Elizabeth Y. Rula, PhD

 To evaluate the Care Transition Solution (CTS) as a means to improve quality through reduction of preventable hospital readmissions among patients with readmission-sensitive conditions subject to penalties imposed by the Affordable Care Act.

 A retrospective quasi-experimental evaluation of the impact of the CTS among admitted patients diagnosed with heart failure, acute myocardial infarction, chronic obstructive pulmonary disease, and/or pneumonia (CMS readmission-penalty diagnoses) in 14 acute care hospitals in Texas. The program, designed for scalable delivery, incorporated identification of high readmission—risk patients, assessment of individual needs, medication reconciliation, discharge planning, care coordination, and telephonic postdischarge follow-up.

The treatment group of program enrollees (N = 560) and the comparison group with no program contact (N = 3340) were matched on 8 coarsened demographic, diagnosis, and severity variables associated with readmission risk. Assessed outcomes included relative risk and odds of readmission within 30 days postdischarge and overall within the 6-month evaluation period. Zero-inflated Poisson multivariate models were used to estimate intervention effects controlling for matching-generated weights, age, disease status, and period of evaluation.

Treatment group risk of readmission was 22% lower overall (incidence rate ratio [IRR], 0.78; 

 <.01) and 30-day readmission risk was 25% lower (IRR, 0.75; 

= .01) relative to the comparison group. Odds of any or 30-day readmission were 0.47 (95% CI, 0.35-0.65) and 0.56 (95% CI, 0.41-0.77), respectively, for treatment relative to comparison.

Participation in the CTS resulted in significantly lower rates of readmissions among patients with readmission-sensitive conditions, offering a scalable and sustainable approach to reduce the number of preventable hospital readmissions.

	There is increasing pressure to reduce hospital readmission rates by improving quality while maximizing efficiency. The evaluated collaborative care transitions program provides a scalable approach to reduce readmissions. 

		Scalable delivery of the Care Transitions Solution (CTS) is achieved through predictive early identification of patients at high risk of readmission and telephonic follow-up. 

		Among program participants, the 30-day readmission risk was reduced by 25% and the odds of any readmission within 30 days were reduced by more than half, relative to matched nonparticipants. 

		The CTS can be efficiently implemented by hospitals and is an effective system to reduce the number of preventable readmissions.

The changing healthcare landscape has amplified the need to improve healthcare quality while also lowering healthcare spending, which reached $2.8 trillion in the United States in 2012.

 A comparison of 12 countries showed the proportion of US gross domestic product spent on healthcare was more than 40% higher than in second-ranked France, yet measures of quality for US healthcare are often lower than in countries that spend far less. High rates of hospitalization for chronic conditions in the United States are one example; they have been attributed to low performance in care coordination and safety, and only average levels of care effectiveness and patient-centered care.

 Although these observations could reflect a dysfunctional system, they are more likely the result of a healthcare system increasingly strained by a surging prevalence of complex conditions and concomitant care needs.

A focus on reducing hospital readmissions—which contribute to overall hospitalization rates&mdash;is recognized as an opportunity to improve care while also reducing avoidable costs. A study of nearly 12 million Medicare hospital discharges in 2004 found that approximately 20% resulted in 30-day readmissions, with only one-tenth of those readmissions likely planned. The total costs of these unplanned readmissions were over $17 billion.

Diagnoses that are particularly readmission-sensitive include heart failure (HF), acute myocardial infarction (AMI), and pneumonia. In one study, 30-day readmission rates for these diagnoses were 24.8%, 19.9%, and 18.3%, respectively, in a Medicare population.

 These conditions also account for a significant amount of hospitalization and cost in Medicaid, commercially insured, and uninsured US populations.

 Readmission rates are concerning for these and other serious conditions across all age and payer groups.

In 2012, the Affordable Care Act established strong financial incentives for hospitals and physicians to reduce readmissions. The law required CMS to establish the Hospital Readmissions Reduction Program, which penalizes, through reduced reimbursement, those hospitals whose risk-adjusted readmission rates for HF, pneumonia, and AMI exceed the national average. In 2015, the program expanded to include chronic obstructive pulmonary disease (COPD) and elective hip or knee replacement as readmission penalty conditions.

 and state Medicaid agencies will likely follow with similar programs to raise quality and reduce readmissions.

High rates of unplanned readmissions are a reflection of the fact that the US healthcare system has traditionally functioned in a fragmented and poorly coordinated fashion, often leaving discharged patients and their family members uneducated, confused, and unprepared for the ongoing management of conditions to avoid future adverse events.

 Inadequate hand-off of patient management among providers, poorly coordinated hospital discharge processes, discharge instructions for the patient that lack sufficient education and follow-up, and lack of medication reconciliation before and after hospitalization all amplify the likelihood that a patient will return to the hospital. Illustrating the consequences of these problems, the Medicare Payment Advisory Commission’s 2007 report to Congress estimated that 76% of 30-day readmissions are potentially preventable.

Although broad efforts to address gaps in the discharge and hand-off process can benefit all hospitalized patients, complex cases at higher risk for readmission often require additional resource-intensive planning and follow-up processes. Full implementation of comprehensive approaches established in the literature

 for large populations or across multiple hospitals is a formidable initiative given resource constraints in the current healthcare environment. Thus, there is a need for scalable approaches delivering effective readmission prevention models that augment what hospitals, health plans, and providers are already doing to prevent unplanned readmissions.

Texas Health Resources (THR), a nonprofit, faith-based health system located in north Texas, is innovating through exploration and testing of scalable approaches to prevent readmissions. Specifically, THR—through partnership with Healthways, a global well-being improvement company&mdash;implemented the Healthways Care Transitions Solution (CTS) in 14 acute care hospitals. CTS uses predictive identification of admitted patients at high risk of readmission, then invokes a collaborative care model for discharge planning and follow-up that extends from the hospital to the home. The current study was conducted to assess the initial effectiveness of CTS participation in preventing readmissions among patients with readmission-sensitive conditions.

A quasi-experimental retrospective cohort study was conducted to test the hypothesis that CTS program participation among admitted patients with readmission-sensitive conditions was associated with reduced readmission risk relative to a matched nonparticipant group. Data originated from 14 acute care hospitals in the Texas Health Resources network; specifically, admission-discharge-transfer data documenting patient admissions during the initial 6-month intervention time period of January 1 to July 1, 2013, were used for this study. Because this study was a retrospective analysis of a quality improvement initiative conducted anonymously, it did not require informed consent from participants and was exempt from institutional review board approval based on exclusion criteria outlined in the US Code of Federal Regulations (45 CFR §46.101).

Implementation of the CTS intervention was staggered over a 6-week period beginning in mid-January 2013; the program initiated at 2 to 3 hospitals each week. By March 2013, the CTS intervention program was fully functioning at all 14 sites. The CTS program was designed to deliver thorough and personalized patient education and discharge planning/preparation, provide regular individualized follow-up, and facilitate care coordination in order to avoid unnecessary visits/time spent in medical facilities or doctors’ offices, while encouraging appropriate medical care to avoid additional exacerbations of the patient’s condition.

Basic principles of the CTS program include the following: 1) Identification of patients at high risk of readmission using a predictive model, and clinical assessment to ensure alignment of resources with need; 2) Early engagement of admitted patients by a nurse transition coach to establish a strong relationship with the patient; 3) Detailed collection of contact information to facilitate postdischarge patient interactions; 4) Assessment of medical, psychosocial, functional, literacy, adherence, and support needs, and of capabilities such as functioning and self-efficacy, to tailor interactions accordingly; 5) Reconciliation of medications before hospitalization to medications after hospital discharge; 6) Provision and review of a patient-oriented Care Transition Record, with documented discharge plan; 7) Coordination of medical providers and service agencies for postdischarge patient care; and 8) Postdischarge telephonic follow-up (4 calls over 4 weeks) to track and support the patient’s recovery and ongoing self-management, and to encourage discharge plan adherence.

The study-eligible population included all patients admitted to 1 of the 14 evaluated hospitals during the study period who were documented to have at least 1 of the following readmission-sensitive conditions that are also CMS readmission-penalty diagnoses: AMI, HF, pneumonia, or COPD. In accordance with intention-to-treat research design, all CTS enrollees were eligible for the treatment group, regardless of their level of participation or completion of the CTS program, to provide a realistic assessment of program effectiveness. The eligible population had an average age of 59.3 years (range = 18-96 years). Eligible comparison group members did not have any documented interaction with the CTS program. The staggered implementation across hospital sites provided a population of program-eligible patients admitted to sites where the program was not yet available, and although the comparison group was not constrained to patients admitted pre-implementation, the availability of these patients in the matching process diminished the potential for selection bias and availed a more equivalent group for comparison.

Drawing from this eligible population, comparable study groups were created using coarsened exact matching (CEM), which exactly matches individuals within a nonparametric framework into strata based on a set of shared characteristics (coarsened variables) chosen to explain selection bias and variance in the outcome. CEM typically requires removal of fewer cases for a given level of bias removal and thus is comparatively a more effective and efficient method for yielding an unbiased estimate of treatment effect.

 Matching variables included age group; gender; dichotomous indication of whether or not initial admission resulted from AMI, pneumonia, HF, or COPD; initial admission length-of-stay group; depression status

; and disease count (range = 0-5, counting discrete indications of COPD, HF, coronary artery disease, diabetes, and chronic kidney disease). After strata assignment, matched comparison members were assigned a weight specific to their stratum and representative of the relative proportion of members in that stratum; treatment group members were assigned a weight of 1.

 CEM weights were used as a covariate to adjust estimates in multivariate statistical modeling.

Study groups were compared on 2 metrics: readmissions within 30 days of prior admission and all readmissions occurring within the 6-month study time period. The initial (index) admission for each CTS group member was defined as the admission from which the member was enrolled in the CTS program. For each comparison group member, the index admission was the first admission record during the study period. A readmission was defined as a hospital admission occurring subsequent to the discharge date, documented for a previous admission during the study period, allowing for more than 1 readmission in the study period. Same-day readmissions, in which a discharge and subsequent admission occurred within 24 hours, were excluded as these cases generally denote transfers as opposed to readmissions. For each study member, hospital admission records were assessed from their index admission to the July 1, 2013, study end point.

The quantitative measure to evaluate the extent of imbalance and heterogeneity between the study groups was the L1 metric, a nonparametric measure generated in the CEM process that quantifies imbalance by comparing relative frequencies of the 2 groups across each of the strata.

 Values of L1 close to 0 indicate a higher quality match with minimal imbalance, whereas an L1 value of 1 indicates complete dissimilarity or disproportionality between the groups (no overlap between groups in the strata assignment). Matching variables are optimized with respect to providing a low L1 with high retention of treatment group members. For additional confirmation of intergroup comparability, a Wald test was estimated to determine if the independent variables in multivariate modeling provided similar explanatory contribution to the dependent variable (readmissions) for both study groups. A small Wald statistic and a large 

 value indicate sufficient similarity between groups to obtain a reliable estimate of intervention effect.

Zero-inflated Poisson multivariate models were used to estimate intervention effects on all readmissions and on 30-day readmissions while accounting for potential confounders. These control variables included age (continuous), evaluation window (time from index admission to end of study period), CEM-generated weights, and disease status for COPD, HF, pneumonia, AMI, diabetes, chronic kidney disease, and coronary artery disease, as well as the status for each of these conditions, as documented during the index admission. Adjusted incidence rate ratios (IRRs; relative risk) were produced from the 0-inflated Poisson model by taking the exponential of the intervention variable coefficient, using the comparison study group as the reference. Poisson multivariate count regression models with least squares means statements were used to calculate adjusted daily readmission rates for each study group. Adjusted daily readmission rates were then converted to annualized readmission rates per 1000. Multivariate logistic regression was used to estimate adjusted odds ratios (ORs) and 95% CIs of intervention effects on the likelihood of having versus not having at least 1 readmission using the same control variables. Data manipulation and analysis were performed using SAS version 9.2 (SAS Institute, Cary, North Carolina).

Descriptive characteristics of the full population of admitted patients meeting study eligibility requirements and the matched treatment and comparison study groups are displayed in 

. CEM matching resulted in the pruning of 12 (2.1%) treatment members and 726 (17.9%) comparison members from the initially eligible populations. Although some improvement in balance was achieved through this pruning, adjustment using CEM weights resulted in near equivalence between the groups across all matching variables. The mean evaluation window (days from index admission to end of the 6-month study period) was also similar for the treatment and comparison groups at 88.9 and 81.8 days, respectively. Remaining differences, including evaluation window, were controlled for using additional covariates in statistical modeling of outcomes.

Further verification of comparability between the study groups was provided by L1 and Wald statistics. 

 displays an L1 postmatch statistic of 4.09 E

indicating minimal imbalance between study groups and a dramatic improvement in balance from the pre-match L1. The summary Wald test statistic and large associated 

 = 1.0) provided confirmation of intergroup comparability with respect to all modeled covariates.

Adjusted IRRs showed significantly lower rates of all readmissions (

 = .0107) in the treatment group relative to the comparison group (

). The treatment group displayed 22% fewer overall readmissions and 25% fewer 30-day readmissions in contrast to the comparison group. As a reference point for the relative rates, the adjusted annualized readmission rate per 1000 patients was 603.0 for the treatment group and 771.4 for the comparison group.

Logistic analysis results indicated that treatment group members were significantly less likely to have had any readmission or a 30-day readmission during the study period, with adjusted odds being 0.47 and 0.56, respectively, relative to comparison group members (

). Changing the reference group, the adjusted odds of having 1 or more readmissions during the study period was 2.1 times higher in the comparison group relative to the treatment group (adjusted OR, 2.1; 95% CI, 1.5-2.9). Results for 30-day readmissions were of similar magnitude (adjusted OR, 1.8; 95% CI, 1.3-2.5).

Across the board, results indicate that the CTS intervention significantly reduced hospital readmissions for THR patients with readmission-sensitive conditions. Participation in the CTS program was significantly associated with lower readmission incidence, both for 30-day readmissions and all readmissions occurring over the 6-month study period. Additionally, admitted patients who did not participate in CTS had approximately twice the odds of having a 30-day readmission or any readmission during the study period relative to the treatment group.

The estimated reduction in readmissions for the entire 6-month period demonstrates the sustained intervention impact, extending beyond the 30-day window that is often the bar for efforts aimed at preventing the immediate ramifications of failures to provide effective discharge instructions and planning. The measured impact of CTS is beyond that of more global hospital initiatives to reduce readmissions from which all patients benefit. The CTS model also intends to maximize cost efficiency and scalability through sophisticated prediction of high-risk cases for selective supplemental support and by outsourcing program operations, data management, predictive modeling, and postdischarge telephonic follow-up.

The CTS program builds on a remote telephonic intervention, also delivered by Healthways, that one study found was associated with a 23% lower likelihood of readmission within 30 days relative to the comparison group.

 CTS represents an enhancement of this telephonic model, one that allows for early intervention and a direct connection to providers. For example, in the prior study, Harrison et al highlighted the delayed delivery of patient discharge information and, consequently, delayed initiation of telephonic follow-up—problems that were due to claims-based identification without any direct data feed from the hospital or communication with providers. Initiating the intervention within a hospital environment with access to timely patient information ensures continuity of patient support across different settings, allows a relationship to be established with the patient prior to discharge, and permits rapid follow-up after discharge. Congruent with Kripalani et al, who concluded that single-component interventions are less impactful,

 the effect of CTS exceeded what was reported for the described patient-centric telephonic-only model while also using a cost-efficient approach.

In addition to prior research supporting the benefit of telephonic follow-up, there is foundational evidence for other aspects of the CTS program reducing readmissions. A randomized study of nurse and pharmacist support to reinforce the discharge plan and review medications found a 0.695 IRR for 30-day readmissions compared with controls.

 A randomized trial of an intensive transitional care model delivered by advanced practice nurses to patients 65 years and older reported a significantly lower 24-week readmission rate for treatment relative to controls (20.3% vs 37.1%).

 Coleman et al reported positive results from the Care Transitions Intervention using both a quasi-experimental approach

 The trial, on admitted patients 65 years and older, showed significant reductions in treatment group readmission rates, with 30- and 90-day readmission adjusted ORs of 0.59 and 0.64, respectively. Similar findings have been published more recently on the real-world effectiveness of the Care Transitions Intervention using a quasi-experimental design.

Comparison of the present study with published randomized trial results indicates the CTS program has performed comparably in readmission avoidance. The 30-day readmission incidence rate ratio of 0.75 for the CTS program was comparable with the 0.695 IRR reported by Jack et al,

 as was the CTS 30-day readmission OR of 0.56 compared with the Coleman et al

 trial result of 0.59. This is especially significant when taking into account that the CTS program was evaluated under “real-world effectiveness” conditions, whereas randomized trials are conducted in a more controlled setting.

The need for real-world effectiveness studies on readmissions has been acknowledged in the literature. Voss et al write, “Patients who agree to participate in randomized controlled trials are a select subset of that population, limiting the generalizability of these observations.”

 Although Coleman’s Care Transition Intervention has been tested extensively, both in randomized trials

 there is value to testing alternative approaches to improve quality of healthcare solutions. A given solution is not necessarily applicable or practical in all environments. The use of a predictive model to selectively deliver this program to patients at higher risk, then outsourcing telephonic follow-up, makes this evaluated program potentially more operationally appealing than programs that are delivered less selectively and that require greater hospital staff time in order to intervene with a larger number of people and to follow up after discharge.

The retrospective design presents a limitation to the study, as it necessitated the creation of a comparison group from a convenience sample as opposed to prospective selection and randomization. Further, generalizability of study results is unknown, but the multisite design strengthens the likelihood that results would translate to other hospitals. Future work should evaluate additional diseases and conditions beyond the penalty diagnoses evaluated here, as well as program effectiveness over a longer duration and in different institutions, demographic groups, and geographic regions.

The changing architecture of healthcare reimbursement is requiring hospitals to quickly find solutions to improve quality metrics. In the first year of the Hospital Readmissions Reduction Program, 2200 hospitals received cumulative penalties of $280 million.

 Given that reported readmission outcomes are publicly available on the CMS Hospital Compare website, the revenue impact of negative press may be as detrimental as the reimbursement penalties. These changes in the healthcare market highlight the importance of the current study results, which indicate that CTS offers an efficient and effective solution for health systems, hospitals, and large provider groups that are seeking support in reducing readmissions among high-risk patients.

Implementation of the CTS program by THR succeeded in significantly reducing readmissions for enrolled patients diagnosed with COPD, HF, AMI, or pneumonia meeting the recognized need to improve care and associated outcomes among a growing number of higher-complexity cases. An innovative approach to improving patient support and care coordination, the CTS program is a collaborative model designed as a scalable and sustainable approach that maximizes efficient resource use while improving quality of care and continuity in the transition from hospital to home. The implications may extend beyond the evaluated diagnoses given that CTS is applicable to additional conditions that may be included within future expansion of the readmission avoidance directive. Overall, CTS is a viable option for institutions wishing to efficiently implement an effective approach to reducing avoidable readmissions. 

 Center for Health Research, Healthways, Inc (BH, EYR, ARW, CC, JEP), Franklin, TN; Texas Health Resources (DV), Arlington, TX.

 The study was funded by Healthways, Inc.

Articles

The 30-day readmission risk was reduced 25% by a collaborative program model employing discharge planning and telephonic follow-up for high-risk patients with CMS penalty diagnoses.

Impact of a Scalable Care Transitions Program for Readmission Avoidance

 To validate a predictive model for identifying Medicare beneficiaries who need end-of-life care planning and to determine the impact on cost and hospice care of a telephonic counseling program utilizing this predictive model in 2 Medicare Health Support (MHS) pilots.

Secondary analysis of data from 2 MHS pilot programs that used a randomized controlled design.

A predictive model was developed using intervention group data (N = 43,497) to identify individuals at greatest risk of death. Model output guided delivery of a telephonic intervention designed to support educated end-of-life decisions and improve end-of-life provisions. Control group participants received usual care. As a primary outcome, Medicare costs in the last 6 months of life were compared between intervention group decedents (n = 3112) and control group decedents (n = 1630). Hospice admission rates and duration of hospice care were compared as secondary measures.

The predictive model was highly accurate, and more than 80% of intervention group decedents were contacted during the 12 months before death. Average Medicare costs were $1913 lower for intervention group decedents compared with control group decedents in the last 6 months of life (P = .05), for a total savings of $5.95 million. There were no significant changes in hospice admissions or mean duration of hospice care.

Telephonic end-of-life counseling provided as an ancillary Medicare service, guided by a predictive model, can reach a majority of individuals needing support and can reduce costs by facilitating voluntary election of less intensive care.

Using a randomized controlled design, we found that telephonic end-of-life counseling

guided by a predictive model successfully reached a majority of Medicare beneficiaries

needing end-of-life support and had an impact on care delivery.

		 Predictive modeling and telephonic delivery allow for the effective widespread provision of end-of-life counseling outside of the physician’s office.

		 Nurse-delivered counseling and education about provisions available to improve quality at the end of life can reduce costs in the last 6 months of life among Medicare beneficiaries.

		Telephonic end-of-life counseling is a promising approach for improving quality and reducing costs in the Medicare program.

Individuals nearing the end of life need support in navigating the choices available to help them maximize the quality of their remaining time. Research shows that the conventional medical system often fails to provide sufficient education and counseling on relevant topics such as palliative care.

Barriers to more widespread end-of-life support in primary and specialty care settings include limited time and resources, discomfort with the topic, and reluctance to give up hope of recovery.

Finding complementary means for this support could overcome these barriers and provide substantial value.

In the absence of documented care choices, life-sustaining measures are provided by default, often contrary to the patient’s wishes.

 These measures, such as artificial ventilation and feeding, are responsible for considerable economic strain, especially for the Medicare program, which uses more than a quarter of its funds for end-of-life care.

 Compared with the costs for survivors in a given year, decedent costs are 6 times higher

 The value of expenditures for aggressive care at the end of life is questionable, because higher end-of-life costs are associated with poorer quality of life in its final stages

 and more difficult bereavement adjustment for caregivers. 

 Alternatively, programs that deliver end-of-life support have been shown both to improve quality of life and reduce costs associated with unwanted care.

Reducing end-of-life costs can be achieved not by denying care, but by facilitating informed, thoughtful, and documented decisions about the value of intensive therapy, often decreasing the demand for aggressive care and life-prolonging measures while increasing hospice enrollment and improving quality of life.

 Advance directives have shown to improve the quality of remaining life and ease family member stress

while also reducing inpatient expenditures.

 Despite these advantages, a study of Medicare decedents found that less than a third had documentation of advance directives,

 although 70% of patients would opt against lifesustaining treatment in the case of incapacitation and poor prognosis.

 Hospice care, similarly underutilized, is associated with better patient quality of life,

decreased likelihood of dying in a hospital,

Unfortunately, only 19% of Medicare decedents use hospice prior to their death,

 and, among all hospice enrollees, 30.8% receive care for 7 days or less.

The Centers for Medicare & Medicaid Services (CMS) is charged with evaluating new approaches to improve the quality of care, increase satisfaction, and reduce cost in the Medicare program. One such evaluation, Medicare Health Support (MHS, formerly Voluntary Chronic Care Improvement Program),

 focused on care management interventions delivered to high-cost Medicare fee-for-service (FFS) beneficiaries with chronic disease, specifically diabetes and/or heart failure, sources of disproportionate spending.

As a focused intervention within the context of the chronic care management programs in 2 MHS pilot programs, we developed and delivered telephonic end-of-life counseling. Initial experience with the pilot population validated the prepilot projections of a high death rate

 and the challenges inherent in timely identification of participants in need of end-of-life support. Because the MHS study design encouraged innovations in the applied solution, we developed and subsequently refined a predictive model to identify individuals in the MHS population who were most likely to be in their last year of life. The final validated predictive model proved to be both sensitive and specific, enabling us to prioritize MHS participants in greatest need of end-of-life support and counseling. Furthermore, a follow-up cost analysis found that

decedents in the intervention group had significantly lower Medicare expenditures in their last 6 months of life compared with controls. This outcome demonstrated that an end-of-life intervention delivered as part of a coordinated-care program can reach and engage the appropriate patients successfully, and enact change.

Development of End-of-Life Predictive Model

A neural network predictive model was developed to identify individuals in the MHS intervention population at greatest risk of death. The development of the model was a staged process, allowing for refinements during the pilot. The model was built with SAS Enterprise Miner, version 4.3 (SAS Institute Inc, Cary, NC), using demographic data and 2 years of Medicare claims data from the MHS intervention cohort (N = 43,497). From the data were extracted 130 predictive modeling factors, classified

into demographic, pharmaceutical, utilization, diagnosis, medical procedure, and financial groupings, then refined to the set of independent variables significantly associated with death in the following year. The model functions to predict which individuals are likely to die over the next 12 months based on an evaluation of these variables in data from the previous 12 months.

The first iteration of the end-of-life model was developed to predict likely decedents among beneficiaries who had an inpatient claim because hospitalizations are a correlate of mortality.

 This preliminary model was implemented in conjunction with the initiation of the focused end-of-life intervention in June 2006. Further development of the model expanded the focus to the entire population, eliminating the hospitalization requirement. The optimized, populationbased, end-of-life predictive model was deployed in November 2007 and was used through the remainder of the pilot programs (see the 

A cross-validation data sampling technique was used during model development that involved holding aside a dataset from the model training and calibration phases to test the model against novel data. This approach was used to identify multiple potential models for consideration, from which the final population-based predictive model was chosen based on performance data from the entire MHS intervention group (N = 43,497). Performance was evaluated through a comparison of model output generated from 12 months of data (predicted risk of death) with the actual decedents in the subsequent 12 months. Predictive power was quantified using receiver operating characteristic (ROC) C statistic and sensitivity analyses.

Each beneficiary prioritized by the model was assigned a designated nurse trained in end-of-life counseling. Prioritization by the model was designed to augment identification of participants needing end-of-life support and did not supersede the clinical judgment of nurses during patient interactions. The intervention was delivered telephonically to the patient and/or the caregiver via outbound or inbound calls. Interactions included, but were not limited to, education and counseling about advance directives; the choice of palliative versus aggressive care, and hospice enrollment; facilitation of interactions with physicians and hospice agencies and referrals to hospice when appropriate; and caregiver support. The intervention was a focused element of the chronic care management provided to MHS participants that addressed

standards of care, appropriate healthcare utilization, and self-efficacy. Under all circumstances, patients were advised to remain adherent to their physician’s care plan and recommended chronic care management. Issues with connectivity (eg, no answer, wrong phone number) prevented successful contact with all identified participants. Similarly, phone call duration and frequency were variable as a function of the participant’s needs.

The primary measure of success for the intervention was the deliberate election of less aggressive care as evidenced by reduced medical spending in the last 6 months of life. Secondary measures of success were an increased rate of hospice enrollment and duration of hospice care prior to death.

The study population was derived from participants in the pilot programs of 2 MHS organizations that were randomized by CMS into intervention and control groups; MHS study design and eligibility requirements are outlined in the eAppendix.16 Briefly, MHS participants were Medicare FFS beneficiaries with diabetes and/or heart failure, and with Hierarchical Condition Category (HCC) scores 

1.35. Participants were randomized based on heart failure diagnosis, HCC risk score category, and dual Medicare/Medicaid eligibility. The MHS pilot programs and evaluation met institutional review board exemption criteria for research and demonstration projects on public benefit and service programs [Protection of Human Subjects, 45 CFR §46.101(b)(5)]. The Secretary of Health and Human Services communicated to each targeted beneficiary that participation was voluntary and could be terminated at any time.

With respect to the end-of life intervention, all CMSdefined intervention group participants were eligible and were prioritized for this intervention using the predictive model. Control group participants were not eligible for the intervention and were not entered into the predictive model. For this analysis, these CMS-defined intervention and control groups were subject to additional eligibility requirements. Individuals eligible for this analysis were those who died during the final year of their respective pilot programs (decedents), who were not enrolled in hospice at the start of the final year of the pilot, who were enrolled from the beginning of the pilot until death or entry into hospice, and who had 1 or more medical claims during the pilot. Based on these requirements, 3112 and 1630 individuals were subject to analysis in the intervention group and control group, respectively. This intention-to-treat analysis included data from all participants eligible for the analysis, irrespective of whether end-of-life interactions occurred.

The cost analysis included all medical claims incurred during the 6 months prior to death with the exception of those incurred after hospice enrollment. These costs were excluded because MHS eligibility ended at the point of hospice enrollment, and complete claims data were not available after this point.

 Costs were calculated according to the CMS modified protocol

for financial performance evaluation of the MHS pilots and included both outlier adjustment and baseline adjustment to compensate for a divergence in Medicare costs between intervention and control groups that emerged from the time of randomization to the start of the pilot. Nonparametric statistical testing was used (Mann- Whitney-Wilcoxon test) to determine the significance of the adjusted costs because the financial data were not normally distributed. Statistical analyses were performed using SAS statistical analysis software version 9.1.3. Analysis of hospice

care was based on the total number of hospice enrollments during the last 6 months of life and the average duration of hospice care among hospice enrollees. This study was a privately funded analysis of data from the MHS pilots, and all authors are employees of the funding institution.

Evaluation of the end-of-life predictive model was performed by comparing model output generated from a historical dataset (N = 43,497) with actual deaths using ROC analysis, a statistical method to assess model accuracy based on sensitivity and specificity. The first iteration of the model was 71% accurate (ROC C statistic = 0.712). This accuracy rate was improved upon in the final model, which was 90% accurate (ROC C statistic = 0.899). The sensitivity of the final model was further tested by determining the percentage of actual deaths in a 12-month time frame that were predicted by the model at defined predicted death score stratification levels (

). For example, individuals with scores in the top 10% accounted for 51.8% of all deaths during this time frame.

A telephonic counseling program, directed by a predictive model, reduced end-of-life costs by 4.5% within 2 Medicare Health Support pilot programs.

Elizabeth Y. Rula, PhD

Articles

Impact of a Scalable Care Transitions Program for Readmission Avoidance

Impact of Predictive Model-Directed End-of-Life Counseling for Medicare Beneficiaries