This multidimensional, retrospective, patient registry–based study explored the methods required to transform and synthesize a raw data set into a suitable database for further analysis of descriptive, predictive, and survival statistics to identify the risk factors that might induce MRPs among discharged, polymedicated older inpatients.

The multidimensional patient registry included 140 variables routinely collected during hospital stays by older adult inpatients aged 65 years old or more, living at home before hospitalization, with at least five prescribed medicines at discharge from hospital. The extracted data set was composed of a sample of 20,422 hospitalizations from 2015 to 2018, with similar numbers of annual hospitalizations: 5134, 5095, 5125, and 5068, respectively.

Medicines prescribed before hospital admission were not considered in the analysis due to a lack of data accuracy and validity. Indeed, information on medication at hospital admission is often collected from patients themselves, who may not accurately report their prescriptions, particularly in cases of unplanned hospitalization.

The hospital data set was extracted from a public teaching hospital’s registry, delivered to the investigators in a CSV (.csv) format file via an encrypted email and saved on a secure server. Finally, the data set was imported into the R statistical package for cleaning, data transformation, and synthesis [20]. Routinely collected data included information derived from patients’ medical and clinical statuses (patient-reported data, clinical examination, medical diagnoses, or medicines prescribed). The data set had to be cleaned up and synthesized to be suitable for analyzing descriptive, predictive, and survival statistics.

Clinical coding was carried out directly by health care professionals during routine daily care, using a pre-established drop-down menu. Official clinical coding of established medical (10th revision of the International Statistical Classification of Diseases and Related Health Problems [ICD-10]) and surgical diagnostics (CHOP) is mandatory under Swiss Federal Office of Public Health regulations. The variables represented by free text in the original database were excluded.

The distributions of each variable in the data set were explored, according to type (categorical and continuous variables), in order to identify any extreme values and obtain a better view of missing values and associations. Our data cleaning and transformation were guided by a literature review on cleaning-up large data sets, the quantity of information available to us, and the study aim [21]. One major challenge was to find a way to select or summarize a significant volume of information so that further descriptive and predictive statistical analyses could be performed (ie, summarize as many variables as possible, while losing the least amount of information). The large number of variables describing an inpatient’s somatic and cognitive status and medical diagnoses represents a significant challenge: we must find a balance between the variability of data and the essential, detailed information they provide without losing the ability to perform descriptive, predictive, and survival analyses [22].

Summarizing the data set was especially challenging because most of the variables documented different parts of inpatients’ overall health status, with all the diverse dimensions of their somatic and cognitive conditions. Special attention was given to the large data set of prescribed medicinal treatments. In many fields, the most common means of coping with such difficulties is the use of statistical clustering, a technique which combines all the available information (all variables) to reveal one or several underlying dimensions or health concepts.

In addition, the data set’s large number of variables and dimensions made it extremely complex to investigate the relationships and interactions between the different somatic and cognitive variables. The data set should allow the analysis of the risks of adverse health outcomes and their relationships with the medicines prescribed. For this reason, computing every variable in the same model may not be the optimal modeling choice if we consider the multidimensional aspect and dependency between those variables. This is especially true if these variables are significant (P<.01) for the discrimination and discovery of mechanisms leading to rehospitalization and a nonreturn home due to medical conditions and MRPs. In the absence of any scientific models, this study used an empirical approach.

Our data analysis strategy applied a hierarchical cluster analysis, using the ClustOfVar R package [37,38]. As with any statistical analysis, results of a hierarchical cluster should not be accepted as they first appear, but should be taken as suggestions or questioned instead. When the final set of groups of variables was defined, a statistical model to cluster the individuals within each group was applied. This created one new variable for each group, indicating the type of characteristics the individual displayed in his/her health status assessment. For example, if we separate the individuals into 3 groups according to their cognitive status, we might obtain a variable indicating that a person belongs to a group with significant, minor, or no cognitive impairment. This type of aggregated variable was used in our final analysis of risk factors.

Our analysis explored several different clustering methods. However, the results displayed here most often used the following variable clustering procedure. First, a one-factor analysis model was typically used; second, the most important latent factors were selected. At this stage, it was essential to obtain accurate clustering rather than reduce the dimensionality, which takes place in the final cluster partition. Third, these factors were considered as variables and served as the input to a k-means clustering algorithm. Finally, the number of clusters was then selected using the Rousseeuw silhouette statistic, also with regard to the interpretability of the resulting partition [39].

In this approach, n and p denote the numbers of the patients and health conditions (indicators), respectively. The data can thus be represented by an n × p matrix, where the observed value for the ith column and the jth row of the data matrix is 1 or 0, indicating the presence or absence of the ith health condition for the jth respondent (i = 1,…, p; j = 1,…, n).

In the 2-step clustering approach, step 1 involves clustering the p conditions into non-overlapping groups of clinical or health conditions. Based on individual patterns in these groups of clinical and medical conditions, step 2 involves clustering the n respondents into clusters which correspond to different patterns of clinical or health conditions.

To thoroughly analyze the data and identify the MRPs leading to adverse health outcomes—such as rehospitalization, nonreturn home, and early death [40,41]—among older adult inpatients, a literature review was conducted [27].

As in every real-life data collection exercise, missing values are unavoidable, and it is important to define how these are integrated into the study. Four approaches were considered: ignoring all observations with 1 or more missing values; defining “NA” as a separate potential variable value; replacing every missing value by the mode of the corresponding variable; or performing multiple imputations on the data set. The first approach was obviously inappropriate, especially in cases where the number of missing data was significant (P<.01). Considering NA as a separate modality for each variable inflates the number of modalities, but it reduces the possibility of bias due to incorrect imputation methods. Nevertheless, for the sake of comparison, it was also tempting to consider the 2 latter approaches. Before choosing between simple replacement using the variable’s mode value and multiple imputation, we had to test for the type of missing data. If data are missing completely at random, we can simply impute using the mode. However, if this possibility is rejected, multiple imputation is theoretically more appropriate. The Little test (1988) [42] examines the null hypothesis H0: the data are missing completely at random. This test was applied to all subclusters of variables and the null hypothesis was rejected for every data set. This indicated that multiple imputation could be performed as an optional solution for estimating missing values.

Finally, defining NA values became our primary choice for the treatment of missing values. By creating an NA variable (an empty variable that does not influence the cluster result), all observations with an NA variable were still taken into account in the cluster analyses. This is why each cluster analysis contains every hospitalization (N=20,422).

The hospital data set was coded and its use was contractually limited by the participating hospital center. Furthermore, because the data sets included highly sensitive electronic patient records from a hospital registry, ethical approval was sought before any synthesis or analysis. Data were stored on a dedicated secure data server, which included a log registry. Each access flow to the secure data environment was documented, and each change required approval. Only users working on the project and requiring access to the data were allowed to use the selected multifactor authentication mechanism in the secure environment. The Human Research Ethics Committee of the Canton of Vaud (CER-VD) (2018–02196) approved the study on February 1, 2019.

The original data set required some adjustments before our plan of analysis could move forward. Four empty variables and 1 observation containing mostly 0 or unavailable values were removed from the data set. The labels for all variables were rewritten and clarified, and many medicine names in French had accents and unreadable symbols corrected.

Tests made using both the BaylorEdPsych and RBtest R packages confirmed that the missing-completely-at-random hypothesis could be rejected [42]. Observations within each subcluster of the data set that only contained missing values were recoded as NA. Their presence might have been due to incorrect inputs, human or software error, or unavailable parts of some questionnaires. Missing data had very little impact on the sample size, appeared to be random, and concerned the first 4300 observations, especially. After recoding these observations, the cognitive status variables showed no more separate missing observations, and we had a complete data set.

Most of the hospital variables were partially independent and gathered into several groups according to the dimension of the patient’s measured/assessed clinical and medical status. We used an empirical approach suggested by health care experts (FP, HV, and AvG) in an attempt to present homogenous groups within the set of variables. In cases involving clear and meaningful clustering, we relied on expert recommendations or opinions taken from a comprehensive literature review [27,33]. However, when evidence was scarce, we clustered variables using statistical methods. The results from statistical methods were compared against those from expert opinion, which served as a validation tool for addressing any possible subjectivity in those expert opinions [27,33].

Seven groups of clusters were developed: somatic/physical health conditions (3 orange groups in Figure 1), cognitive health conditions (green textbox in Figure 1), total number of prescribed medications based on the ATC classification, diagnoses based on the ICD-10 (yellow textbox in Figure 1), and the surgical interventions based on CHOP (gray textbox in Figure 1). Besides these more apparent distinctions between variables, other underlying subclusters may be present within these groups. This point is beyond the scope of this paper, however, and will be documented elsewhere with a complementary, within-group analysis (the presence of an interpretable clustering of variables within a group before clustering individuals). An examination of the place of discharge variable confirms this: of 20,422 hospitalizations, only 131 patients (<1%) were documented to have died during hospitalization. Bearing in mind that there was no explicit variable indicating this worst outcome, we developed indicators that were suggestive of imminent death or a highly and irreversibly deteriorated health condition. Based on a literature review of polymorbidity, 6 clinical indicators from the data set were associated with a functional deterioration leading to progressive decline and poor health status [43]: (1) restricted mobility, (2) incapacity to change position, (3) altered alertness, (4) altered orientation, (5) altered gait, and (6) reduced or absent cognitive skills necessary to carry out the activities of daily living. Each of these variables indicated a deteriorating health status. To ensure that only severely deteriorating health problems were captured, we only considered patients to be endangered if they had multiple problems. We therefore created a variable indicating the number of problems present, with values ranging from 0 to 6 (Multimedia Appendix 7). More than half of the sample presented with at least one deteriorated health condition. However, only a small fraction of the older adult patients had 4 or more deteriorated health conditions at discharge.

Multiple variables showed modalities that did not correspond exactly to those described in the list (Multimedia Appendices 1-6). The risk of falling variable in the list of somatic data (orange textbox, Figure 1) is continuous, and it was thus recoded into a 3-modality factor as no risk (0 falls), moderate risk (1-4 falls), and high risk (≥5 falls in the last year).

The number of somatic variables is large and heterogeneous, making the direct clustering of individuals challenging. We considered the hypothesis that there were probably dissimilarities in this whole set of somatic variables, and starting from this assumption, we split the variables into subclusters.

In the absence of any validated techniques, tools, or evidenced-based literature, we developed an empirical subcluster clustering strategy. The initial separation of the variables was guided by information retrieved from a literature review of communicable somatic diseases completed with the authors’ experiences and expertise in patterns of somatic illness [27,28]. Four subclusters of somatic variables were constructed: mobility, health difficulties, capacities for the activities of daily living, and other health risks (orange textbox in Figure 1). The mobility subcluster was composed of the clinical variables of movement, changing position, altered gait, balance disorders, and past and recent falls. The general health status subcluster included exhaustion, hearing, vision, verbal expression, drowsiness, sleep rhythm, sleep impairment, pain intensity, and chronic pain. The capacities for the activities of daily living subcluster were composed of upper- and lower-body care, upper- and lower-body (un)dressing, eating, drinking, and micturition- and defecation-related movements. The other health risks subcluster was composed of clinical variables assessing the risks of sores, wounds, malnutrition, and falling during hospitalization. To reinforce the authors’ opinions, a statistical validation model of the variable clustering was computed using the hierarchical clustering functions of the R ClustOfVar package (Figure 4).

Findings showed some differences between the authors’ opinions and the statistical model. To optimize the composition of somatic health status variable subclusters, an adapted version was selected for further data analysis following discussion and a consensus agreement. Three subclusters of somatic variables were considered. The mobility subcluster was composed of the movement, changing position, and altered gait variables. The general health impairments subcluster included exhaustion, hearing, vision, verbal expression, risk of falling, chronic pain, and pain intensity. The capacities for the activities of daily living subcluster included upper- and lower-body care, upper- and lower-body (un)dressing, eating, drinking, and micturition- and defecation-related movements.

After separating the variables, the somatic health status subclusters of mobility, health impairments, and capacities for the activities of daily living were themselves partitioned, with the aim of discovering any possible underlying groupings of inpatients.

Clustering the large data set with more than 2000 different ICD-10 and 800 different CHOP diagnoses into general clusters was not interpretable. To make it suitable for further analysis, the ICD-10 data set was recoded into 4 groups: physiological systems, mental illnesses, oncological diseases, and others. The CHOP diagnoses were also recoded into 4 groups: physiological systems, sensorial, other, and measurement instruments for diagnostics (Table 6).

The different clustering and recoding methods resulted in the data set presented in Table 7.

This paper describes the rationale and methods used to synthesize a large, routinely collected data set of clinical and medical information concerning polymedicated home-dwelling older adults during hospitalization. The electronic patient records from a hospital center provided a valuable data resource for researchers wishing to perform a variety of analyses to explore health risk determinants, medication prescribing, rehospitalization, and death rates. Prospectively collecting research data is often time-consuming and expensive, resulting in biased samples of highly selected individuals, who are often unrepresentative of real-life patients [21]. Data that are already available for use in anonymized electronic patient records provide a valuable opportunity for a variety of different research designs and are particularly useful in the design of registries for evaluating patient outcomes [44]. In some situations, using population-based registries is even preferable to collecting primary data because selection bias due to nonresponders is not a problem [21]. However, large patient registries are sometimes also inconvenient as they frequently present raw data sets and, for several different reasons, they may not be immediately suitable for performing advanced statistical analyses [22]. Those large data sets usually need to be transformed, cleaned-up, and synthesized to be usable for advanced descriptive and predictive statistical analyses.

Our 4-year population-based data set was composed of polymedicated home-dwelling older inpatients with multiple chronic conditions, hospitalized and perhaps rehospitalized in a hospital center in the French-speaking part of Switzerland. The data came from multiple data set sources and were not easily exploitable for advanced statistical analyses, forcing the research team to explore and develop a synthesizing strategy for a large set of variables so as to respond to our research question. Synthesizing a large number of heterogeneous variables in a finite set of specific medical, clinical, and medication data groups was carried out using the principles of cluster methodologies [30,32] and following Olsen’s recommendations for best practices in the analysis of population-based registries [22]. Most of the variables documenting patients’ health status fulfilled the criteria for clustering into different groups according to the dimensions of their health status. Despite the existence of a large number of clustering algorithms, we observed that clustering variables remains a challenge [37]. First, our data set covered a large number of different domains, and it is often the case that clustering algorithms must be applied to heterogeneous sets of variables, creating an acute need for robust, scalable clustering methods for mixed continuous and categorical-scale data [45]. Current clustering methods for mixed-type data are generally unable to equitably balance the contributions of continuous and categorical variables without strong parametric assumptions. Second, stable cluster analysis is strongly dependent on the data set, especially on how well separated and how homogeneous the clusters are. In the same clustering exercise, some clusters will be more or less stable than others [46]. To overcome this challenge, our study used a combined empirical and statistical approach. In the empirical approach, the variables in the clusters and subclusters were selected following expert opinion (FP, HV, and AvG), presenting the most homogeneous groups possible within the set of variables described in the literature [47]. In the statistical approach, we used the most appropriate clustering methods and compared the results with the experts’ opinions, which served as a validation tool to address any possible subjectivity in those opinions. Both methods were implemented independently and compared. This approach was similar to that used in 2 recent studies exploring frailty and comorbidity patterns [27,28]. Although this study developed 6 clusters based on best practices and the previously mentioned empirical statistical approach, other underlying subclusters could also be present within them. This was also noted in the study by Newcomer et al [48] which used agglomerative hierarchical clustering methods to identify clinically relevant subclusters based on groupings of coexisting conditions in a large sample of hospitalized adults.

This study demonstrated that constructing subclusters should not rely solely on an explicit statement indicating the worst outcome, such as death. Clinical indicators documenting functional deterioration which led to a progressive decline and a poor health status were integrated into the 7 clustered data sets. A recent population-based registry study by Vuik et al [49] confirmed the utility of this kind of approach and concluded that health status could not only be based on sociodemographic characteristics and medical diagnoses such as age or morbidity, but should also consider specific assessments of clinical care and patient function.

The procedure used in this study can be summarized as a 7-step approach to transforming and synthesizing a raw, multidimensional, hospital patient registry data set into an exploitable database:

The strengths of our retrospective registry study lie in its huge sample, allowing us to explore the data’s variability and homogeneity in depth. Clustering data risks reducing their variability and the information that can be extracted from them, and some clinical variables showed a significant number of missing values. This fact raises questions about the accuracy and quality of the clinical data assessed, which would require measures of interrater reliability among the health care professionals inputting data into the registry. However, because this was beyond the study’s aims, we did not explore interrater scores of clinical assessments or health care professionals’ scoring of routinely assessed clinical data.

Another limitation to our study was that the sample was restricted to inpatients aged 65 years or older. Because this retrospective, register-based study was part of a larger project [50] focused on medication management among polymedicated, home-dwelling older adults with multiple chronic conditions, we did not have the ethics committee’s approval to extend our extraction of data from the hospital register to all hospitalized adults. Furthermore, our analysis did not consider medicines prescribed before hospital admissions due to a lack of data accuracy and validity.

Finally, and surprisingly, our hospital data set revealed a low mortality rate. Considering the incidence of death in the region, our database showed that it was limited in its representativeness of mortality. Older inpatients presenting with a severe functional decline or at the end of their life probably left the hospital early to die at home or in a nursing home/intermediate care clinic.

Transforming and synthesizing electronic health records is an intermediate stage in the process of subsequently investigating risk profiles and predictive and survival outcomes. Proceeding to these types of analyses requires that each patient has a personal identifier (PID) for computing survival, predictive risk factors, re-admission rates, unplanned institutionalization, and other clinical outcomes explored in cohort and case–control studies. In addition, survival analysis must be performed up to 18 months after discharge—beyond our data analysis cut-off point. Within the framework of a trajectory analysis of health care, all the longitudinal data on 1 patient should be on the same horizontal line in the spreadsheet used for calculations. To do this, each patient must have a unique code allowing data to be linked across multiple hospitalizations. Risk and predictive analyses could be organized using multiple linear logistic regression models (generalized estimating equation [GEE statistics]).

In this study, the data synthesized to date will enable our research to be completed with additional longitudinal survival analyses. The construction of sequences of hospitalizations and rehospitalizations will allow us to better understand the impact of certain events from a longitudinal perspective. The registry data have some limitations because observations are equally spaced in time and all start from the same point, in 2015. However, this study promises to provide valid and robust results, because, despite the sample period, the next hospitalization may in fact be the best measure of treatment impact. For instance, the consequences of treatment decisions taken during one hospitalization (such as medications prescribed or surgical interventions) might only be measurable when the older inpatient needs to be rehospitalized. Yet those unequal periods between hospitalizations may actually prove to be advantageous because they provide a period of effect—that is, a period selected naturally by the evolving health status specific to each older inpatient (eg, inappropriate treatments make inpatients return to hospital at the exact moment their health worsens). A survival analysis would need to be performed to measure the impact of each important intervention (medical act or medication prescription).

This retrospective registry analysis study delivered a method to transform and synthesize a large, raw data set, which included patients’ health records with sociodemographic, clinical, medical, health status, and medication data. Data were cleaned-up and the most appropriate approach for managing missing values was applied. The multicomponent data synthesis strategy integrated recoding together with empirical and evidence-based statistical clustering methods. Seven clusters were constructed to present the health status of hospitalized older adult inpatients. Medical status, comorbidity, and medication data were recoded to summarize the large data set. Finally, our overall strategy delivered an exploitable, population-based database for the advanced analysis of descriptive, predictive, and survival statistics for older inpatients.