Multimorbidity patterns of and use of health services by Swedish 85-year-olds: an exploratory study

All eligible inhabitants were individuals born in 1922 and residing in Linköping municipality, Sweden (n = 650; 235 men and 415 women). The inclusion period for this study was one year (between March 2007 and March 2008). Postal questionnaires and invitation letters were posted at the beginning of each month during the inclusion period. The letters invited individuals to participate in the study 2 months after their 85^th birthday. In the case of no response to an invitation letter, a reminder was sent after 2 weeks. All responses were sent to the Department of Geriatric Medicine, Linköping University Hospital.

The postal questionnaire included questions on socio-demographics (housing, marital status, living situation, level of education, and previous occupation). Working status, measured by previous occupation, was classified into the following categories: low (blue collar), intermediate (white collar), and high (self-employed or academic profession) [18]. For non-participants, information about housing type was checked using the registered address.

The individuals were asked about their use of assistive technology (wheelchair, walker, crutch, vertically adjustable bed, bath/shower technology, adapted toilet, portable toilet, and gripper) and assistance needed (community assistance, transportation service, personal alarm, and food delivery). To evaluate the individual’s self-rated health, a visual analogue scale (VAS) was used ranging from 0 (worst imaginable health status) to 100 (best imaginable health status) [19]. Finally, the individuals provided information on the presence of chronic diseases.

The patients’ medical records are part of the electronic medical report system of the County Council of Östergötlands containing all health-care records (both inpatient and outpatient data) for all citizens of Linköping and the County of Östergötland. Older medical history was also checked in older paper medical records kept for all individuals at the central hospital archives of Linköping University Hospital. This procedure was performed by an experienced geriatrician. The research team compared the documentation in the medical records with the self-reported information in terms of diseases and drug treatments. The self-reported information was the response to two separate questions in the postal questionnaire, regarding chronic and acute medical conditions/diseases. All diseases/conditions indicated were noted for each patient. A disease or condition was only registered if there was clear documentation of the disease and its treatment, regardless of the patients’ self-reports. The 10^th version of the International Classification of Diseases (ICD-10) was used [20]. The presence of chronic disease was then registered if the disease fulfilled one or more of three criteria: the disease was permanently present, the disease was caused by an irreversible pathological condition, and treatment for the disease required rehabilitation or a long period of care. A predetermined list was made for disease categories: cardiovascular disease, cerebrovascular disease, respiratory disease, musculoskeletal disease, mental disease, neurological disease, digestive disorders, urological disorders, endocrine disease, hematological disorders, autoimmune disease, infection, skin changes and malignancy (solid and blood). In the present study, we chose a prevalence of more than 5% as the criterion for a common morbidity.

Data for the use of health services by each individual were provided by the local health care register. The records included visits to a GP, visits to an ER and hospitalization during 2007.

The SPSS Statistical package (version 20.0) was used for the data analyses. The differences between men and women were assessed using the Student’s t-test for normally distributed continuous variables, Mann–Whitney U-test for non-Gaussian distributed variables, and Pearson Chi-square test for categorical variables. Effect size was calculated using Cohen’s d for Student’s t-test, rank-biserial correlation coefficient r (r_rb) for the Mann–Whitney U-test and Cramér’s phi (φ _c ) for the Pearson Chi-square test.

A hierarchical cluster dendrogram was generated using Yule’s Q as the similarity measure between clusters, with a higher value indicating greater similarity measurement. Yule’s Q correlation matrix was calculated as a transformation of the odds ratio (OR) between two variables from (0 to infinity) to (-1 to +1): Q = (OR – 1) / (1 + OR) [21, 22]. We chose the average linkage between groups for the agglomeration because this method takes into account the cluster structure and is relatively robust [23].

To determine predictors of an ER visit or hospitalization, logistic regression was performed with a forward stepwise method (using a likelihood ratio, with entrance/exit tolerances of 0.05/0.10). Model 1 used all single morbidities as candidate variables. Model 2 substituted cluster scores for single morbidities. Interaction of morbidity clusters was included in Model 3. According to John et al. [13], the effects of multimorbidity patterns are evaluated in the form of cluster scores (a count of all morbidities in one cluster) and their interactions (multiplication of two cluster scores, to determine if the clusters’ effects are dependent on each other). Other candidate variables such as socio-demographic factors, individuals’ needs, self-rated health, and the number of visits to a GP were included for model fitting [10]. Collinearity and correlation were analyzed before model fitting. Marital status and living situations were not included concurrently in the analysis owing to high collinearity (r > 0.6). The Nagelkerke R² (Cox and Snell R² adjusted, range 0–1) was used to estimate the amount of variance in the outcomes explained by the predictors [24].

The local Ethical Committee approved the study (Dnr141-06, Linköping), and written informed consent was obtained from all participants and/or their relatives. All participants were informed that taking part in the project was voluntary and participation could be terminated at any moment.

Twelve individuals died before completing the questionnaire and 52 (8%) individuals did not respond to the invitation letter even after the reminder. A total of 496 individuals (189 men and 307 women, 76%) completed the questionnaire. No gender difference was found between participants and non-participants (men vs. women, 46 vs. 108, χ² = 3.452, df = 1, p = 0.063). A larger proportion of non-participants (45/154, 30%) than participants (55/496, 11%) lived in sheltered accommodation/nursing homes (χ² = 29.679, df = 1, p < 0.001). Table 1 summarizes baseline characteristics of the participants. More women than men were living by themselves, had lower education, had lower working status, and used more assistive technology and assistance. Despite the statistical significance, the differences correspond to a small effect size. The most frequently used assistive technology—a walker—was related to improving mobility (40% of all participants; men vs. women, 23% vs. 52%). Food delivery was the only item of assistance reported by few elderly (men vs. women, 10/189 vs. 26/307, χ² = 1.64, df = 1, p = 0.2). The elderly perceived themselves to be in general good health (score of self-rated health ≥ 60) and men were even more positive than women in this study. During the observation year, over three-quarters of the elderly (men vs. women, 138 vs. 242, χ² = 2.205, df = 1, p = 0.138) had visited a GP, but less than one-third had visited an ER or been hospitalized. In absolute numbers, almost twice as many women as men visited an ER or had been hospitalized.

Table 2 gives the rates of most common morbidities according to gender. The significant differences were the greater proportions of men with myocardial infarction and malignancy and the greater proportions of women suffering urinary incontinence, affective disease, dementia and osteoporosis.

Using the measure of similarity (Yule’s Q) and the cluster algorithm (average linkage between groups), we found a large decline in agglomerative coefficients between 0.2 and 0.3, indicating an increase in heterogeneity between clusters. A cut-off in this range of coefficients provided three–five clusters for men (Figure 1) and four–six clusters for women (Figure 2). A higher cut-off coefficient resulted in several smaller clusters whereas a lower cut-off coefficient provided larger clusters. We evaluated that a five-cluster structure identifies most clinically meaningful multimorbidity for both genders. To show the magnitude of similarity between clusters/variables, we took Cluster 1 as an example and read off the distance for each node in Cluster 1 in the dendrograms.

In the men’s dendrogram, Cluster 1 was identified as a vascular cluster. Heart and pulmonary conditions were structured in Cluster 2 (cardiopulmonary) and Cluster 4 (cardiac). Two clusters were related to aging: Cluster 3 containing a somatic–mental combination and Cluster 5 aggregating malignancy with osteoarthritis.

In the women’s dendrogram, the vascular cluster (Cluster 1) was similar to that in the men’s dendrogram but included hyperlipidemia. The cardiopulmonary cluster (Cluster 3) was larger than that of men; myocardial infarction, arrhythmia, and heart failure were connected, and chronic obstructive pulmonary disease (COPD)/asthma was associated with osteoporosis. There were combinations related to aging in Cluster 2 where urinary incontinence was combined with osteoarthritis and in Cluster 4 where malignancy and thyroid dysfunction were merged. Finally, a mental disease cluster (Cluster 5) comprised dementia and affective disorders.

As illustrated in Tables 3 and 4, single-morbidity models (Model 1) explained more variance than did morbidity-cluster models (Model 2). Heart failure was the most significant factor associated with ER visits for both men and women (Model 1). The men’s cardiac cluster (Cluster 4) and women’s cardiopulmonary cluster (Cluster 3) led to an increased likelihood of an ER visit (Model 2). Model 3 for cluster interaction was not constructed, because there was no significant cluster interaction.

Morbidity clustering (Model 2) and cluster interactions (Model 3) explained more variance than the single-morbidity model (Model 1) (Tables 5 and 6).

No single morbidity was significantly related to men’s hospitalization. The cardiac cluster (Cluster 4) and its combination with the cardiopulmonary cluster (Cluster 2) were significant with respect to hospitalization. The variance had an overall increase of 3.2% from Model 1 to Model 3.

For women, heart failure was positively associated with hospitalization and urinary incontinence had an inverse association. The clusters (Cluster 2 and 3) including these two morbidities appeared in Model 2. In Model 3, the cardiopulmonary cluster (Cluster 3) had a stronger effect than that in Model 2. However, Cluster 2 dampened the effect via a cluster interaction with Cluster 3.

Beyond the statistical results from cluster analysis, some patterns of multimorbidity are expected and supported with findings from other studies. First, in Cluster 1, all morbidities shared the common pathophysiological mechanism of vascular disorders except diabetes. However, we still have good reason to believe that in the long run very old people who have complications associated with diabetes have other co-occurring vascular morbidities. Similar findings were also reported by previous studies using cluster analysis [13, 15]. Second, the cardiopulmonary cluster is another expected cluster. Heart failure in the men’s cardiopulmonary cluster was only related to COPD/asthma. The cluster was closer to vascular diseases (Cluster 1) than the cardiac cluster (Cluster 4). The women’s cardiopulmonary cluster contained all heart conditions as well as COPD/asthma. COPD/asthma was first linked to osteoporosis, suggesting osteoporosis was a consequence of long-term treatment of corticosteroids for COPD/asthma patients [25, 26]. This cardiopulmonary pattern was also reported by Marengoni et al. [14] and John et al. [13] but with no gender specificity. A third finding is the clusters of mental diseases. The women’s mental and somatic morbidities were independent of each other. Comparatively, men had a somatic–mental cluster as only affective disorder was included in the analysis. Its association with urinary incontinence was not formally documented in any psychiatric journal according to Vasudev et al. [27] even though the impact of urinary incontinence on mental health has been reported by other researchers [28, 29].

Some morbidities emerged in the same cluster but did not seem to follow pathophysiological pathways such as urinary incontinence and osteoarthritis (Cluster 2 in women). In women, osteoarthritis-related disability may negatively affect urinary control [30]. Another exception is the comorbidity of malignancy. It is difficult to anticipate which comorbidity coexists with a certain type of malignancy, since cancer patients manifest multiple health problems [31]. One reflection from daily clinical practice is that patients with a malignancy diagnosis usually have received complete clinical and laboratory examinations, and therefore, some comorbidity such as osteoarthritis and thyroid dysfunction would not be missed. Another hypothesis is based on the selection of survivals of concurrent ailments. Among cancer patients, some co-occurrences (e.g., severe heart disease) are more likely than others (e.g., osteoarthritis) to cause a high risk of mortality.

Patients using ER services are heterogeneous with respect to the medical services they require. The slightly lower R² in the morbidity cluster models reveals that the selected morbidity cluster (men’s cardiac cluster and women’s cardiopulmonary cluster) did not improve explained variance. In other words, single-morbidity models are more precise in predicting ER visits. A reflection of real clinical practice is that a single morbidity (e.g., heart failure) as a medical condition may already be enough for an ER visit. Unexpectedly, several common morbidities such as COPD/asthma, stroke and even myocardial infarct were not significantly related to ER visits in this study population. Seemingly, in this very old population, these diagnoses were not clearly related to exacerbations or new attacks, but more possibly suggested permanent chronic conditions in patients’ medical records.

In terms of hospitalization, our results are consistent with those of other studies that multimorbid patients were more likely to be hospitalized [32, 33]. The advantage of our approach is that morbidity cluster and cluster interaction models provide more information. Unlike the counts of morbidities, where all morbidities are equally scored irrespective of their inner relationship, morbidity cluster and cluster interaction models address what morbidity cluster was the leading cause of hospitalization. For both men and women, the cardiac and pulmonary condition was a major factor associated with hospitalization. For women, urinary incontinence and its comorbidity with osteoarthritis suggests that old women with certain conditions might be treated using care services other than hospitalization (e.g., primary care).

There is no consensus about how to best measure multimorbidity. According to the theory that the associations among morbidities must be involved when comorbidity rates exceed those that are statistically expected (coincidental) [34], hierarchical cluster analysis helps identify the conjunction between morbidities in a small population with a high prevalence of multimorbidity. Cluster score and cluster interactions have revealed synergistic effects on associative morbidities [13]. However, we realize that very different results may be obtained from the same data using different hierarchical clustering methods [23]. It is of great importance to relate the statistical results to real-life clinical practice so as to verify the interpretable clusters.

In the logistic regression models, the low R² reminds us that reasons for the use of health services are multifaceted phenomena. According to Andersens’ behavioral model, the use of health services is determined by predisposing characteristics (e.g., demographics, social structure, and health belief), enabling resources (e.g., the number of medical personnel and facilities), or a need for health care (health conditions including mortality, morbidity, and disability) [35]. Even if need is a dominant reason why older people use the ER [36, 37], the measures of need as well as other contextual factors can vary [38]. In the present study, an increased number of GP visits reflected the medical needs of very old peoples. Greater use of assistive technology by older men provides information about their severe physical disability or illness because men are more reluctant than women to use assistive technology that brings them shame, embarrassment, and feelings of victimization [39]. Working status and education were measured separately instead of transforming to socio-economic status. The effect of socio-economic status on the use of health service is not consistent in all studies [37, 40, 41], probably owing to the use of different measures and the different financing of health care.

A number of limitations of the present study should be mentioned. First, some morbidities (e.g., arthritis, anemia, and hip fracture) that were not included in the analysis had higher prevalence in other studies [13‐15, 42]. We cannot draw any gender-specific conclusion in the present study. Heterogeneity among populations needs to be considered. Second, diseases with no treatment and asymptomatic conditions could be missed by self-reported surveys and neglected by doctors when recording a medical history; e.g., anemia and osteoporosis. In particular, among non-participants having a high frequency of living in nursing homes, the extreme underestimation of dementia results in health services not being provided to individuals with cognitive impairment. Third, the financing and organization of health care in Sweden limits the generalizations of the findings as other countries may have different social or health care policy. Different welfare regimes affect the priorities of public resources and address inequality issues relating to the use of health services. Individuals with supplemental private health insurance may use private health services, but in this age group, the consumption of private health care is not common practice and it is not the focus of this paper.