and March 31, 2017. Within a 70% random sample (Training cohort) we developed hierarchical, proportional odds models to predict length of stay and hierarchical logistic regression models to predict discharge to a skilled nursing facility. Models were validated in the remaining 30%. Of 633,737 patients in the training cohort, 16.8% had a prolonged LOS (7+ days) and 7.8% were discharged to a post-acute facility (extended care, a transitional care unit, or rehabilitation). Model discrimination was moderate in the validation dataset for predicting LOS (C statistic = 0.640) and strong for predicting discharge to a skilled nursing facility (C statistic = 0.827). For both models, discrimination was similar in STEMI and NSTEMI subgroups and calibration was excellent.
Conclusions:
These models developed in a national registry can be used at the time of initial hospitalization to predict LOS and discharge to post-acute facilities. Prospective testing of these models is needed to establish how they can improve care coordination and lower costs.
Introduction
As alternative payment models (APMs) shift reimbursement from “payment for volume” to “payment for value,” health care systems will be increasingly at risk for a greater proportion of their patients’ cost of care. Most strikingly, the Centers for Medicare and Medicaid Services (CMS) had initially discussed mandatory episode payment models (EPMs) for acute myocardial infarction (AMI), commonly referred to as “bundled payments”.1 In these models, hospitals would be paid a fixed price for a hospital stay for an AMI and the related care 90 days following discharge, adjusted for performance on specific quality metrics.2 Importantly, these costs include those incurred with other providers, such as post-acute care. Although mandatory EPMs were later cancelled, CMS is starting voluntary EPMs on October 1, 2018, called Bundled Payments for Care Improvement Advanced (BPCI Advanced) which will include AMI.3
Given this imminent start of BPCI Advanced, there is an urgent need to develop tools to support implementation of EPMs for health care systems. These EPMs represent a dramatic shift in the way that hospitals receive payment for AMI care as compared with traditional fee-for-service payment. Conceptually, EPMs represent an effort to incentivize value in cardiology,4 with the hope that they will improve the quality of care during the index hospitalization and encourage better care transitions and post-discharge monitoring, thus reducing the total cost of care.5 For these models to improve value, hospitals and clinicians need to be able to quickly identify patients who are at higher (or lower) risk of prolonged hospitalization and referral to post-acute services following discharge.6 Identifying high-risk patients early during hospitalization can theoretically support a ‘fast-track’ discharge program that avoids critical care admissions and accelerates education and health transitions for low-risk patients, while enabling better communication with post-acute care facilities for patients at high risk for prolonged LOS and needing skilled nursing care after discharge. To address these evolving changes in healthcare and to provide tools to help providers succeed in AMI alternative payment models, we sought to develop predictive models that estimate a patient’s likelihood of prolonged hospitalization and need for post-acute services from data available at the beginning of the index hospitalization.
Methods
Per policy of the National Cardiovascular Data Registry, the data, analytic methods, and study materials will not be made available to other researchers for purposes of reproducing the results of this analysis. Nevertheless, if requested, the authors will provide samples of the analysis to assist other, future research efforts. Furthermore, requests for additional analysis with these data can be made through https://cvquality.acc.org/NCDR-Home/research/submit-a-proposal.
Study Population
The ACTION registry is a US multicenter registry of pateints hospitalized with either ST-segment elevation MI (STEMI) and non-ST segment elevation MI (NSTEMI), regardless of treatment strategy.7 Administered by the American College of Cardiology as part of the National Cardiovascular Data Registry (NCDR), ACTION has a rigorous data quality program consisting of (1) a data quality report, (2) a set of internal quality metrics, and (3) a yearly audit program designed to ensure completeness, consistency, and accuracy of data.8 Participating hospitals receive quarterly quality reports. This suite of offerings can support quality assessment through benchmarking on hospital’s performance with that of other hospitals throughout the US. Specific data included in the registry are listed at https://www.ncdr.com/webncdr/action/home/datacollection. Participating hospitals include more than a quarter of all hospitals caring for patients with AMI in the United States.9
Patients in ACTION who were discharged alive from July 1, 2008 and March 31, 2017 were eligible for analysis. We excluded patients discharged to hospice care, those transferred to a different acute care facility, and patients who were discharged from the hospital against medical advice. We also excluded patients with data collected on the limited data collection form in ACTION, given lack of needed data elements. Numbers of excluded and included patients are shown in Figure 1.
Figure 1.
Included and Excluded Patients
Analyses from ACTION were approved by Chesapeake Research Review, Inc., an independent institutional review board, and conducted by the NCDR analytic center at Saint Luke’s Mid America Heart Institute.
Outcomes and Covariates
We aimed to create predictive models for 2 clinical outcomes: length of stay and discharge to a post-acute skilled nursing facility. Length of stay was defined as an ordinal variable by subtracting discharge date from admission date and rounding to the nearest whole number. We defined a short length of stay as 0–2 days, an intermediate length of stay as 3–6 days, and a prolonged length of stay as 7 days or greater. We chose these categories simply to describe the distribution of characteristics of AMI hospitalizations in different LOS ranges, although the ordinal model allows for predicting hospitalization within any specific LOS threshold. Discharge to post-acute facility was defined as any discharge to extended care, a transitional care unit, or rehabilitation.
Initially, we examined all available variables and their distributions related to medical history, presentation, clinical status at admission, and demographics. Candidate variables for the prediction models were selected based on clinical relevance10 and had to be available to clinicians at admission, as the goal was to develop models that could be used at time of initial hospitalization. Given this clinical purpose, we did not consider predictor variables not known to clinicians upon admission. Considered variables included age, sex, body mass index, self-transport (yes/no), race (white, black, Asian, American Indian or Alaskan Native, Native Hawaiian or Pacific Islander), insurance status (Medicare, Medicaid, no insurance, or private insurance), heart failure at first medical contact (yes/no), cardiogenic shock at first medical contact (yes/no), systolic blood pressure at first medical contact (mm Hg), pulse at first medical contact (beats/minute), ST-segment elevation MI (vs. non-ST segment elevation MI), smoking (yes/no), hypertension (yes/no), dyslipidemia (yes/no), end-stage renal failure requiring dialysis (yes/no), diabetes (yes/no), GFR (mL/min), hemoglobin (g/dL), prior heart failure (yes/no), prior myocardial infarction (yes/no), prior percutaneous coronary intervention (yes/no), prior coronary artery bypass surgery (yes/no), prior atrial fibrillation (yes/no), prior cerebrovascular disease (yes/no), and prior peripheral arterial disease (yes/no).
Statistical Analysis
Demographic and clinical characterstics were compared among patients in the 3 LOS categories using one-way analysis of variance for continuous variables and chi-square tests for categorical variables. Demographic and clinical characterstics were compared between those who did versus did have require post-acute care using independent t-tests for continuous variables and chi-square tests for categorical variables. For the LOS model (with a continuous outcome), we used a cumulative proportional logit model, which predicts the likelihood of a LOS for each distinct number of days. This approach is appropriate for continuous data11 and clinically useful because it allows modeling of short vs. long LOS as well as produces a predicted mean LOS for each patient. Nevertheless, we explored the implications of this statistical approach in a sensitivity analysis described below. For the post-acute model (with a binary outcome), we used logistic regression. We used split-sample methods for training and internal validation, with the models derived in a 70% random sample of the data and then internally validated in the remaining 30% of data. Although the split proportion is arbitrary, other NCDR risk models have used 70/3012,13 as well as 80/20 splits.14 Since our intent was to create parsimonious models useful to clinicians for bundled payments, we used backward selection procedures according to the method of Harrell.15 In this approach, the total adjusted variability (R-squared) of an initial model considering all candidate variables is first estimated. Then, sequential backwards elimination is performed, with estimation of total model adjusted R-squared at each step. When the adjusted R squared falls below 90% of the R-squared of the initial model, the selection procedure is terminated and the remaining variables are retained in the final model. Both models were hierarchical, with hospital-level random effects to account for clustering of patients within specific hospitals and to minimize the influence of local practice variation on patient-level predictors of the 2 outcome variables. Cubic splines were considered for all continuous variables. Missing data were minimal for all variables (<0.5%) and was addressed by imputation with sequential regression using the IVEWare software (Ann Arbor, MI). No patients were excluded because of missing data.
To facilitate the utility of the risk models for clinicians at the bedside as well as for automated decision support tools embedded within electronic health records, we created simplified risk scores for both prediction models.16 We then validated the 2 reduced models in the Validation cohort. Model discrimination was assessed via the c-statistic,17 and calibration was assessed by plotting observed vs. predicted rates of the 2 outcomes in the Validation cohort. Model performance was also tested separately among patients admitted with STEMI and NSTEMI. In both models, collinearity was assessed with variance inflation factors. To test potential influence of our variable selection methodology and candidate variables on our results, we conduced sensitivity analyses using a forward selection procedure, and forcing history of cancer and history of chronic lung disease into the models. To assess the effect of secular trends, we included year as a predictive variable in both models as an additional sensitivity analysis. To assess the effect of a cumulative proportional logit model (ie, treating the LOS data as ordinal data), we also conducted a sensitivity analysis with a negative binomial distribution (treating the LOS data as count data). Since results of all these sensitivity analyses were substantively unchanged, they are presented in the Supplementary Appendix. All statistical analyses were performed using SAS Version 9.2.
Results
Study Population
Among 1,247,251 patients who were admitted with a STEMI or NSTEMI to an ACTION participating hospital between July 1, 2008 and March 31, 2017, we excluded 237,987 (19.1%) because the hospitals used the short form, 43,172 (3.5%) because they died during hospitalization, 45,410 (3.6%) because they were transferred to another acute care facility, 6772 (0.5%) because they were discharged against medical advice, and 7586 (0.6%) because they were discharged to hospice. As such, our final analytic cohort included 906,324 patients (Figure 1), of which 633,737 patients (70.0%) were in the training cohort and 272,587 patients (30.0%) were in the validation cohort.
Of the 633,737 patients in the training cohort, 260,980 patients (41.2%) had a short LOS and 106,245 patients (16.8%) had prolonged LOS according to the definitions here. Relative to patients with short LOS, patients with prolonged LOS were older (67.9 vs. 61.5 years old, p < 0.001), more often female (37.5% vs. 30.2%, p < 0.001), and insured by Medicare (66.6% vs. 46.1%, p < 0.001). Relative to patients with short LOS, patients with prolonged LOS were also more likely to have congestive heart failure on initial presentation (25.5% vs. 4.8%, p < 0.001) and lower systolic blood pressure (142.9 mm Hg vs. 150.8 mm Hg, p < 0.001), as well as a history of atrial fibrillation (11.7% vs. 5.0%, p < 0.001), peripheral arterial disease (13.1% vs. 6.2%, p < 0.001), end-stage renal failure requiring dialysis (4.4% vs. 1.2%, p < 0.001), and cancer (12.0% vs. 8.7%, p < 0.001). Clinical characteristics of patients, divided by LOS category, appears in Table 1.
Table 1.
Clinical characteristics, divided by length of stay categories
| Total | Length of stay | P-Value | |||
|---|---|---|---|---|---|
| n = 633737 | 1 LOS 0–2 Days | 2 LOS 3–6 Days | 3 LOS 7+ Days | ||
| n = 260980 | n = 265377 | n = 106245 | |||
| Age | 64.2 ± 13.6 | 61.5 ± 13.0 | 65.4 ± 14.0 | 67.9 ± 12.7 | < 0.001 |
| Sex | < 0.001 | ||||
| Male | 417421 (65.9%) | 182227 (69.8%) | 167983 (63.3%) | 66443 (62.5%) | |
| Female | 216316 (34.1%) | 78753 (30.2%) | 97394 (36.7%) | 39802 (37.5%) | |
| BMI | 29.8 ± 11.6 | 30.1 ± 13.5 | 29.7 ± 9.9 | 29.7 ± 10.3 | < 0.001 |
| Missing | 3096 | 1194 | 1407 | 484 | |
| Race | |||||
| White | 539964 (85.2%) | 226827 (86.9%) | 223379 (84.2%) | 88800 (83.6%) | < 0.001 |
| Black or African American | 71370 (11.3%) | 25476 (9.8%) | 32646 (12.3%) | 13104 (12.3%) | < 0.001 |
| Missing | 6 | 1 | 1 | 4 | |
| Asian | 11716 (1.8%) | 4465 (1.7%) | 4916 (1.9%) | 2317 (2.2%) | < 0.001 |
| Missing | 6 | 1 | 1 | 4 | |
| American Indian or Alaskan Native | 4589 (0.7%) | 1880 (0.7%) | 1884 (0.7%) | 819 (0.8%) | 0.133 |
| Missing | 6 | 1 | 1 | 4 | |
| Native Hawaiian or Pacific Islander | 998 (0.2%) | 356 (0.1%) | 416 (0.2%) | 224 (0.2%) | < 0.001 |
| Missing | 6 | 1 | 1 | 4 | |
| Hispanic or Latino Ethnicity | 35758 (5.7%) | 13509 (5.2%) | 15354 (5.8%) | 6856 (6.5%) | < 0.001 |
| Missing | 2664 | 1093 | 1088 | 476 | |
| Means of Transport to First Facility | |||||
| Self/Family | 362283 (57.2%) | 164858 (63.2%) | 143082 (54.0%) | 53563 (50.5%) | < 0.001 |
| Ambulance | 263721 (41.7%) | 93549 (35.9%) | 118637 (44.7%) | 51190 (48.2%) | |
| Mobile ICU | 2375 (0.4%) | 598 (0.2%) | 1199 (0.5%) | 576 (0.5%) | |
| Air | 4710 (0.7%) | 1700 (0.7%) | 2199 (0.8%) | 808 (0.8%) | |
| Missing | 648 | 275 | 260 | 108 | |
| Clinical Presentation | < 0.001 | ||||
| STEMI or STEMI Equivalent Heart Failure at First Medical Contact Missing Heart Rate at First Medical Contact Missing Systolic Blood Pressure at First Medical Contact Missing |
242931 (38.3%) 75434 (11.9%) 568 84.0 ± 23.0 1671 148.0 ± 33.1 2000 |
103162 (39.5%) 12549 (4.8%) 220 80.9 ± 20.2 675 150.8 ± 30.3 809 |
109383 (41.2%) 35707 (13.5%) 248 84.8 ± 23.3 693 147.3 ± 33.2 829 |
30235 (28.5%) 27120 (25.5%) 97 89.5 ± 27.1 296 142.9 ± 38.3 354 |
|
| Cardiac Arrest at First Medical Contact Missing Cardiogenic Shock at First Medical Contact Missing |
14223 (2.7%) 105808 15724 (2.5%) 666 |
2132 (0.9%) 35391 1908 (0.7%) 233 |
5977 (2.8%) 50419 6236 (2.4%) 306 |
6110 (7.1%) 19975 7574 (7.1%) 125 |
|
| Medical History | < 0.001 | ||||
| Current/Recent Smoker (w/in 1 year) Missing Hypertension Missing Dyslipidemia Currently on Dialysis Missing Chronic Lung Disease Missing Diabetes Mellitus Missing Atrial Fibrillation or Flutter Missing Cerebrovascular Disease Missing Peripheral Arterial Disease Missing |
217435 (34.3%) 173 468440 (73.9%) 86 388396 (61.3%) 14939 (2.4%) 299 58299 (13.9%) 214232 210617 (33.2%) 220 49040 (7.7%) 573 73867 (11.7%) 267 55370 (8.7%) 332 |
100360 (38.5%) 59 182104 (69.8%) 37 155842 (59.7%) 3163 (1.2%) 106 17009 (10.4%) 97304 71883 (27.6%) 85 12977 (5.0%) 225 21547 (8.3%) 93 16186 (6.2%) 128 |
86549 (32.6%) 73 199582 (75.2%) 32 163406 (61.6%) 7083 (2.7%) 123 26993 (14.7%) 81874 91469 (34.5%) 96 23543 (8.9%) 249 34299 (12.9%) 115 25194 (9.5%) 141 |
30181 (28.4%) 39 85870 (80.8%) 17 68369 (64.4%) 4681 (4.4%) 69 14276 (19.8%) 34120 46892 (44.2%) 38 12421 (11.7%) 93 17877 (16.8%) 58 13914 (13.1%) 63 |
|
| Laboratory Values | < 0.001 | ||||
| Initial Creatinine Value Missing Initial Hemoglobin Value Missing |
1.3 ± 1.2 3167 13.9 ± 2.1 3401 |
1.1 ± 0.9 1674 14.3 ± 1.9 1884 |
1.3 ± 1.2 1204 13.7 ± 2.2 1239 |
1.6 ± 1.6 288 13.2 ± 2.4 277 |
|
Of the 633,737 patients in the training cohort, 49941 patients (7.8%) were discharged to post-acute facilities. Relative to patients not discharged to post-acute facilities, patients discharged to post-acute care were older (76.8 years vs. 63.1 years, p < 0.001), more often female (51.4% vs. 32.7%, p < 0.001), and more often insured by Medicare (84.7% vs. 51.9%, p < 0.001). They were also more likely to have congestive heart failure on initial presentation (30.7% vs. 10.3%, p < 0.001) and lower systolic blood pressure (139.7 vs. 148.7, p < 0.001), as well as a history of atrial fibrillation (17.6% vs. 6.9%, p < 0.001), peripheral arterial disease (15.8% vs. 8.1%, p < 0.001), end-stage renal failure requiring dialysis (5.1% vs. 2.1%, p < 0.001), and cancer (16.0% vs. 9.6%, p < 0.001). Clinical characteristics of patients, divided by post-acute utilization, appears in Table 2.
Table 2.
Clinical characteristics, divided by post-acute utilization
| Total | SNF | P-Value | ||
|---|---|---|---|---|
| n = 633737 | 1 n = 49941 | 0 n = 583796 | ||
| Age | 64.2 ± 13.6 | 76.8 ± 11.6 | 63.1 ± 13.2 | < 0.001 |
| Sex | < 0.001 | |||
| Male | 417421 (65.9%) | 24261 (48.6%) | 393160 (67.3%) | |
| Female | 216316 (34.1%) | 25680 (51.4%) | 190636 (32.7%) | |
| Body Mass Index | 29.8 ± 11.6 | 28.6 ± 11.6 | 30.0 ± 11.6 | < 0.001 |
| Missing | 3096 | 466 | 2630 | |
| Race | ||||
| White | 539964 (85.2%) | 43194 (86.5%) | 496770 (85.1%) | < 0.001 |
| Black or African American | 71370 (11.3%) | 5328 (10.7%) | 66042 (11.3%) | < 0.001 |
| Missing | 6 | 1 | 5 | |
| Asian | 11716 (1.8%) | 753 (1.5%) | 10963 (1.9%) | < 0.001 |
| Missing | 6 | 1 | 5 | |
| American Indian or Alaskan Native | 4589 (0.7%) | 235 (0.5%) | 4354 (0.7%) | < 0.001 |
| Missing | 6 | 1 | 5 | |
| Native Hawaiian or Pacific Islander | 998 (0.2%) | 66 (0.1%) | 932 (0.2%) | 0.137 |
| Missing | 6 | 1 | 5 | |
| Hispanic or Latino Ethnicity | 35758 (5.7%) | 2129 (4.3%) | 33629 (5.8%) | < 0.001 |
| Missing | 2664 | 240 | 2424 | |
| Means of Transport to First Facility | < 0.001 | |||
| Self/Family Ambulance Mobile ICU Air Missing |
362283 (57.2%) 263721 (41.7%) 2375 (0.4%) 4710 (0.7%) 648 |
15857 (31.8%) 33349 (66.9%) 413 (0.8%) 262 (0.5%) 60 |
346426 (59.4%) 230372 (39.5%) 1962 (0.3%) 4448 (0.8%) 588 |
|
| Clinical Presentation | < 0.001 | |||
| STEMI or STEMI Equivalent Heart Failure at First Medical Contact Missing Heart Rate at First Medical Contact Missing Systolic Blood Pressure at First Medical Contact Missing Cardiac Arrest at First Medical Contact Missing Cardiogenic Shock at First Medical Contact Missing |
242931 (38.3%) 75434 (11.9%) 568 84.0 ± 23.0 1671 148.0 ± 33.1 2000 14223 (2.7%) 105808 15724 (2.5%) 666 |
242931 (38.3%) 15317 (30.7%) 53 89.9 ± 26.7 169 139.7 ± 37.3 185 2473 (6.0%) 8839 3217 (6.4%) 60 |
230091 (39.4%) 60117 (10.3%) 515 83.5 ± 22.6 1502 148.7 ± 32.6 1815 11750 (2.4%) 96969 12507 (2.1%) 606 |
|
| Medical History | < 0.001 | |||
| Current/Recent Smoker (w/in 1 year) Missing Hypertension Missing Dyslipidemia Currently on Dialysis Missing Chronic Lung Disease Missing Diabetes Mellitus Missing Atrial Fibrillation or Flutter Missing Cerebrovascular Disease Missing Peripheral Arterial Disease Missing |
217435 (34.3%) 173 468440 (73.9%) 86 388396 (61.3%) 14939 (2.4%) 299 58299 (13.9%) 214232 210617 (33.2%) 220 49040 (7.7%) 573 73867 (11.7%) 267 55370 (8.7%) 332 |
7805 (15.6%) 20 43017 (86.1%) 5 32656 (65.4%) 2550 (5.1%) 31 7788 (23.2%) 16355 22483 (45.0%) 17 8759 (17.6%) 46 12988 (26.0%) 19 7891 (15.8%) 35 |
209630 (35.9%) 153 425423 (72.9%) 81 355740 (60.9%) 12389 (2.1%) 268 50511 (13.1%) 197877 188134 (32.2%) 203 40281 (6.9%) 527 60879 (10.4%) 248 47479 (8.1%) 297 |
|
| Laboratory Values | < 0.001 | |||
| Initial Creatinine Value Missing Initial Hemoglobin Value Missing |
1.3 ± 1.2 3167 13.9 ± 2.1 3401 |
1.6 ± 1.5 174 12.5 ± 2.2 171 |
1.2 ± 1.1 2993 14.0 ± 2.1 3230 |
|
LOS Model Development and Validation
After multivariable adjustment and stepwise elimination, 9 variables were retained in the ordinal LOS model. Factors independently associated with a longer LOS included older age (OR 1.15 per 10 years, 95% CI 1.14–1.15), heart failure on admission (OR 2.29, 95% CI 2.26–2.33), higher heart rate on admission (OR 1.10 per 10 bpm, 95% CI 1.10–1.10), systolic blood pressure on admission (nonlinear relationship [U-shaped]: SBP <150 mmHg: OR 0.92 per 1 mm Hg, 95% CI 0.91–0.92; SBP >150 mmHg: OR 1.01 per 1 mm Hg, 95% CI 1.01–1.01), shock on admission (OR 3.67, 95% CI 3.57–3.78), diabetes mellitus (OR 1.28, 95% CI 1.27–1.30), lower glomerular filtration rate (OR 1.02 per 5 mL/min/1.73 m2 lower, 95% CI 1.02–1.02), and lower hemoglobin (OR 1.10 per 1 g/dL lower, 95% CI 1.09–1.10; Figure 2A). Model discrimination was moderate in the validation dataset (C-statistic = 0.640 [0.638–0.641]) and was similar in both STEMI (C-statistic = 0.636 [0.632–0.640]) and NSTEMI (C-statistic = 0.640 [0.637–0.642]) patients. Variation inflation factors did not suggest high collinearity (full results are shown in the Supplementary Appendix). Calibration of expected with observed data was excellent (see “Calibration of the LOS Model” in the Supplementary Appendix).
Figure 2.
A, Model for Prolonged Length of Stay. B, Model for Post-Acute Utilization
The simplified risk model for prolonged LOS appears in Figure 4A. Patients in the low risk cohort (points fewer than 10) had an observed mean LOS of 3.4. Patients in the intermediate risk cohort (points between 11 and 16) had an observed mean LOS of 4.5. Patients in the high risk cohort (points 17 and higher) had an observed mean LOS of 6.4.
Figure 4.
A, Simplified Risk Score for Predicting Length of Stay. B, Simplified Risk Score for Predicting Post-Acute Care
Post-Acute Care Model Development and Validation
After multivariable adjustment and stepwise elimination, 9 variables were retained in the post-acute care model. Factors independently associated with a greater odds of discharge to post-acute facility included older age (OR 2.01 per 10 years, 95% CI 1.99–2.03), heart failure on admission (OR 1.79, 95% CI 1.75–1.84), higher heart rate on admission (OR 1.08 per 10 beats per minute, 95% CI 1.08–1.09), shock on admission (OR 3.84, 95% CI 3.67–4.02), prior cerebrovascular disease (OR 1.71, 95% CI 1.67–1.75), and lower hemoglobin (OR 1.16 per 1 g/dL lower, 95% CI 1.16–1.16). Relative to private insurance, patients with no insurance were less likely to go to post-acute care (OR 0.51, 95% CI 0.47–0.55) while patients with Medicare and Medicaid insurance had a greater odds of discharge to post-acute care (OR 1.35, 95% CI 1.32–1.38; OR 1.41, 95% CI 1.33–1.50, respectively; Figure 2B). Model discrimination was strong in the validation dataset (C-statistic = 0.827 [0.826–0.829]) and was similarly strong in both STEMI (C-statistic = 0.844 [0.840–0.847]) and NSTEMI (C-statistic = 0.809 [0.807–0.811]) patients. Variation inflation factors did not suggest high collinearity (full results are shown in the Supplementary Appendix). Calibration of expected with observed data was excellent (Figure 3).
Figure 3.
A, Observed and Expected Length of Stay in the Validation Cohort. B, Observed and Expected Post-Acute Utilization in the Validation Cohort
The simplified risk model for post-acute care appears in Figure 4B. Patients in the low risk cohort (points between −3 and 3) had an observed rate of post-acute care of 1.6% Patients in the intermediate risk cohort (points between 4 and 9) had an observed rate of post-acute care of 11.2% Patients in the high risk cohort (points 10 and higher) had an observed rate of post-acute care of 27.4%.
Discussion
As healthcare reimbursement increasingly shifts from volume- to value-based, hospitals will be at increasing financial risk to manage patients more efficiently. In this work, we sought to build tools for clinicians and health systems to predict risks of patients for prolonged LOS and need for post-acute care. In doing so, we have demonstrated that with clinical characteristics known at the time of initial hospitalization for AMI, both LOS and post-acute utilization can be predicted with moderate and strong predictive accuracy, respectively. With implementation of episode-based payment in October 2018, which is intended to incentivize reduction in costs, these risk models will be essential for identifying higher- (or lower-) risk patients for whom interventions may be implemented. Furthermore, by facilitating the ability of clinicans to respond to changing payment models in ways that improve targeting resources to patients at higher risk of incurring unnecessary costs, these models can improve the likelihood that payment reforms will improve value in AMI care.
For a model to be actionable, the patient risk must be known as soon as possible, so care pathways can be implemented to streamline care and post-care needs. We thought there would be great clinical utility in knowing, on the day of admission, if a patient were likely to be able to be discharged within a short time frame. For example, knowing at the time of admission that a patient is likely to have a short LOS, it might be possible to redesign care pathways and hospital spaces that avoid intensive care unit stays for these patients and prevent the time, workload and delays of inter-unit transfers within a hospital. Furthermore, patients with expected short LOS could have early coordination with family and potentially early referral to cardiac rehabilitation. Conversely, knowing the projected LOS for a sicker patient and the likelihood that they would need post-acute care after discharge could enable earlier discussions with specific skilled nursing facilities to enable a smoother transition in care. Although we established thresholds to define short and prolonged LOS to describe our data, our model considers LOS as an ordinal variable, so this model allows any thresholds to be defined by health delivery systems and specific clinicians.
Previous work has identified predictors of both post-acute care and prolonged LOS in other procedures and disease conditions. Following joint replacement, a model including age, gender, race, disability, dual eligibility for Medicare and Medicaid, length of stay, intensive care unit use, number of acute hospitalizations in the past year, catheter-associated urinary tract infection, and venous thromboembolism predicted post-acute care with moderate discrimination (C statistic = 0.66).18 Importantly, this model included a number of factors present during the hospitalization (e.g., LOS, intensive care, complications) that render its ability to impact care much lower. In a general medical population, a model including 5 functional status variables (patient’s partner inability to provide home help, inability to self-manage drug regimen, number of active medical problems on admission, dependency in bathing, and in transfers from bed to chair) predicted discharge to post-acute care with strong predictive ability (C statistic = 0.82).19 Unfortunately, we were unable to include some functional status or frailty measures in our models, which could have improved model performance even further. Frailty is known to be associated with poor prognosis in patients with AMI20, so markers of frailty such as serum albumin may have been useful in risk prediction. Hemoglobin was considered, retained in the selection procedures, and included in both final models. Furthermore, integrating these clinical registry data with administrative data, such as the Johns Hopkins Ambulatory Care Group (ACG) measure21 or the inpatient Charlson-Deyo22 or Elixhauser measures23 may further improve model discrimination. Furthermore, since prior studies suggest that prediction of post-acute care may depend on different variables for different disease conditions and procedures. For example, variables found to be influential in our models such as shock and heart failure on presentation are unlikely to be influential in predicting post-acute care following elective procedures such as joint replacement. Our post-acute model combines predictor variables that may be specific to AMI (heart failure, heart rate, and shock at first contact) with predictor variables associated with post-acute care in other settings (age, hemoglobin, and insurance status). We are reassured that the collinearity of these AMI-specific and non-AMI specific predictors is low, strengthening our confidence in the estimates identified in the model.
Among these 2 models, discrimination is stronger for predicting post-acute care than LOS. While some clinical outcomes, such as mortality, are straightforward to predict, others are not, including LOS and readmission. In this specific case, potential variability in patients’ LOS is likely influenced by many non-patient-related factors, including poor communication among providers or between providers and caregivers on when a patient is ready for discharge,inter-physician variability in the assessment of when a patient is ‘safe’ to be discharged, pretreatment with thienopyridines in NSTEMI patients who are then found to require CABG, poor anticoagulation management or the use of excessive contrast such that patients develop acute renal injury all may diminish the predictive accuracy of the model. Thus while the model’s discrimination is comparable to other routinely used models such as readmission models, we believe that its lower discrimination underscores the opportunity to reduce variability and inefficiency in AMI care and to improve the value of treatment. Moreover, we hope that widespread use of this LOS model could enable the development of novel protocols that could decrease unwarranted variations in care, ultimately leading to better performing LOS models over time. We did not include year as a candidate variable, given that the intent was to develop models useful to clinicians, but since the incidence and prognosis of AMI has been changing over time24, the models’ discrimination should be reassessed in the future. We are reassured that including year as a predictive variable in sensitivity analysis did not change the discrimination of the models.
We believe that these models may substantially improve the ability of health care providers and health delivery organizations to improve performance in EPMs and other types of payment mechanisms that incentivize judicious use of resources and costs, although further research into how best to use these models is needed. Developing risk-based protocols, implementing the tools and evaluating their impact on care, outcomes and costs is an important direction for future research. In addition to BPCI Advanced, many similar incentives exist for providers to reduce unnecessary health care expenditures. Cardiologists increasingly are delivering care within accountable care organizations, both within Medicare and with private insurers, which offer shared savings to providers if costs do not exceed specific benchmarks.25
While we believe these models will be most useful to hospitals seeking to improve the efficiency of their care, we also recognize that these models could be used to risk-adjust for patient-complexity. On one hand, risk-adjustment in pay-for-performance models may limit provider avoidance of higher-risk patients, a potential adverse consequence of pay-for-performance.26 However, risk-adjustment can also actually create unintended incentives and threaten the validity of quality metrics. For example, although the Hospital Readmissions Reduction Program was associated with lower risk-standardized and absolute readmission rates27–29, much of the improvement in risk-adjusted readmission rates was related to increased reporting of comorbidities.30 Perhaps because of those concerns, BPCI Advanced does not currently use risk-adjustment for AMI EPMs.3 In the specific case of our models presented here, the outcome predicted is a blend of a clinical outcome and a clinical decision. For example, LOS and utilization of post-acute care after discharge are influenced both by clinical circumstances and provider decisions. Conceptually, therefore, using these metrics for public reporting or payment adjustment, without valid risk adjustment, could create misplaced incentives to avoid post-acute care when patients need it or reducing length of stay inappropriately. Because the detailed clinical data used in these models are not available in administrative databases, we believe our models are best used to design clinical interventions, based on patients’ risks, to optimize quality and performance in EPMs rather than report or adjust payments based on the outcomes of LOS or post-acute utilization themselves.
Our results here should be interpreted in the setting of important limitations. First, as an analysis of observational data in a clinical registry, there is always the possibility of confounding with unmeasured variables associated with both predictors and the outcome variables. Second, the elements we selected were constrained to those collected in ACTION and other important predictors of these outcomes (e.g. frailty) will need to be tested in future research studies. Markers of fraility, including serum albumin and serum hemoglobin levels, might improve the discrimination of these models. Third, since these data are derived from a voluntary registry, the external validity of our findings to all hospitals is uncertain. In particular, hospitals more interested in quality improvement may be more likely to participate in the registry. We are reassured, however, by our large sample size of patients with AMI throughout the United States. Finally, our model cannot account for events that occur during the course of AMI hospitalization. Adding such information might improve the discrimination of the models. We did not consider clinical events during hospitalization for the models, however, since we viewed the clinical utility of the models as guiding interventions to specific patients at time of initial hospitalization.
Conclusions
We demonstrate here that for AMI patients with variables known at the time of initial admission, prolonged LOS can be predicted with moderate discrimination and that need for post-acute care can be predicted with strong discrimination. These models can be used at the bedside by providers both to improve the quality of care and improve performance in APMs such as bundled payments, although further development of risk-based protocols and the impact of implementing these models require more study.
Supplementary Material
What is known
Voluntary episode payment models (“bundled” payments) within Medicare are going to begin soon.
Cardiology episodes of care including acute myocardial infarction are prominent within the new payment models.
For these models to improve value, clinicians need to be able to quickly identify patients who are likely to require more intense utilization, including prolonged length of stay and post-acute care after discharge.
What the study adds
The models presented here can prospectively predict both length of stay and post-acute utilization for acute myocardial infarction patients at the time of initial hospitalization.
Integrating these models into clinical practice might lead to clinical innovations that improve quality and value.
Acknowledgements:
We are grateful to Ian O’Leary who assisted with preparation of the manuscript.
Funding sources: Dr. Wasfy is supported by a career development award from the National Institutes of Health through Harvard Catalyst (KL2 TR001100). This work was primarily funded through the National Cardiovascular Data Registry of the American College of Cardiology.
Footnotes
Disclosures: Jason H. Wasfy: Research grant, National Institutes of Health/Harvard Catalyst, Amount: >$10,000. Frederick A. Masoudi: Contract with the American College of Cardiology through the University of Colorado for role as Chief Science Officer of the NCDR (significant). P. Michael Ho: Consultant (Clinicail Trial Steering Committee), Janssen, Inc: <$10,000; Deputy Editor, Circulation: Cardiovascular Quality and Outcomes:>10,000; Research grant, Department of Veterans Health Administration:>10,000; Research grant, School of Medicine, University of Colorado Anschutz Medical Campus:>10,000. John Spertus: Research grant, American College of Cardiology Foundation, Amount: >$10,000; Equity, Health Outcomes Sciences, Amount: <$10,000
References:
- 1.Episode Payment Models: General Information. 2017;Available from: https://innovation.cms.gov/initiatives/epm. Access date: 16 Aug, 2018.
- 2.Notice of proposed rulemaking for bundled payment models for high-quality, coordinated cardiac and hip fracture care. 2017;Available from: https://www.cms.gov/Newsroom/MediaReleaseDatabase/Fact-sheets/2016-Fact-sheets-items/2016-07-25.html. Access date: 16 Aug, 2018.
- 3.BPCI Advanced. 2018; Available from https://innovation.cms.gov/initiatives/bpci-advanced. Access date: 17 Aug, 2018
- 4.Gilstrap L, Joynt KE. MACRA and Cardiology-The Devil Is in the Details. JAMA Cardiol. 2017; 2:1177–1178. [DOI] [PubMed] [Google Scholar]
- 5.Song Z, Blumenthal DM. Expanding Payment Reform in Medicare: The Cardiology Episode-Based Payment Model. JAMA. 2016;316:1973–1974. [DOI] [PubMed] [Google Scholar]
- 6.New Cardiac Bundles Could Produce Some Big Winners and Losers. Available from: http://avalere.com/expertise/managed-care/insights/new-cardiac-bundles-could-produce-some-big-winners-and-losers. Access date: 16 Aug, 2018.
- 7.Peterson ED, Roe MT, Chen AY, Fonarow GC, Lytle BL, Cannon CP, Rumsfeld JS. The NCDR ACTION Registry-GWTG: transforming contemporary acute myocardial infarction clinical care. Heart. 2010;96:1798–802. [DOI] [PubMed] [Google Scholar]
- 8.Messenger JC, Ho KKL, Young CH, Slattery LE, Draoui JC, Curtis JP, Dehmer GJ, Grover FL, Mirro MJ, Reynolds MR, Rokos IC, Spertus JA, Wang TY, Winston SA, Rumsfeld JS, Masoudi FA. NCDR Science and Quality Oversight Committee Data Quality Workgroup. The National Cardiovascular Data Registry (NCDR) Data Quality Brief: the NCDR Data Quality Program in 2012. J Am Coll Cardiol. 2012;60:1484–8. [DOI] [PubMed] [Google Scholar]
- 9.Masoudi FA, Ponirakis A, de Lemos JA, Jollis JG, Kremers M, Messenger JC, Moore JWM, Moussa I, Oetgen WJ, Varosy PD, Vincent RN, Wei J, Curtis JP, Roe MT, Spertus JA. Trends in U.S. Cardiovascular Care: 2016 Report From 4 ACC National Cardiovascular Data Registries. J Am Coll Cardiol. 2017;69:1427–1450. [DOI] [PubMed] [Google Scholar]
- 10.Tolles J, Meurer WJ. Logistic Regression: Relating Patient Characteristics to Outcomes. JAMA. 2016;316:533–4. [DOI] [PubMed] [Google Scholar]
- 11.Liu Q, Shepherd BE, Li C, Jr FEH. Modeling continuous response variables using ordinal regression. Stat Med. 2017; 36:4316–4335. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Tsai TT, Patel UD, Chang TI, Kennedy KF, Masoudi FA, Matheny ME, Kosiborod M, Amin AP, Weintraub WS, Curtis JP, Messenger JC, Rumsfeld JS, Spertus JA. Validated contemporary risk model of acute kidney injury in patients undergoing percutaneous coronary interventions: insights from the National Cardiovascular Data Registry Cath-PCI Registry. J Am Heart Assoc. 2014;3:e001380. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Dodson JA, Reynolds MR, Bao H, Al-Khatib SM, Peterson ED, Kremers MS, Mirro MJ, Curtis JP, NCDR. Developing a risk model for in-hospital adverse events following implantable cardioverter-defibrillator implantation: a report from the NCDR (National Cardiovascular Data Registry). J Am Coll Cardiol. 2014;63:788–96. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14.Rao S V, McCoy LA, Spertus JA, Krone RJ, Singh M, Fitzgerald S, Peterson ED. An updated bleeding model to predict the risk of post-procedure bleeding among patients undergoing percutaneous coronary intervention: a report using an expanded bleeding definition from the National Cardiovascular Data Registry CathPCI Registry. JACC Cardiovasc Interv. 2013;6:897–904. [DOI] [PubMed] [Google Scholar]
- 15.Harrell FE. Regression modeling strategies : With applications to linear models, logistic regression, and survival analysis. New York: Springer; 2001. [Google Scholar]
- 16.Sullivan LM, Massaro JM, D’Agostino RB. Presentation of multivariate data for clinical use: The Framingham Study risk score functions. Stat Med. 2004;23:1631–60. [DOI] [PubMed] [Google Scholar]
- 17.Pencina MJ, D’Agostino RB. Evaluating Discrimination of Risk Prediction Models: The C Statistic. JAMA. 2015;314:1063–4. [DOI] [PubMed] [Google Scholar]
- 18.Kumar A, Karmarkar A, Downer B, Vashist A, Adhikari D, Al Snih S, Ottenbacher K. Current Risk Adjustment and Comorbidity Index Underperforms in Predicting Post-Acute Utilization and Hospital Readmissions after Joint Replacements: Implications for Comprehensive Care for Joint Replacement Model. Arthritis Care Res (Hoboken). 2017;69:1668–1675. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Louis Simonet M, Kossovsky MP, Chopard P, Sigaud P, Perneger T V, Gaspoz J-M. A predictive score to identify hospitalized patients’ risk of discharge to a post-acute care facility. BMC Health Serv Res. 2008;8:154. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Alonso Salinas GL, Sanmartin M, Pascual Izco M, Rincon LM, Pastor Pueyo P, Marco Del Castillo A, Garcia Guerrero A, Caravaca Perez P, Recio-Mayoral A, Camino A, Jimenez-Mena M, Zamorano JL. Frailty is an independent prognostic marker in elderly patients with myocardial infarction. Clin Cardiol. 2017;40:925–931. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Weiner JP, Starfield BH, Steinwachs DM, Mumford LM. Development and application of a population-oriented measure of ambulatory care case-mix. Med Care. 1991;29:452–72. [DOI] [PubMed] [Google Scholar]
- 22.Deyo RA, Cherkin DC, Ciol MA. Adapting a clinical comorbidity index for use with ICD-9-CM administrative databases. J Clin Epidemiol. 1992;45:613–9. [DOI] [PubMed] [Google Scholar]
- 23.Elixhauser A, Steiner C, Harris DR, Coffey RM. Comorbidity measures for use with administrative data. Med Care. 1998;36:8–27. [DOI] [PubMed] [Google Scholar]
- 24.Yeh R, Sidney S, Chandra M, Sorel M, Selby J, Go A. Population trends in the incidence and outcomes of acute myocardial infarction. N Engl J Med. 2010;362:2155–2165. [DOI] [PubMed] [Google Scholar]
- 25.Ojeifo O, Berkowitz SA. Cardiology and Accountable Care. Circ Cardiovasc Qual Outcomes. 2015; 8:213–7. [DOI] [PubMed] [Google Scholar]
- 26.Eijkenaar F Key issues in the design of pay for performance programs. Eur J Heal Econ. 2013;14:117–31. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27.Zuckerman RB, Sheingold SH, Orav EJ, Ruhter J, Epstein AM. Readmissions, Observation, and the Hospital Readmissions Reduction Program. N Engl J Med. 2016;374:1543–51. [DOI] [PubMed] [Google Scholar]
- 28.Wasfy JH, Zigler CM, Choirat C, Wang Y, Dominici F, Yeh RW. Readmission Rates After Passage of the Hospital Readmissions Reduction Program: A Pre-Post Analysis. Ann Intern Med. 2017;166:324–331. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29.Desai NR, Ross JS, Kwon JY, Herrin J, Dharmarajan K, Bernheim SM, Krumholz HM, Horwitz LI. Association Between Hospital Penalty Status Under the Hospital Readmission Reduction Program and Readmission Rates for Target and Nontarget Conditions. JAMA. 2016;316:2647–2656. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30.Ibrahim AM, Dimick JB, Sinha SS, Hollingsworth JM, Nuliyalu U, Ryan AM. Association of Coded Severity With Readmission Reduction After the Hospital Readmissions Reduction Program. JAMA Intern Med. 2018;178:290–292. [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.