Development and validation of 15-month mortality prediction models: a retrospective observational comparison of machine-learning techniques in a national sample of Medicare recipients.

OBJECTIVE: The objective is to develop and validate a predictive model for 15-month mortality using a random sample of community-dwelling Medicare beneficiaries. DATA SOURCE: The Centres for Medicare & Medicaid Services' Limited Data Set files containing the five per cent samples for 2014 and 2015. PARTICIPANTS: The data analysed contains de-identified administrative claims information at the beneficiary level, including diagnoses, procedures and demographics for 2.7 million beneficiaries. SETTING: US national sample of Medicare beneficiaries. STUDY DESIGN: Eleven different models were used to predict 15-month mortality risk: logistic regression (using both stepwise and least absolute shrinkage and selection operator (LASSO) selection of variables as well as models using an age gender baseline, Charlson scores, Charlson conditions, Elixhauser conditions and all variables), naïve Bayes, decision tree with adaptive boosting, neural network and support vector machines (SVMs) validated by simple cross validation. Updated Charlson score weights were generated from the predictive model using only Charlson conditions. PRIMARY OUTCOME MEASURE: C-statistic. RESULTS: The c-statistics was 0.696 for the naïve Bayes model and 0.762 for the decision tree model. For models that used the Charlson score or the Charlson variables the c-statistic was 0.713 and 0.726, respectively, similar to the model using Elixhauser conditions of 0.734. The c-statistic for the SVM model was 0.788 while the four models that performed the best were the logistic regression using all variables, logistic regression after selection of variables by the LASSO method, the logistic regression using a stepwise selection of variables and the neural network with c-statistics of 0.798, 0.798, 0.797 and 0.795, respectively. CONCLUSIONS: Improved means for identifying individuals in the last 15 months of life is needed to improve the patient experience of care and reducing the per capita cost of healthcare. This study developed and validated a predictive model for 15-month mortality with higher generalisability than previous administrative claims-based studies.

Clinical metric and medication persistency effects: evidence from a Medicaid care management program.

The objective was to estimate clinical metric and medication persistency impacts of a care management program. The data sources were Medicaid administrative claims for a sample population of 32,334 noninstitutionalized Medicaid-only aged, blind, or disabled patients with diagnosed conditions of asthma, coronary artery disease, chronic obstructive pulmonary disease, diabetes, or heart failure between 2005 and 2009. Multivariate regression analysis was used to test the hypothesis that exposure to a care management intervention increased the likelihood of having the appropriate medication or procedures performed, as well as increased medication persistency. Statistically significant clinical metric improvements occurred in each of the 5 conditions studied. Increased medication persistency was found for beta-blocker medication for members with coronary artery disease, angiotensin-converting enzyme inhibitor/angiotensin receptor blocker and diuretic medications for members with heart failure, bronchodilator and corticosteroid medications for members with chronic obstructive pulmonary disease, and aspirin/antiplatelet medications for members with diabetes. This study demonstrates that a care management program increases the likelihood of having an appropriate medication dispensed and/or an appropriate clinical test performed, as well as increased likelihood of medication persistency, in people with chronic conditions.

Dose-response effects for depression and Schizophrenia management on hospital utilization in Illinois Medicaid: a multivariate regression analysis.

BACKGROUND: The prevalence of schizophrenia and depression in the United States is far higher among Medicaid recipients than in the general population. Individuals suffering from mental illness, including schizophrenia and depression, also have higher rates of emergency department utilization, which is costly and may not generate the positive health outcomes desired. Disease management programs strive to help individuals suffering from chronic illnesses better manage their condition(s) and seek health care in the appropriate settings. The objective of this manuscript is to estimate a dose-response impact on hospital inpatient and emergency room utilizations for any reason by Medicaid recipients with depression or schizophrenia who received disease management contacts. METHODS: Multivariate regression analysis of panel data taken from administrative claims was conducted to test the hypothesis that increased contacts lower the likelihood of all-cause inpatient admissions and emergency room visits. Subjects included 6,274 members of Illinois' non-institutionalized Medicaid-only aged, blind or disabled population diagnosed with depression or schizophrenia. The statistical measure is the odds ratio. The odds ratio association is between the monthly utilization indicators and the number of contacts (doses) a member had for each particular disease management intervention. RESULTS: Higher numbers of intervention contacts for Medicaid recipients diagnosed with depression or schizophrenia were associated with statistically significant reductions in all-cause inpatient admissions and emergency room utilizations. CONCLUSIONS: There is a high correlation between depression and schizophrenia disease management contacts and lowered all-cause hospital inpatient and emergency room utilizations.

Dose-response effects for disease management programs on hospital utilization in Illinois Medicaid.

The objective of this study is to estimate a dose-response impact of disease management contacts on inpatient admissions. Multivariate regression analysis of panel data was used to test the hypothesis that increased disease management contacts lower the odds of an inpatient admission. Subjects were 40,452 members of Illinois' noninstitutionalized Medicaid-only aged, blind, or disabled population diagnosed with asthma, coronary artery disease, chronic obstructive pulmonary disease, diabetes, and/or heart failure. All members are also in the state's Illinois Health Connect program, a medical home strategy in place for most of the 2.4 million Illinois Medicaid beneficiaries. The statistical measure is the odds ratio, which is a measure of association between the monthly inpatient admission indicator and the number of contacts (doses) a member has had for each particular disease management intervention. Statistically significant contacts are between 8 and 12 for heart failure, between 4 and 12 contacts for diabetes, and between 8 and 13 contacts for asthma. Total inpatient savings during the study period is estimated to be $12.4 million. This study shows the dose-response pattern of inpatient utilization improvements through the number of disease management contacts.

An intervention for VA patients with congestive heart failure.

OBJECTIVE: To assess the effect of a telephone intervention to improve quality of life among patients with congestive heart failure (CHF). STUDY DESIGN: Prospective randomized study. METHODS: Single-site recruitment of 458 patients using Veterans Health Administration care into a randomized controlled trial with a 1-year preintervention data collection period and a 1-year intervention and follow-up period. To compensate for imbalanced study groups, propensity scores were included in adjusted models of quality of life, satisfaction with care, inpatient utilization, survival, and costs of care. RESULTS: Patients aged 45 to 95 years participated in the study; 22% were of Hispanic race/ethnicity, and 7% were African American. All but 5 were male, consistent with the older population among veterans. At baseline, 40% were in Goldman Specific Activity Scale class I, 42% were in class III, 6% were in class II or IV, and 12% were unclassified. Patients scored a mean (SD) of 14 (1.5) points below the norm on the physical component score. After the yearlong intervention, no differences in clinical outcomes were noted between the intervention group and the control group. The CHF-related costs were higher for the intervention group, as were overall costs that included the cost of the intervention. Intervention group patients reported better compliance with weight monitoring and exercise recommendations. CONCLUSIONS: A risk-stratified intervention for patients with CHF resulted in potential behavioral improvements but no survival benefit. A high-cost high-intensity intervention may be required to improve survival for patients with CHF. Inclusion of the costs of interventions is recommended for future researchers.

Diabetes disease management results in Hispanic Medicaid patients.

OBJECTIVES: To investigate outcomes of a telephonic nursing disease management program for Medicaid patients with diabetes residing in Puerto Rico. STUDY DESIGN: A 12-month, matched-cohort study. STUDY POPULATION: Four hundred and ninety (490) intervention group members matched to 490 controls. INTERVENTION: Disease management diabetes program. For those in the intervention group, the disease management program customized a self-management intervention plan. MAIN OUTCOME MEASURES: Medical service utilization, including hospitalizations, emergency department visits, physician evaluation and management visits, selected clinical indicators, and financial impact. RESULTS: The intervention group showed significant effects compared with the control group, including a 48% reduction in inpatient bed days, and a 23% increase in ACE inhibitor use, resulting in a return on investment estimate of 3.8:1. CONCLUSIONS: The study demonstrates that a nursing disease management program for diabetes can significantly improve hospitalizations, drug compliance, and vaccinations in a Hispanic Medicaid population.

Cost and utilization avoidance with mail prompts: a randomized controlled trial.

OBJECTIVE: To study the medical service utilization changes and return on investment from a health plan's direct mailings that either encouraged members to receive influenza vaccinations or encouraged members to call a nurse advice service. STUDY DESIGN: Randomized controlled trial with 2 intervention groups and 1 control group consisting of all members over age 65 years who were enrolled in 5 states in the Blue Cross and Blue Shield Government-wide Service Benefit Plan. Sample size was 134,791 individuals. METHODS: Administrative claims-based influenza, pneumonia, heart failure, and respiratory inpatient bed days, emergency department (ED) visits, physician evaluation and management visits, other outpatient visits, and nurse advice call rates were compared between the intervention and control groups. RESULTS: The influenza mailing intervention group experienced 2.87% (P = .033) fewer conditionrelated inpatient bed days and 7.25% (P = .101) fewer condition-related ED visits. The nurse advice service mailing intervention group experienced 7.65% (P <.001) fewer condition-related inpatient bed days and 6.75% (P = .125) fewer condition-related ED visits. Per dollar spent, the return on investment was estimated to be $2.51 for the influenza mailing intervention and $24.24 for the nurse advice mailing intervention. CONCLUSIONS: Administrative claims data suggest that members respond to health plan mailings. By mailing information to their members, health plans can affect rates of medical service utilization and generate cost savings.

Understanding patient risk and its impact on chronic and non-chronic member trends.

A common method of performing commercial and government (ie, Medicare, Medicaid) disease management (DM) program savings evaluations--and the basis of DMAA's Guidelines--is the adjusted historical control method. This method uses a trend adjustment to adjust for the effects of utilization and unit cost changes over time. An appropriate trend adjuster is one that is based on a population with a constant-risk profile, so that utilization and price effects may be measured without being confounded by population changes. Previous literature has demonstrated that the method of identification of chronic and non-chronic members and the timing of the member's transfer between populations has a significant influence on the measured trends in the 2 populations, and thereby on the measured savings from a DM intervention program. The application of risk-adjustment methods to the non-chronic population can correct for this change in risk profiles and ensure a constant-risk population. This method may be used for the non-chronic trend estimation, and will result in an unbiased population trend estimate. However, the chronic population presents different problems for trend adjustment. Because the chronic population is subject to intervention, the application of risk adjustment to this population would potentially neutralize the effect of the outcome that the evaluation is attempting to capture. This paper addresses an alternative method of performing the standard DM savings calculation, which aims to avoid confounding from changes in the chronic population risk profiles due to extrinsic factors.

Evaluation of a medicaid asthma disease management program.

This study evaluates 1-year outcomes of an asthma disease management program implemented in an Oregon Medicaid population. A non-randomized pre-post study, a matched case-control study, and a "programmatic effects" analysis were conducted. Compared to matched controls, the treatment cohort had significantly fewer emergency room visits per thousand (7 vs. 28, P < 0.001) and higher office visits per thousand (57 vs. 7, P < 0.0001) but no significant difference in hospital admission rates. The programmatic effects model identified the participants' initial severity levels and the number of various communications they received as the most important variables in explaining the change in asthma severity from baseline to 12 months. These findings are supportive of the DM design, which is to reduce acute services by improving coordination of care between patients and their providers. Additionally, it appears that there is a close association between the number of patient contacts and their subsequent change in health status.

Opening the door to coordination of care through teachable moments.

The challenge for care coordination is to identify members at a moment in time when they are receptive to intervention and provide the appropriate care management services. This manuscript describes a pilot program using inbound nurse advice calls from members to engage them in a care management program including disease management (DM). Annual medical claims diagnoses were used to identify members and their associated disease conditions. For each condition group for each year, nurse advice call data were used to calculate inbound nurse advice service call rates for each group. A pilot program was set up to engage inbound nurse advice callers in a broader discussion of their health concerns and refer them to a care management program. Among the program results, both the call rate by condition group and the correlation between average costs and call rates show that higher cost groups of members call the nurse advice service disproportionately more than lower cost members. Members who entered the DM programs through the nurse advice service were more likely to stay in the program than those who participated in the standard opt-in program. The results of this pilot program suggest that members who voluntarily call in to the nurse advice service for triage are at a "teachable moment" and highly motivated to participate in appropriate care management programs. The implication is that the nurse advice service may well be an innovative and effective way to enhance participation in a variety of care management programs including DM.

Health services outcomes for a diabetes disease management program for the elderly.

Our objective was to investigate the utilization, drug, and clinical outcomes of a telephonic nursing disease management (DM) program for elderly patients with diabetes. We employed a 24-month, matched-cohort study employing propensity score matching. The setting involved Medicare + Choice recipients residing in Ohio, Kentucky, and Indiana. There were 610 intervention group members over the age of 65 matched to a control group of members over the age of 65. The DM diabetes program employed a structured, evidence-based, telephonic nursing intervention designed to provide patient education, counseling, and monitoring services. Measurements consisted of Medical service utilization, including hospitalizations, emergency department visits, physician evaluation and management visits, skilled nursing facility days, drug utilization, and selected clinical indicators. Among the results, the intervention group had considerably and significantly lower rates of acute service utilization compared to the control group, including a 17.5% reduction in hospitalizations, 22.4% reduction in bed days, 12.3% increase in physician evaluation and management visits, 23.7% increase in angiotensin-converting enzyme (ACE) inhibitor use, 13.3% increase in blood glucose regulator use, 11.8% increase in hemoglobin A1c (HbA1c) tests, 10.3% increase in lipid panels, 26.0% increase in eye exams, and 35.5% increase in microalbumin tests. In conclusion, the study demonstrates that a commercially delivered diabetes DM program significantly reduces hospitalizations and bed-days while increasing the use of ACE inhibitors and blood glucose regulators along with selected clinical procedures such as HbA1c tests, lipid panels, eye exams, and microalbumin tests.

A matched-cohort study of utilization outcomes for an adult medicaid population enrolled in an asthma disease management program.

Few studies have examined the clinical and utilization impact of asthma disease management programs for Medicaid beneficiaries. This study examines utilization and clinical outcomes for an adult group of low- to moderate-risk patients with asthma. Propensity scores are used to construct matched samples of treated-control pairs in order to establish equivalent comparison groups and evaluate the effects of program participation. During the program period, the participants experienced 33.3% fewer hospitalizations, 42% fewer bed days, 87% fewer asthma-related admissions, fewer ED visits, and higher rates of medication usage than those for matched controls, suggesting the beneficial impact of participation for Medicaid program participants.

What can a pilot congestive heart failure disease management program tell us about likely return on investment?: A case study from a program offered to federal employees.

In 1999, the Blue Cross and Blue Shield Federal Employee Program (FEP) implemented a pilot disease management program to manage congestive heart failure (CHF) among members. The purpose of this project was to estimate the financial return on investment in the pilot CHF program, prior to a full program rollout. A cohort of 457 participants from the state of Maryland was matched to a cohort of 803 nonparticipants from a neighboring state where the CHF program was not offered. Each cohort was followed for 12 months before the program began and 12 months afterward. The outcome measures of primary interest were the differences over time in medical care expenditures paid by FEP and by all payers. Independent variables included indicators of program participation, type of heart disease, comorbidity measures, and demographics. From the perspective of the funding organization (FEP), the estimated return on investment for the pilot CHF disease management program was a savings of $1.08 in medical expenditure for every dollar spent on the program. Adding savings to other payers as well, the return on investment was a savings of $1.15 in medical expenditures per dollar spent on the program. The amount of savings depended upon CHF risk levels. The value of a pilot initiative and evaluation is that lessons for larger-scale efforts can be learned prior to full-scale rollout.

The National Program of Cancer Registries: explaining state variations in average cost per case reported.

INTRODUCTION: The Centers for Disease Control and Prevention's National Program of Cancer Registries is a federally funded surveillance program that provides support and assistance to state and territorial health departments for the operation of cancer registries. The objective of this study was to identify factors associated with the Centers for Disease Control and Prevention's costs to report cancer cases during the first 5 years of the National Program of Cancer Registries. METHODS: Information on expenditures and number of cases reported through the National Program of Cancer Registries was used to estimate the average cost per case reported for each state program. Additional information was obtained from other sources, and regression analyses were used to assess the contribution of each factor. RESULTS: Average costs of the National Program of Cancer Registries differed substantially among programs and were inversely associated with the number of cases reported (P < .001). The geographic area of the state was positively associated with the cost (P = .01), as was the regional cost of living (P = .08), whereas the program type (i.e., enhancement or planning) was inversely associated with cost (P = .08). CONCLUSION: The apparent existence of economies of scale suggests that contiguous state programs might benefit from sharing infrastructure and other fixed costs, such as database management resources, depending on the geographic area and population size served. Sharing database management resources might also promote uniform data collection and quality control practices, reduce the information-sharing burden among states, and allow more resources to be used for other cancer prevention and control activities.

A matched-cohort study of health services utilization outcomes for a heart failure disease management program.

Chronic disease is the leading cause of illness, disability, and death in the United States, affecting nearly 100 million Americans. Heart failure alone affects nearly 4.9 million Americans, with another 550,000 newly diagnosed cases each year. The aim of this study was to investigate the program effects of a heart failure care support program. A two-group cohort study matching on propensity scores was used to investigate 277 heart failure care support program participants and corresponding matched non-participants. Measures used were rates of hospitalizations, emergency department visits, physician office visits, and heart failure-related prescription drug use and procedures. Relative to the matched control group, program participants experienced 26.3% (p = 0.023) fewer inpatient admissions, 37.9% (p = 0.018) inpatient bed days, 33.3% (p = 0.059) more beta blocker use, 76.7% (p = 0.048) more alpha blocker use, 22.2% (p = 0.006) more lipid panels, 13.4% (p = 0.019) more electrocardiographies, 50.0% (p = 0.008) fewer cardiac catheterizations, and 94.6% (p = 0.014) more pneumonia vaccinations. The current study employs a propensity score matching methodology to select a subset of comparison patients most comparable to treatment patients, and documents the beneficial health services outcomes of participation in a heart failure care support program.

Reducing medical service utilization by encouraging vaccines: randomized controlled trial.

BACKGROUND: Vaccination against influenza is associated with reductions in hospitalizations for heart disease, cerebrovascular disease, pneumonia, or influenza, and the risk of death from all causes during the influenza season. DESIGN: Randomized controlled trial. PARTICIPANTS: All members enrolled in the Blue Cross Blue Shield Association's Government Wide Service Benefit Program in the states of Oklahoma, Rhode Island, Kentucky, California, Arizona, Utah, and Colorado in October 2002. The sample size was 339,220 members. INTERVENTION: Two identical influenza/pneumonia direct mail marketing pieces that encouraged members to receive influenza and pneumococcal vaccinations. The study period was October 15, 2002 through March 15, 2003 when most influenza cases occur. Data were collected in July 2003 and analyzed during August 2003. MAIN OUTCOME MEASURES: Administrative claims based on influenza/pneumonia inpatient admissions and emergency department (ED) visits. RESULTS: The intervention group experienced a 2.62% (p=0.010) higher rate of influenza vaccinations; 4.61% (p=0.080) higher rate of pneumonia vaccinations; 9.67% (p=0.136) lower rate of influenza/pneumonia inpatient admissions; and 22.64% (p=0.002) lower rate of influenza/pneumonia ED visits compared to the control group. The benefit-cost ratio (return on investment) from this intervention was estimated to be US dollar 2.21 per dollar spent. CONCLUSIONS: Administrative claims data suggest that members respond to health plan mailings with an increase in influenza vaccination rates. Health plans can cost-effectively impact medical service utilization and vaccination rates by mailing information to their members.

Endogeneity bias in the absence of unobserved heterogeneity.

PURPOSE: To demonstrate that endogeneity bias can still arise even when no unobserved heterogeneity exists. METHODS: A formal mathematical proof and a Monte Carlo simulation are used to demonstrate that ordinary estimation techniques will generate biased parameter estimates. RESULTS: The Monte Carlo results support the formal proof. Even in the absence of unobserved heterogeneity, ordinary least squares estimation that does not account for the endogenous nature of an explanatory variable resulted in a parameter estimate for the endogenous variable that was significantly biased (by a factor of 1.42 for the simple model and 1.98 for the saturated model). Alternatively, controlling for endogeneity using the instrumental variables approach led to an unbiased parameter estimate. CONCLUSIONS: Endogeneity bias can still occur even when unobserved heterogeneity is not present.

A matched-cohort study of health services utilization and financial outcomes for a heart failure disease-management program in elderly patients.

OBJECTIVES: To investigate the utilization and financial outcomes of a telephonic nursing disease-management program for elderly patients with heart failure. DESIGN: A 1-year concurrent matched-cohort study employing propensity score matching. SETTING: Medicare+Choice recipients residing in Ohio, Kentucky, and Indiana. PARTICIPANTS: A total of 533 program participants aged 65 and older matched to nonparticipants. INTERVENTION: Disease-management heart failure program employing a structured, evidence-based, telephonic nursing intervention designed to provide patient education, counseling, and monitoring services. MEASUREMENTS: Medical service utilization, including hospitalizations, emergency department visits, medical doctor visits, skilled nursing facility (SNF) days, selected clinical indicators, and financial effect. RESULTS: The intervention group had considerably and significantly lower rates of acute service utilization than the control group, including 23% fewer hospitalizations, 26% fewer inpatient bed days, 22% fewer emergency department visits, 44% fewer heart failure hospitalizations, 70% fewer 30-day readmissions, and 45% fewer SNF bed days. Claims costs were 1,792 dollars per person lower in the intervention group than in the control group (inclusive of intervention costs), and the return on investment was calculated to be 2.31. CONCLUSION: The study demonstrates that a commercially delivered heart failure disease-management program significantly reduced hospitalizations, emergency department visits, and SNF days. The intervention group had 17% lower costs than the control group; when intervention costs were included, the intervention group had 10% lower costs.

Determinants of hospital length of stay for cervical dysplasia and cervical cancer: does managed care matter?

OBJECTIVE: To examine whether type of health insurance plan, among other variables, affects the length of stay for cervical cancer-related hospitalizations. STUDY DESIGN, PATIENTS, AND METHODS: Inpatient admission claims records for cervical dysplasia and cervical cancer were selected for 1994-1997 from the MarketScan private health insurance claims database. After identifying records by stage of disease and deleting records for pregnant women, 1145 unique patient records were used in a truncated count regression model to analyze the predictors of hospital length of stay. RESULTS: All later stages of disease were associated with a longer hospital stay. After controlling for other variables, the coefficients showed an increase in predicted length of admission ranging from 2.5 days for stage I to 6.3 days for stage IV cervical cancer compared with dysplasia/carcinoma in situ (all stages, P < .01). There was no significant statistical difference in the lengths of stay for patients covered under comprehensive fee-for-service plans vs other types of health insurance plans, including managed care. CONCLUSIONS: Managed care plans are often thought to contain healthcare costs by shortening the hospital length of stay. Our findings show no association between managed care plans and hospital length of stay for women with cervical cancer or its precursors.

Diabetes disease management in a community-based setting.

PURPOSE: The medical cost of diabetes in the United States in 1997 was at least $98 billion. This study illustrates the behavioral change and medical-care utilization impact that occurs in a community-based setting of a diabetes disease-management program that is applied to program participants in a health insurance plan's health maintenance organization and preferred provider organization. DESIGN: A historical control comparison of diabetes-management participants. METHODOLOGY: One hundred twenty-seven identified diabetes patients are followed from baseline through 1 year. Differences in behavior are compared at program intake and at a 6-month reassessment. Differences in medical-service utilization are compared in the baseline year and the year subsequent to program enrollment. Poisson multivariate-regression models are estimated for counts of inpatient, emergency department, physician evaluation and management, and facility visits, while also controlling for potential confounders. PRINCIPAL FINDINGS: Behaviors improved between program intake and the 6-month reassessment. From patient reports, the number of participants having a hemoglobin A1c test increased by 44.9 percent (p < .001), and there was a 53.2-percent decrease in symptoms of hyperglycemia (p = .002). From medical claims after program enrollment, a drop occurred during the program year in every dimension of medical-service utilization. Regression results show that in-patient admissions decreased by 391 (p < .001) per 1,000 for each group, while controlling for age, length of membership, and the number of comorbid claims for congestive heart failure. In the analysis of costs that were pre- and post-enrollment, which included disease-management program costs, a 4.34:1 return on investment was calculated. CONCLUSION: The diabetes program provides patients with comprehensive information and counseling relative to practicing self-management of diabetes through a number of integrated program components. This study strongly suggests that the implementation of such a program is associated with positive behavioral change and, thus, with substantial reduction in medical-service utilization. In addition, the intervention resulted in a net decrease in direct medical costs.