Learning from disease management programs

With the legal modification of risk structure compensation in 2001 (BGBI. I:3465), the basis for the implementation of disease management programs (DMP) (SGB V §137f) was created in Germany. The DMP instrument promised to contribute to improving the quality of care for chronically ill patients (1, 2). The legally required proof of effectiveness and performance savings contain very high methodological requirements and are increasingly controversial (3). Despite all the ambiguities, there are still indications that DMPs can be viewed as an effective strategy for changing the behavior of chronically ill patients and service providers (4). The research approach presented here builds on this finding: Firstly, we assume that the patient's behavioral change triggered by the DMP enrollment has a significant potential for reducing the probability of the occurrence of a secondary disease, and secondly, we assume that with the experience of DMP-guided treatment processes change medical behavior in such a way that patients who are treated in the same practice context without DMP can also benefit from them.

>> The essential characteristics of a DMP are the progression orientation of patient care and the emphasis on coordination between the different service providers who care for chronically ill patient groups (3, 5). Ideally, this leads to the optimization of the course of the disease, observable in the quality of the care processes (e.g. through regular screenings), improved control of the disease (e.g. improved health status) and thereby to a higher quality of life, higher life expectancy, lower costs for the health insurance companies and a stronger orientation of the Providers of evidence-based guidelines. A DMP contains various elements: “… a care path, various interface definitions, care orders, a concept for patient information, a documentation grid and a specification for outcome or evaluation data” (6).
Diabetes mellitus type II (T2D) is a widespread disease with a 5 to 6 percent prevalence in Germany (7). The rate of new cases of T2D is generally considered difficult to determine, as the number of unreported cases of the unrecognized disease is estimated to be very high (7, 8). DMPs are useful tools when the disease they treat is considered to be well researched and is characteristic (3):
• “…with great variation in treatment patterns from patient to patient and from doctor to doctor…” and beyond
• “… the high costs of the disease, an increase in avoidable complications and low patient compliance …”

Therefore, the clinical picture T2D has been selected into everyday outpatient care since the implementation of the DMP in 2002 with the aim of improving the quality of the treatment processes.
International and German diabetes DMP effectiveness studies focus on various dimensions of treatment quality. In addition to evaluating the structure and process of the treatments, measuring the quality of the results represents a very high level of quality assurance (7). To operationalize the care process, regularity of screenings (9, 10), reduction in hospitalization rates (11) and behavioral changes in patient empowerment, satisfaction, well-being and distress (12) are tested. Doctors' adherence, i.e. their willingness to follow EBM guidelines (13), is also the subject of DMP studies. A central criterion for the effectiveness of DMPs is the cost reduction (14-16) that can be achieved through the intended network effects. Laboratory values (such as HbA1c) (9, 10, 17), body mass index (12) or smoking status (15) are usually defined as surrogate parameters for the quality of the results. This is obvious because “Surrogate parameters (…) can naturally be normalized more quickly than the long-term consequences (e.g. cardiovascular events) of inadequate care.” (6).
There are indications in American meta-analyses that care processes and the control of chronic diseases can be improved by DMP, but it does exist
• “…no conclusive support for its effect on health outcomes…” and
• “… no conclusive evidence that disease management leads to a net reduction in direct medical costs” (18).
• Likewise, “Effects on mortality […] remain unclear and little systematic analysis was performed on the cost-effectiveness of integrated care programs” (19).

In recent German studies on the effectiveness of diabetes DMP, the main quality goals set were to improve disease control (20, 21) and the care process (21-23). The results largely show positive effects of the DMP on the desired goals, with the exception of one study by the GEK, the database of which was telephone interviews (22). There was no effect of DMP on HbA1c. To date, there are only a few reliable evaluations regarding the goals of improving outcomes (preventing long-term consequences or early death) and reducing costs in diabetic care. Exceptions here are the studies by Barmer GEK (23) and by Stock et al. 2010 (24). Both studies showed positive effects on long-term consequences, mortality and reduction in the cost of care.
The treatment process guided by a DMP necessarily involves the behavior of doctors and patients. Uncertainties regarding the quality of results that can be achieved through the use of DMPs could be addressed by specifying distinguishable intervention elements of a DMP in relation to the interaction partners (treating doctors vs. patients) (13). In particular, the German interpretation of DMP implementation, in which the family doctor was conceived as the central figure in the management of a patient's DMP (24), requires research designs that shed light on the learning process of patient and doctor.
This study therefore examines both the direct effect of DMP enrollment and the practice-specific context dependency of the probability of late sequelae following an initial diagnosis of T2D. The patient's context is a practice with below-average, average or above-average proportions of DMP patients among all T2D patients. With the varying proportion of DMP per practice, the treating doctor is assumed to have varying degrees of experience in using this instrument. The research question is therefore: To what extent does the proportion of diabetics (type 2) enrolled in DMP within a practice have an indirect effect on the occurrence of late diabetological sequelae as an indicator of an outcome due to optimal care?
Our study results show that, firstly, the evaluation methodology commonly used in DMP studies can lead to a systematic underestimation of their effectiveness because indirect medical learning effects can also be related to patients who are not enrolled in a DMP. In addition to this methodological implication, the recommendation to the cost bearers to further advance the implementation of the DMP can be derived from the direct effect of the DMP in reducing late effects and the indirect learning effect for the treating physicians.
In the first part of this paper we explain the database used for this study, show the variable construction and the methodological approach. The research results presented in the second part suggest both direct and indirect effects of the DMP on the manifestation of the “late sequelae” outcome. In the third part, the results are classified in the discussion and, fourthly, their significance for, firstly, the practice of health services research and, secondly, for future DMP implementation is concluded.
Data, variables and methods
Origin of the data
This research is based on data from the IMS Health Disease Analyzer patient database. It is “one of the largest pharmaceutical epidemiological databases in Europe” (25, 26).
Based on the population of practicing general practitioners, internists and specialists in Germany known from data from the German Medical Association, the sampling plan is updated annually by IMS Health. According to the stratification criteria set by the company (according to specialist group, region, community size and doctor's age), those medical practices that are prepared to provide encrypted information on diagnoses, prescriptions, risk factors and laboratory values on a monthly basis via standardized interfaces are then acquired (27). As an incentive for the doctors, expense allowances are paid and a regular evaluation of their own prescribing behavior compared to that of other doctors is offered.
Through the strategies described above, the data set can be considered representative of the medical population. The comparison with the medical statistics from the German Medical Association confirms this (26). Comparability can also be assumed for the distribution of diagnoses, as this was shown by the similarity of the incidence of some oncological diseases in 2004, recorded by the Robert Koch Institute, with the data from the “Disease Analyzer”. With regard to the diabetic patient type that is the focus of this study, the age structure of a GEK data set from 2005 roughly corresponds to that of the “Disease Analyzer”. However, people aged 70 and older are somewhat underestimated in the “Disease Analyzer”. There is also convincing representativeness with regard to prescriptions of antidiabetic drugs in comparison with other sources (26).
Typical questions about health services research such as pharmaceutical epidemiology (28-30), health economics (31-33), questions of compliance (34, 35) and drug safety (36) as well as the prescribing behavior of doctors (35, 37) were, among others examined in recent years with the help of the “Disease Analyzer”.

Variables and inclusion criteria
The following description of patient inclusion for this study is graphically shown in a flow chart (Fig. 1).
The focus of the present analysis is on diabetics (type II) who were treated in practices run by general practitioners, practitioners or internists, including diabetology-focused practices. A patient was always included in the study as a diabetic (type II) if the coding in the practice corresponded to E 11 to E 14. This procedure is based on the ICD coding guidelines according to DIMDI (37). The complications listed as subcategories under E11 to E14 are considered long-term consequences. The subcategories .0 (coma) and .1 (ketoacidosis) as well as .9 (no complication) were excluded as late consequences.
The entry into force of the reform of the risk structure equalization between the health insurance companies (RSA) with the introduction of structured treatment programs (DMP) (SGB V; §137f) from January 1st, 2002 determined the start of the period to be observed on January 1st, 2002 - the start The T2D-DMP was only in the summer of 2003. The initial diagnoses were therefore partly before the DMP registration. Despite the decreasing economic importance of DMPs for health insurance companies since January 2009, they can still be used (38), so that the observation end of October 2009 is plausible. The morbidity-oriented RSA came into force on January 1, 2009 and reduced the economic attractiveness of the provision of the DMP from the perspective of health insurance companies. Despite their weakening, these can still be provided. Financial compensation for the health insurance companies per enrolled patient is also still possible.
For the period January 2002 to October 2009, those patients who continuously provided data for at least nine months were selected from 1,102 practices. In the next step, only those who could be observed continuously for at least five years and appeared in practice at least two months of the year were included (Figure). The selection of the 5-year panel fidelity is based on the consideration that the T2D patients developed a late sequelae on average after 2.5 years with a standard deviation of 1.9 years.
To confirm the initial diagnosis as the beginning of a T2D “career”, the diagnoses marked as confirmed were used. In addition, it was checked that the patient had been recorded in the practice within six months before the initial diagnosis of T2D without being diagnosed with T2D.
The next selection criterion was the long-term blood sugar “HbA1c”, which is central to diabetic care. Since the mean HbA1c value was only available for 3,172 patients, the panel selected so far was reduced to this number of patients.
The final selection resulted in a sample of 3,072 T2D patients in 101 practices. The gender of the patients is evenly distributed among approximately 50 %. The average age is just under 64 years. The mean long-term blood glucose value is 6.7% with a standard deviation of 1.1%.
Of the 101 practices, 23 have a treating doctor with further training in diabetology. The doctors, who are on average around 53 years old, have 6.7 years of previous experience in hospitals with a wide range of 4.7 years.
Almost 40% of the patients (n=1,196) are recorded as registered diabetological DMP patients through their treating practice. When enrolling individually, only DMPs that were started before the onset of the long-term sequelae were taken into account so that the secondary preventive effect can be measured.
In order to measure an effect of DMP enrollment, it is important that the two patient groups are structurally similar. This can be largely confirmed in the descriptive analysis: the gender distribution, the patient age, the time that has passed since the initial diagnosis, the proportion of patients in urban regions and treated by diabetologically trained doctors are very similar in both groups (Table 1). An interesting indication of direct positive effects of DMP enrollment in the treatment processes of diabetics (type 2) is, firstly, the comparatively lower proportion of late complications in the DMP group (16% vs. 21% in the non-DMP patients) and, secondly, the very high proportion longer time until late effects occur (3.5 years vs. 1.6 years in non-DMP patients).
The presumably systematic underestimation of the proportion of late complications among all diabetics of 19 % - there are hardly any studies in the literature for Germany that enable data comparison (38, 39) - is due to the setting of the data source: they are not ophthalmologists, nephrologists or surgeons Here the first diagnoses of a long-term consequence are general practitioners, internists with further training in diabetology in outpatient care.
In addition, long-term consequences such as myocardial infarction (I-21) or stroke (I-64) are not listed here due to the selection of diabetics via Chapter IX of the ICD code.
The proportion of DMP per practice was determined from the data set after the practice selection and before the patient selection. Treatment processes supported by DMP are seen here as a learning impulse for the doctor. This means that all recorded T2D patients should be included in this variable, regardless of their panel fidelity or diagnostic certainty.
The average proportion of individual DMP registrations in all treatment processes within a practice is 33 %. Of the 101 practices identified, 18 can be identified in whose T2D treatment processes there was never a DMP registration. This corresponds to 18 % of the practices and affects 16 % of the individual treatment processes. The complete lack of just one registration suggests a negative attitude on the part of the practice owner. Therefore, it would only make limited sense to integrate this type of practice into the model, because a learning effect through DMP cannot be assumed here. Therefore, in order to construct DMP proportion variables, it is also necessary to reduce the treatment processes to those in whose treatment context at least a minimum of DMP-guided treatments took place.
The DMP proportions within the medical practices that are basically DMP-ready are on average 41 % and do not follow a specific distribution. With the help of the calculated standard deviation of 21 %, dummies can now be constructed for the categorical characteristics of a below-average, average, above-average and extremely above-average proportion of DMP within a practice (Fig.). With these DMP share variables, different contexts are now operationalized as characteristics of the individual treatment processes:
• No DMP enrollments at all in T2D treatments: not a single treatment process is accompanied by a DMP. The doctor cannot and possibly does not want to learn about DMP criteria (478 cases, corresponding to 16 %). This context group is not included in the model because it cannot be learned from DMP.
• Below-average proportion of DMP-T2D patients in treating practice: treatment context leads to the doctor's first experiences with controlled T2D treatment management by DMP (500 cases, equivalent to 16 %)
• Average proportion of DMP-T2D patients in treating practice: treatment context with regular contact with DMP-guided treatment processes (611 cases, equivalent to 20 %)
• Above-average proportion of DMP-T2D patients in treating practice: treatment context with extensive experience of the doctor that is strongly focused on this instrument (1079 cases, equivalent to 35 %)
• Almost exclusive DMP enrollment in T2D treatments: treatment context with treatment processes dominated by DMP (404 cases, corresponding to 13 %). This context group is not included in the model interpretation because there are no longer any different contexts.

methods
The time until late effects occur differs in the two patient groups. In principle, one could therefore use the analysis method of Cox regression. However, the Kaplan-Meier estimation carried out as a test showed us that the effect of DMP enrollment is not constant over time (Figure) and therefore a basic requirement for this type of event data analysis is not met (39-41). About four years after the initial diagnosis of T2D, the time-delaying effect of DMP wears off, and the long-term effects then appear earlier compared to non-DMP patients. It can therefore be assumed that the effect of the DMPs is particularly effective in the initial period after the initial diagnosis. Therefore, we use a logistic regression model because the dependent variable “status” is a binary coded dummy variable. Because time is of considerable importance for the probability of late sequelae occurring, we include the time variable (year of initial diagnosis) in the logistic regression model in addition to the other control variables.
The likelihood ratio test was carried out to determine the quality of the model fit, the deviation was checked and the Hosmer-Lemeshow test was carried out to assess the classification results (for the procedure see Backhaus et al. (2006: 445-455); Allison (2005: 31-48). The results shown in the table consistently show an appropriate model fit. The adjusted r² is in the acceptable range at 0.22 (42), although a contribution to the quality of the overall model could certainly be made by adding better explanatory variables. Testing of the individual covariates using the Wald statistic showed no overfitting due to superfluous variables. The test for multicollinearity showed satisfactory tolerance values for all included variables.
Results
The results (Model II) show that enrollment in a DMP means an approximately 40 % lower chance of a late sequela being diagnosed for the individual patient. This effect increases to a 55% lower chance of a late consequence occurring when the model is expanded to include the practice's own proportions of DMP patients.
In practices with fewer than average DMP patients, the probability of all patients treated there of having a long-term sequelae drops by almost 70 %. In practices working with DMP patients on average there is also a reduction in this probability, but only 57 %. This effect is comparable to the reduction in the chance of late effects occurring when treated in practices with an above-average number of DMP patients. With the binary variables of the practice characteristic “DMP proportion among T2D patients” the explained variance (r²=0.22) of the overall model (Model III) improves.
As the patient ages, the chance of long-term complications increases slightly; as long-term blood sugar increases (HbA1c in %), it also increases, as expected, by 33%. Physician-specific factors such as the doctor's age at the time of initial diagnosis (increases the chance of a late sequelae) and years of experience in the hospital (reduces the chance of a late sequelae) play a rather moderate role. The care of a patient by a doctor with further training in diabetology increases the patient's chance of developing a long-term sequelae by 61 %.
discussion
Disease management programs are extremely complex managed care instruments with different elements. Despite the strong generalization made here, an attempt is made to do justice to these programs by including aspects from several levels of analysis.
Another limitation is the focus on the diagnosis of diabetes mellitus (type 2). It was not possible to integrate all comorbidities into the model for methodological reasons. Nevertheless, the widespread association of T2D with other diseases is a potentially confounding factor, as can be illustrated by the example of the difficult training of a diabetic patient who also suffers from depression (43). To make matters worse, the underlying form of data collection in outpatient settings can lead to the underestimation of serious long-term consequences caused by T2D. Serious complications (such as myocardial infarctions or strokes) are more likely to be recorded as inpatients.
The relatively reinforcing effect of treating a T2D patient by a doctor with further training in diabetology also needs to be discussed. This result can be plausibly justified by a selection bias, since diabetologists generally treat cases with more difficult initial conditions. The apparently increasing effect of the higher degree of urbanity on the onset of long-term consequences can be assessed in a similar way. Diabetics in larger cities have a 93% higher chance of having a long-term sequelae from the diagnosed T2D. The cause here could be better infrastructural provision in urban regions, which goes hand in hand with more technological possibilities. This means that certain diagnoses (e.g. fundus measurements) are more likely.
Over the period from 2003 to 2009, the probability of having a long-term episode decreases, which is probably due to the unavoidable right-wing censorship. The later the disease starts, the shorter the time period in which long-term effects can occur. In addition, effects of improving T2D care cannot be completely ruled out. It is difficult to causally separate these.
Despite all the limiting factors of this study that are worth considering, the following results can be noted due to the large number of cases, the long time series and the underlying representativeness:
The direct positive effect of DMP enrollment on the onset of late effects measured in this study confirms the results of studies that examined the effectiveness of DMPs (20, 21, 23, 24). Previous findings that pointed to the HbA1c value as a significant predictor for the development of a micro- or macrovascular long-term sequelae were also confirmed (44-46).
The non-linear, indirect effect of the proportion of DMP-guided treatments of diabetics (type 2) on all diabetological treatment processes within a doctor's practice has proven to be very stable in several model variations we have carried out. This sheds light on doctor behavior as an element in T2D care processes. This indirect effect implies that there is the meaningful aspect of physician learning that goes beyond the direct isolated physician interaction with a patient enrolled in a DMP.
conclusion
The implementation of a process innovation, such as a DMP, is subject to complex contexts. This impacts the development of research designs to evaluate the DMP. Therefore, questions remain regarding the research designs commonly used in DMP studies. In most cases, the Randomized Controlled Trials (RCT) known from clinical research are used here. For this purpose, the innovation and control contexts are often assigned and then observed separately. This adoption of RCT logic raises the question of external validity. The reality in German medical practices tends to show a mix of patients who are enrolled in a DMP or not and are cared for by the same doctor. Therefore, in addition to the direct effects of DMPs, their indirect effects should also be of interest in order to avoid systematic underestimation of the effect of DMPs.
While the evaluation situation with regard to the direct effects of DMP-guided treatment processes is generally controversial, the results shown here indicate that the use of DMP can be an effective strategy with regard to changing the behavior of patients and doctors. It can be deduced from this that from the perspective of the cost bearers it may be worthwhile in the long term to continue implementing DMP, even if the direct economic incentive is no longer as high with the introduction of the Morbi-RSA in 2009.
DMP can be understood as an impulse for controlled medical reflection within T2D treatment processes. This means that an indirect effect on the quality of care can also be assumed, since changes in the doctor's behavior usually affect all patients. <<