Health care systems worldwide are facing considerable challenges from the increasing number of chronic and multimorbid patients, characterized by complex needs and frequent transitions between care settings [1]. In Switzerland, 2.2 million people report a chronic disease and nearly 20% of the population older than 50 years have multiple chronic diseases (multimorbidity) [2]. Although the trend of progressing multimorbidity is widely recognized [3-6], it is still unclear how best to take care of patients with multimorbidity and which interventions would be effective. For more than two decades, integrated and coordinated care have been developed worldwide [7]. Nevertheless, integrated and coordinated care faces continuing challenges such as scaling-up, implementation, and sustainability difficulties. Additionally, integrated and coordinated care requires development of novel approaches to evaluate and measure patients multimorbidity and complexity. This is key to stratify the targeted population and adapt the intervention to the needs of the patients. Often, such evaluations and measures rely on morbidity indices (eg, Charlson and Elixhauser) or on the number of (self-reported) chronic conditions or comorbidities [8]. Whereas the former were developed in an inpatient setting as predictors of mortality, the latter may not comprehensively reflect the patient’s disease burden and complexity. Despite these limitations, they remain often used because of their relative accessibility and simplicity. In settings where electronic medical (health) records, national disease registries, or data on chronic conditions are unavailable, administrative health insurance claims data represent a potentially useful source of information. In fact, they are increasingly used in health services research, especially to express multimorbidity using pharmacy-based cost groups (PCGs) [9,10]. PCGs, based on use of prescribed drugs rather than on clinical information, were developed as a proxy for morbidity measure [11]. Although the approach has limitations related to underestimation of medicines used, unclaimed, or paid out-of-pocket and thus not present in the data or the assumption that the drug is used exclusively for treating the particular condition [11,12], it allows mapping patient profiles to reflect their morbidity status. As such mapping approaches and comorbidity counts are considered simplistic [13], researchers may consider alternative methods to investigate patient complexity more exhaustively. One such method is cluster analysis, which relies on the idea that many common conditions cluster together in the population in predictable patterns [13]. It has been shown that cluster analysis of real-world data for drug use research can be used for detecting clinically plausible subgroups [14]. Similar approaches of classifications based on multimorbidity patterns have been applied in the literature [14-16], but using PCGs as the multimorbidity indicator for cluster analysis is novel. In that context, the aim of our study is to investigate patient multimorbidity and complexity beyond simple mapping and counts of PCGs, using clustering methodology in claims data of Swiss residents aged ≥50 years.

After discarding individuals with missing information, our data set comprised 18,732 individuals (points). An initial examination of the data set exhibited three large “single” clusters that we extracted prior to the clustering procedure, showing no PCGs, only hypertension PCGs, and only mental disease PCGs, representing 67.9% (n=12,720), 9.7% (n=1813), and 4.1% (n=765) of the population, respectively. Clustering analyses, performed on the remaining 3434 patients not included in the latter “single” clusters, identified four distinct clusters: Cluster 0 to Cluster 3, numbered in the order in which they are detected while applying HDBSCAN (Figure 1). The clusters can be clearly visualized from this tree (Figure 2); and a good persistence of 0.29, 0.24, 0.15, and 0.24, respectively, was found. The average OOB misclassification error from the 10 RFs was 0.51, which is quite high, showing that RF does not differentiate well between the original and the synthetic data, and there is not much structure in the data. Regarding the performed MDS, the normalized stress was 0.31, indicating reasonable preserving of the distances between points.

The 4 detected clusters encompass different mixes of PCGs (Table 1 and Figure 3): Cluster 0 comprises a large mix of PCGs (mental + hypertension + pain + asthma [chronic obstructive pulmonary disease]) often appearing jointly; Cluster 1 comprises PCGs (thyroid, hypertension, glaucoma, and mix of others) appearing jointly less often; Cluster 2 comprises asthma, Parkinson, cardiac diseases, and pain rarely appearing jointly; and Cluster 3 comprises a large mix of PCGs almost never appearing jointly (single diseases).

The following description and interpretation of clusters is based on the descriptive statistics of health care use and costs data (Table 1), which help to understand the underlying principle of grouping individuals into PCG clusters. First, the members of Cluster 0 (n=817, 4.4%) had the highest number of PCGs and highest costs and health care use, and were referred to as “complex high-cost high-need patients” (for a detailed description, see Table 1). The degree of complexity in these settings was reflected as the combination of the following characteristics interpreted from descriptive statistics (Table 1): average number of PCGs, percentage of multimorbid patients, levels of health care use (eg, number of doctor consultations and hospital stays), and costs in the population subgroup. The members of Cluster 1 (n=709, 3.8%), although having multiple PCGs, had health care costs and use lower than in Cluster 0; thus, they were referred to as “slightly complex with inexpensive low-severity PCGs.” The members of Cluster 2 (n=531, 2.8%) were of the oldest age and presented especially high use of hospitalizations and visits to the generalist doctor and, thus, were referred to as “oldest at high risk.” High risk, interpreted in these settings from the descriptive statistics, was reflected by relatively high use of hospital care, yet lower than in the most complex cluster: long length of stay (5.6 and 6.6 nights for clusters “Oldest at risk” and “Complex high-cost high-need,” respectively) and high inpatient costs (CHF 2749 [US $2950] and CHF 3109 [US $3333], respectively). The members of Cluster 3 (n=1056, 5.6%) were characterized by a relatively small number of PCGs (close to 1) and the highest costs of medications and, thus, were referred to as “patients with 1 costly disease.”

Our study shows that performing cluster analysis to explore patient multimorbidity and complexity is feasible. We demonstrated that individuals with single PCGs of mental diseases or hypertension, individuals with multiple PCGs, or individuals with a single high-cost PCG have different health care use patterns and represent different complexity groups.

Earlier studies focusing on chronic conditions identified from electronic health records evidenced the existence of systematic associations between chronic diseases, whereby chronic diseases, often from dissimilar disease categories, coappeared within a multimorbidity pattern or cluster [24-26]. Importantly, though, these studies showed that the complexity of multimorbidity patterns in terms of diseases and associated drug use increased with age, which holds true for both genders. Moreover, in line with our findings, multiple earlier studies used cluster analysis for identifying clinically homogenous multimorbidity patterns in the population, where clusters were composed of diagnosis-related groups [16,27-30]. However, these studies used measures of multimorbidity and comorbidity or clinical diagnosis data rather than PCGs from claims data. This makes direct comparison of results challenging, due to the differences in methodologies and level of diagnosis details. A recent systematic review confirmed that analytical methods used to identify patient profiles with multimorbid conditions are heterogeneous (including factor analysis, multiple correspondence analysis, hierarchical clustering, and three-step unified-clustering method), which may explain the variation in the multimorbidity patterns reported in various studies [31]. Despite those differences, the observed most prevalent clusters or groups are similar across studies and included hypertensive or metabolic diseases [28,29] and mental and behavioral diseases [16]. The greater prevalence of and similarities in metabolic and mental clusters were confirmed by a systematic review of multimorbidity patterns, whereby these clusters were identified in 9 and 10 of 14 reviewed articles, respectively [32]. One study compared multimorbidity patterns between populations of two European countries (Spain and the Netherlands) and found that, indeed, the highest similarities were observed in the cardio metabolic cluster, even though the populations are likely to differ across countries [26].

The existing literature on the use of cluster analysis to identify homogenous segments based on health care use and expenditures is limited [33-37]. Specifically, the study by Nnoaham and Cann [33] identified segments (or clusters), similar to ours, based on health care use (expressed by visits to the physicians, medications, and admissions) and complexity (expressed by long-term conditions). Other studies used cluster analysis to identify groups with high expenditures and deduced that, despite having a lot of heterogeneity, the high expenditures cluster typically exhibited fair or poor health with more medical conditions or comorbidities [34,35]. These findings confirm ours; they nevertheless need to be interpreted with caution due to differences in methodologies, age of the population, and level of details available for background individual characteristics and diagnoses. There is evidence that cluster analysis may provide more information to decision makers than a list of possible statistically significant variables or a list of individuals who are the highest users [35].

To our knowledge, this is the first study using cluster analysis to explore patients’ multimorbidity and complexity, reflected by the mix of PCGs and health care use patterns. In addition, it benefits from the richness of health care use data, a large sample size, and advanced clustering methods. However, the study has certain limitations. The first limitation stems from the process of multiple parameters configuration, which increases complexity while not allowing results validation. Thus, the cluster interpretation has to rely on metrics from the algorithms, descriptive statistics, and clinical relevance. Second, as the data were lacking clinical information, we only relied on PCGs mapping, which may give an incomplete picture of drug data [9,11,12].

Our study shows that PCG-based cluster analysis of health care use claims data allows diverting from an approach of simple comorbidity counts and can identify the population profiles with increased health care use and costs. Such results may provide insightful information for policy making, care planning, and care delivery to facilitate the transformation from procedures and guidelines focusing on a single disease toward development of integrated and better coordinated care.