Applying a data-driven population segmentation approach in German claims data

The data analysis was carried out as part of a project which aimed at developing an innovative, data-supported primary care model for a rural district (“Landkreis”) in the south of Germany. Pseudonymised claims data was provided by one big German statutory health insurance fund (AOK Baden-Wuerttemberg) that covered more than 50% of the population in this district. In 2019, the rural district had a total population of about 287,000 (50.43% female), with 279 persons per square kilometre [12]. While the gender distribution was similar to the total population in Germany (50.66% female), the rural district was more densely populated compared to the German average of 233 persons per square kilometre [13]. The cross-sectional study included data on all adults (18 years of age and above) who had AOK insurance for at least one day in 2019 and resided in the rural district. For each patient, the following healthcare utilisation variables were selected: GP visits, specialist ambulatory visits, emergency visits, non-elective inpatient admissions, elective inpatient admissions, and ambulatory hospital visits. Specialist ambulatory visits were divided into general specialist care and specialised specialist care, as providers are considered in different planning schemes for the regional distribution of physicians in Germany depending on their specialisation. In total, there are four different provider groups: GPs, general specialists (e.g., ophthalmologists or gynaecologists), specialised specialists (e.g., anaesthetists or cardiologists) and separate specialists (e.g., pathologists or human geneticists). The fourth provider group, however, was not considered in the analysis as the group of separate specialists is organised on a much larger planning scale than the other provider groups (and immediate accessibility is rather secondary) [14]. Since doctor-patient contacts are covered by a per capita payment which in Germany is usually billed at the first visit in a quarter, ambulatory practice visits were considered in a quarterly period. In addition, healthcare utilisation included both visits in- and outside the defined region. Data on patient characteristics including age, gender, disease patterns, long-term care grade, residential care status and costs were also extracted (Table 1).

The legal requirements for the scientific use of case-related or personal data from the statutory health insurance in Germany were applied in accordance with the Social Code Book (“Sozialgesetzbuch”, SGB). The basis for the use of the pseudonymised claims data in this study is paragraph § 75, Sect. 1, number 1, SGB X (research projects based on social data) and was approved by the responsible supervisory authority. Contractual arrangements were made for the data collection and use after approval. The contract outlined each party’s duties and rights regarding the extent (including data scope) and length (including naming of deletion periods) of use as prescribed in paragraph § 75 SGB X. In compliance with data protection laws, the AOK Baden-Wuerttemberg has provided pseudonymised data by replacing personal identifying information (in this case health insurance number) with a pseudonym.

The variables used for segmentation included age and the seven healthcare utilisation variables described above. Age gives implications on the demand for and utilisation of health and social services [17, 18], whereas the seven healthcare utilisation variables represent the use of different healthcare providers and settings [5]. Both age and five of the utilisation variables have been used by previous studies for a data driven segmentation approach, whose segmentation method was also applied in this study [5, 6, 8]. Furthermore, the number of specialist ambulatory visits was separated into general specialist care and specialised specialist care. The number of ambulatory hospital visits was included to give an additional understanding on the population’s demand for healthcare services. Due to different scales across the variables, all segmentation variables were z-standardised by subtracting the mean of each variable and dividing it by its standard deviation.

For the segmentation of the population a combined approach was selected. First, the hierarchical cluster analysis (Ward’s method) was applied as stopping rules are readily available to determine the optimal number of clusters (k) [19]. Its aim is to minimise the cluster sum of squares and thereby maximise homogeneity within the clusters [20]. However, hierarchical cluster methods are not suitable for large datasets [21]. Therefore, it was followed by a non-hierarchical k-means clustering algorithm with a Euclidean distance which can handle large datasets in an efficient way [22].

In accordance with previous studies, Ward’s method was run on ten random subsets of 3,000 patients to define the optimal k [5, 6, 8]. For each of the subsets, two- to 15-cluster-solutions were compared by calculating the Calinski and Harabasz pseudo-F index [19] and the Duda-Hart Je (2)/Je (1) index F [23]. The Calinski and Harabasz pseudo-F index compare the between-cluster variation to the within-cluster variation, thereby assessing the cluster tightness. The Duda-Hart Je (2)/Je (1) index F uses the ratio of the two within sum of squares to decide whether the current cluster can be further split. High values of the pseudo-F index and the Duda-Hart index, with its corresponding low pseudo T-squared value and high pseudo T-squared values on either side, indicate distinct clustering and hence potential cluster solutions [19, 23, 24]. Across the ten subsets, this method resulted in four- to seven-cluster solutions. For these four k values, the k-means clustering algorithm was applied to the full dataset. Each cluster solution was evaluated by its clinical relevance and interpretability [5, 6], resulting in k = 6 to be the optimal cluster solution. To confirm that the cluster solution did not appear by random chance, a split-sample analysis was conducted [5]. The full dataset was split in two equal-sized subsets which were analysed with the k-means clustering algorithm for six clusters. The segments showed the same results as the ten random subsets. All cluster analyses were performed using R Statistical Software, including the packages cluster, clusterSim and NbClust to conduct the cluster analyses and the package comorbidity to compute the Charlson index [25].

The segments were then analysed to identify significant differences across the segments. For the healthcare utilisation variables, a Kruskal–Wallis test was used, because they did not meet the normality assumptions. For age, a one-way ANOVA test was computed. For the variables gender, residential care status, long-term care grade and disease patterns a chi-square test was calculated.

A pairwise segment comparison was then performed on the variables which differed significantly using Mann–Whitney U tests, Student t-tests, and z-tests. Furthermore, a Bonferroni correction was conducted at the significance level of 0.05 for the pairwise tests to counteract the problems of multiple testing occurring when comparing the segments.

As shown in Table 2, the study included 126,046 (51.85% female) patients with an average age of 50.80 years and a standard deviation (SD) of 19.72. About 4% of the study population changed their health insurance fund throughout the year. The k-means cluster analysis revealed six distinct clusters. For each cluster, a label was chosen that best reflects the insured’s healthcare utilisation in the specific segment. Thus, the following names for the segments were set: “Low overall care use”, “High primary care use”, “High emergency care use”, “High specialist care use”, “High hospital care use” and “High overall care use.” Age and all seven healthcare utilisation variables differed significantly across the six segments with p < 0.001, reflecting the purpose of cluster analysis to maximise the distance between the clustering variables.

Segment one “Low overall care use” included not only most of the study population (42.89%) but also the youngest patients (mean 37.76, SD 13.02), followed by segment three “High emergency care use” (mean 41.24, SD 17.02). The oldest patients were in segment six “High overall care use” (mean 69.01, SD 18.59), making up the smallest segment with only 2.03% of the population. In addition, the differences in the variables gender, residential care status and long-term care grade were also statistically significant with p < 0.001 (Table 2). While only 0.20% of the population in segment one lived in residential care homes, 14.94% of the population in segment six did so. Moreover, more than half of the population in segment six had a long-term care grade (54.40%).

The disease patterns were also found to differ between the segments (Table 3). The most frequent comorbidity listed in the Charlson index in the study population was chronic pulmonary disease (15.45%), followed by diabetes without chronic complications (11.44%) and congestive heart failure (8.34%). The prevalence of most of the comorbidities was lower than population average in segments one and three and higher than population average in the other segments. With 99.36%, most people living with a moderate or severe liver disease were in segment five. For this disease, segment five is the only segment with a prevalence higher than population average.

More than half of the study population (59.47%) had a Charlson index of 0, mostly documented in segments one and three. Only 2.57% of the population had a Charlson index of 5 or higher which mostly appeared in segments two, four, five and six, showing the tendency toward multimorbidity in these segments.

The proportion of healthcare utilisation by each segment is presented in Fig. 1. All utilisation variables differed significantly across the segments with p < 0.001. Pairwise comparison between the segments demonstrated that for all the variables, this difference existed across all other segments or four other segments. Overall, the segments showed distinct characteristics (Fig. 2).

Segment one consisted of patients with a low overall care use. Not only were they younger than the population mean, but they also used less healthcare services in all types of healthcare settings assessed, whereas segments two to four all had an above-average use of certain healthcare settings. Similar to segment one, patients in segment two had a lower healthcare utilisation in almost every type of healthcare setting except for GP practice visits, which correlated with older age and higher prevalence of comorbidities in this segment (Tables 2 and 3). These were the two largest segments with about three quarters of the study population. Together, they accounted for 33.37% of total cost.

The patients in segment three were similar in age to those in segment one and had fewer healthcare provider visits than the average population when it came to specialised care practice visits and ambulatory hospital visits. However, their number of GP practice visits, general specialist care practice visits and elective inpatient admissions exceeded those of segment one. They were characterised by a high number of emergency care visits compared to the population mean (68.68% of all emergency visits) and accounted for 10.29% of total cost.

In segment four, patients were older, the number of GP practice visits, general specialist care practice visits, specialised care practice visits and elective inpatient admissions were higher than the population mean, but lower for emergency visits and non-elective inpatient admissions. This correlated with a higher prevalence of comorbidities. Overall costs of this segment comprised 20.74% of the population total cost.

For segments five and six, healthcare utilisation in all settings was higher than the population mean. In detail, segment five had the highest number of ambulatory hospital visits among all segments with 60.21% of all visits. Segment six is characterised by a very high overall healthcare utilisation, with GP practice visits, non-elective inpatient admissions and elective inpatient admissions being the highest among all segments. Covering only 2.03% of the study population, the segment accounted for 24.04% of the total cost. Both segments five and six showed a high rate of comorbidities.

There are several limitations when interpreting the results of this study. As health insurance claims data is collected routinely for billing purposes, only restricted information was available on the severity of the diseases and further determinants of health, such as the social-economic status. Future research should link data from different data sources to provide more detailed insights into the population needs. Also, this study is limited by using a proxy for the number of physician visits. In Germany, multiple doctor-patient contacts in one quarter are usually billed at the first visit in the quarter, therefore, the number of visits was assessed in a quarterly period underestimating the actual number of visits. Insureds have the opportunity to voluntarily enrol in GP-centred models of care (“Hausarztzentrierte Versorgung”, HZV), which requires them, for instance, to consult their GP first to get a referral for specialist care. The GP is the first point of contact for patients taking the role of a gatekeeper, which may lead to an increased number of GP visits but less uncoordinated encounters with specialists [42, 43]. The data used in this study was collected over a one-year period and might thereby be sensitive to random variation due to influenza waves, for example. Longitudinal data will be needed to give more robust information on the patient’s healthcare utilisation and disease patterns as well as to show patient movements between the respective segments. Another potential limitation is the limited generalisability of the results due to historically determined different structures of insureds in statutory health insurance funds in Germany, e.g., income and educational status may differ between health insurance funds [44]. Thus, including the whole population covered by various health insurance funds in further studies might lead to different segment sizes and segment types. Finally, the district presented in this study is an example of a rural region in Germany. Urban regions may show different healthcare utilisation patterns.