In the fall of 2009, a comprehensive sample of all beneficiaries of the General Regional Health Fund (Allgemeine Ortskrankenkasse [AOK]) from 10 primary care practices in south-western Germany were screened for potential participants of a CM intervention comparing patient identification by PCP or PM. Further details of the selection process and patients’ characteristics at baseline have been published previously . The 10 primary care practices (5 single-handed practices and 5 group practices) were recruited from rural areas (3 practices) and (sub-)urban areas (7 practices). We obtained de-identified insurance claims data including medical and pharmacy claims from January 2007 to December 2010 from AOK beneficiaries of all ages. This study was part of a series of studies to develop a CM program for high-risk patients in primary care . The University Hospital Heidelberg Institutional Review Board approved the study.
Predictive modelling software
For PM-based case finding, the AOK used the commercial software package Case Smart Suite Germany (Verisk Health, Munich, Germany) . This program is an extension of diagnostic cost group-PM, which has been applied previously in comparative case finding studies . Information from the past 2 years (2007–2008) served as inputs for PM, including all International Statistical Classification of Diseases, (10th Revision, German modi-fication, ICD-10-GM) diagnostic codes assigned in outpatient and inpatient settings, prior costs, hospital admissions, and demographic data. Based on multivariate logistic regression analysis, the software computes the likelihood of at least one hospitalization (LOH) for each individual within the next 12 months (2010). In our study, patients with a LOH score above the 90th percentile of the sample were classified as ‘high-risk’ and identified as potential participants of CM without any restrictions of the number of patients selected per practice. Due to the design of the planned CM intervention we focused on patients with at least one of the following ICD-10-GM index conditions : type 2 diabetes mellitus (codes E11-E14), chronic obstructive pulmonary disease (J43-J44), asthma (J45), chronic heart failure (I11.0, I13.0, I13.2, I25.5, I50), and late-life depression (F32-F33 [>60 years]). Patients with dementia (F00-F03), dialysis (Z49, Z99.2), or active cancer disease (C00-C97) were excluded.
Fourteen PCPs from the 10 participating primary care practices were asked to screen a list of all AOK beneficiaries in their practice and to identify up to 30 patients for future participation in a CM program aimed at reducing avoidable hospitalizations. PCPs were informed about the aims and intervention elements of the planned CM intervention but no explicit selection criteria were given in addition to the inclusion and exclusion criteria above. PCPs were blinded to results of PM until they submitted their final list of selected patients.
Intervention and outcomes
In the year following patient selection no intervention beside best generally available primary care took place in either of the groups (the CM intervention started by the end of 2010). Data on hospital admissions and mortality were obtained from insurance claims thereby minimizing reporting bias. PCPs were blinded for the assessment of hospitalization rates in 2010.
Statistical analyses were carried out with R version 2.15.1 . Continuous data were summarized using medians and interquartile ranges (1st and 3rd quartiles) and categorical data using frequency counts and percentages. Because of the hierarchical structure of the data, multilevel analysis was applied that took into account the dependence between patient outcomes (level 1) within primary care practices (level 2). The number of hospitalizations was analysed by a multilevel Poisson regression model , with ‘Group’ (not selected, PM, PCP, both) as a categorical predictor, and random intercepts accounting for overdispersion due to differences across practices. In longitudinal data analyses, an additional covariate ‘Year’ was used in the regression model, and an additional random intercept accounted for the correlated outcomes within patients. Results are reported as risk ratios (RR) with 95% Wald-type confidence intervals (CI). Some of the patients died within 2010. Since the exact date of death was not included in the data we did not adjust for reduced exposition in these patients. However, in a sensitivity analysis, we included an offset variable in the Poisson regression for patients who died in 2010, thereby assuming that these patients survived 6 months (until 30 June 2010). Results of this sensitivity analysis are reported when they differ from the main analysis.
Mortality rates were investigated using a multilevel binomial logistic regression model, with ‘Group’ as a categorical predictor and random intercepts for practices. Results are reported as odds ratios (OR) with 95% CIs. In line with the exploratory nature of this study, the significance level was set to 5% (two-sided), and we performed neither adjustment for multiple testing nor imputation of missing values.