Open access Research Soft clustering using real-world data for BMJ Open: first published as 10.1136/bmjopen-2019-029594 on 30 August 2019. Downloaded from http://bmjopen.bmj.com/ on August 2, 2021 by guest. Protected by copyright. the identification of multimorbidity patterns in an elderly population: cross- sectional study in a Mediterranean population Concepción Violán,‍ ‍ 1,2 Quintí Foguet-Boreu,1,3,4 Sergio Fernández-Bertolín,1,2 Marina Guisado-Clavero,1,2 Margarita Cabrera-Bean,5 Francesc Formiga,6 Jose Maria Valderas,1,7 Albert Roso-Llorach‍ ‍ 1,2 To cite: Violán C, Foguet- Abstract Boreu Q, Fernández-Bertolín S, Strengths and limitations of this study Objectives The aim of this study was to identify, with et al. Soft clustering using soft clustering methods, multimorbidity patterns in the real-world data for the ►► Studies focus on diseases rather than individuals electronic health records of a population ≥65 years, identification of multimorbidity as the unit of analysis in assessing multimorbidity patterns in an elderly and to analyse such patterns in accordance with the patterns (hard clustering forces each individual to population: cross-sectional different prevalence cut-off points applied. Fuzzy cluster belong to a single cluster, whereas soft clustering study in a Mediterranean analysis allows individuals to be linked simultaneously allows elements to be simultaneously classified into population. BMJ Open to multiple clusters and is more consistent with clinical multiple cluster). 2019;9:e029594. doi:10.1136/ experience than other approaches frequently found in ►► Reliable and valid identification of disease clusters bmjopen-2019-029594 the literature. is needed for the development of evidence-based ►► Prepublication history and Design A cross-sectional study was conducted based on clinical practice guidelines and pathways of care additional material for this data from electronic health records. for patients that correspond to the wide spectrum of paper are available online. To Setting 284 primary healthcare centres in Catalonia, diseases in patients with multimorbidity. view these files, please visit Spain (2012). ►► Soft clustering analysis allows for diseases to be the journal online (http://​dx.​doi.​ Participants 916 619 eligible individuals were included linked simultaneously to multiple clusters and is org/​10.​1136/​bmjopen-​2019-​ (women: 57.7%). 029594). more consistent with clinical experience than other Primary and secondary outcome measures We approaches frequently found in the literature. CV and QF-B contributed extracted data on demographics, International ►► The different cut-off points (prevalence filters) ap- equally. Classification of Diseases version 10 chronic diagnoses, plied to obtain multimorbidity patterns permitted the prescribed drugs and socioeconomic status for patients identification of common nuclear diseases which Received 07 February 2019 aged ≥65. Following principal component analysis of remained independent of their prevalence. Revised 24 July 2019 categorical and continuous variables for dimensionality ►► The literature provides support for the aetiopatho- Accepted 29 July 2019 reduction, machine learning techniques were applied for physiological and epidemiological associations be- the identification of disease clusters in a fuzzy c-means tween conditions forming part of the same cluster. analysis. Sensitivity analyses, with different prevalence cut-off points for chronic diseases, were also conducted. Solutions were evaluated from clinical consistency and significance criteria. tailored approaches to multimorbidity management and Results Multimorbidity was present in 93.1%. Eight further research. © Author(s) (or their clusters were identified with a varying number of disease employer(s)) 2019. Re-use permitted under CC BY-NC. No values: nervous and digestive; respiratory, circulatory and Introduction commercial re-use. See rights nervous; circulatory and digestive; mental, nervous and The term multimorbidity widely refers to the and permissions. Published by digestive, female dominant; mental, digestive and blood, existence of numerous medical conditions in BMJ. female oldest-old dominant; nervous, musculoskeletal a single individual.1 In many regions of the For numbered affiliations see and circulatory, female dominant; genitourinary, mental end of article. and musculoskeletal, male dominant; and non-specified, world there is evidence that a substantial, and youngest-old dominant. Nuclear diseases were identified probably growing, proportion of the adult Correspondence to for each cluster independently of the prevalence cut-off population is affected by multiple chronic Dr Concepción Violán; conditions. Moreover, the association of point considered. ​cviolan@​idiapjgol.o​ rg Conclusions Multimorbidity patterns were obtained multimorbidity with increasing age leading Dr Quintí Foguet-Boreu; using fuzzy c-means cluster analysis. They are clinically to a two-fold prevalence in the final decades ​42292qfb@​comb.​cat meaningful clusters which support the development of of life has been proven.2 Multimorbidity has Violán C, et al. BMJ Open 2019;9:e029594. doi:10.1136/bmjopen-2019-029594 1 Open access BMJ Open: first published as 10.1136/bmjopen-2019-029594 on 30 August 2019. Downloaded from http://bmjopen.bmj.com/ on August 2, 2021 by guest. Protected by copyright. been estimated to be at around 62% between 65 and 74 (2012). The Catalan Health Institute provides universal years, and around 81.5% after 85 years.3 Its true extent is, coverage and operates 284 primary healthcare centres however, difficult to gauge as there is no agreed defini- (PHC). tion or classification system.4–7 Most of the published literature focuses on diseases Data sources rather than individuals as the unit of analysis in assessing Since 2006 the Information System for Research in multimorbidity patterns.8 Orienting the analysis of Primary Care (SIDIAP) database includes anonymised multimorbidity patterns at an individual level, and not longitudinal electronic health records from primary of disease, could have crucial implications for patients. and secondary care which gather information on demo- In the current context of limited evidence on interven- graphics, diagnoses, prescriptions and socioeconomic tions for unselected patients with multimorbidity, such an status.18 In our study the inclusion criteria were individ- approach–would allow better understanding of popula- uals aged 65–99 years on 31 December 2011 with at least tion groups, and facilitate the development and imple- one PHC visit since 2012. Only participants who survived mentation of strategies aimed at prevention, diagnosis, until 31 December 2012 (index date) were included in treatment and prognosis. It would also elicit essential the analysis. information for the development of clinical guidelines, pathways of care, and lead to better understanding of the Variables nature and range of the required health services.9 10 Diseases were coded in the SIDIAP using the Interna- Cluster analysis involves assigning individuals so that tional Classification of Diseases version 10. An opera- the items (diseases) in the same cluster are as similar as tional definition of multimorbidity was the simultaneous possible, while individuals belonging to different clusters presence of more than one of the selected 60 chronic are as dissimilar as possible. The identification of clus- diseases previously identified by the Swedish National ters is based on similarity measures and their choice may study of Aging and Care in Kungsholmen (SNAC-K).19 depend on the data or the purpose of the analysis.11 12 Additional variables included in the study were socio- Hard clustering forces each element to belong to a single demographics (age, sex, socioeconomic status (MEDEA cluster, whereas soft clustering (also referred to as fuzzy index20)), clinical variables (including number of chronic clustering) allows elements to be simultaneously classi- diseases and invoiced drugs) and use of health services fied into multiple clusters. (number of visits to family physicians, nurses and emer- Empirical evidence is needed on how both estab- gency services). lished and novel techniques influence the identification of multimorbidity patterns. A recent systematic review Statistical analysis recommended that future epidemiological studies Descriptive statistics were used to summarise overall cover a broad selection of health conditions in order information. Disease prevalence was computed for all to avoid missing potentially key nosological associations the included population. Descriptive analyses were strat- and enhance external validity. When many conditions ified by the presence of multimorbidity. Comparison was are considered, the clustering of individuals based on performed using Student’s t-test or Mann-Whitney test morbidity data will encounter high-dimensional issues. for continuous variables and χ2 test for categorical ones. This is particularly important when a clustering-based In order to obtain the most representative clusters approach is adopted to assess the impact of multimor- all patients were included irrespective of whether they bidity on individual health outcomes and health service presented multimorbidity or not. Sex and age variables, uses.2 8 13–15 together with chronic diseases selected by prevalence, The identification of multimorbidity patterns seems to were included in the analysis. The number of features to be implicitly dependent on the prevalence of the included be considered varied from the 62 original ones (no prev- diseases.2 8 16 17 However, to the best of our knowledge no alence filtering applied) to 54 and 49, for a 1% and 2% previous study has analysed the identification of multi- prevalence threshold, respectively. morbidity patterns explicitly based on the prevalence of Due to the large number of diseases, a principal the diseases. component analysis for categorical and continuous data The aim of this study was to identify, with soft clus- (PCAmix) was implemented to reduce complexity. With tering methods, multimorbidity patterns in the electronic this technique both continuous and dichotomous vari- health records of a population ≥65 years, and to analyse ables were simultaneously processed through the appli- such patterns in accordance with the different prevalence cation of Multiple Correspondence Analysis to the binary cut-off points applied. variables and PCA to the continuous ones. Using Karl- is-Saporta-Spinaki criterion to select the optimal number of dimensions to retain, the data set of 49 features per Methods individual per 2% prevalence cut-off was transformed to Study population a new dimensionally reduced data set of 13 continuous A cross-sectional analysis was carried out in Catalonia features per individual, which concentrated most of the (Spain), a Mediterranean region of 7 515 398 inhabitants variability of the newly transformed data set.21 2 Violán C, et al. BMJ Open 2019;9:e029594. doi:10.1136/bmjopen-2019-029594 Open access BMJ Open: first published as 10.1136/bmjopen-2019-029594 on 30 August 2019. Downloaded from http://bmjopen.bmj.com/ on August 2, 2021 by guest. Protected by copyright. Once the transformed data set was obtained, clusters of chronic conditions at baseline were identified using the fuzzy c-means clustering algorithm.22 This Machine learning technique forces every individual to belong to every cluster in accordance with its characteristics and by assigning a membership degree factor in (0, 1) to each individual with respect to each pattern. This provides the flexibility enabling patients to belong to more than one multimorbidity pattern.23 The main parameters in this clustering procedure were the number of clusters and a fuzziness parameter, denoted m, which ranged from just above 1 to infinity. High m values produce a fuzzy set of clusters, so that individuals are equally distributed across clusters, whereas lower ones generate non-overlapped clusters. Further details on the stability and validation techniques applied to obtain the best fuzzy c-means parameters and the set of centroids are presented in online supplementary additional file 1. Figure 1 Study population flow chart. *See 60 chronic To describe the multimorbidity patterns, frequencies diseases group defined in Swedish National study of Aging and percentages of diseases (P) in each cluster were and Care in Kungsholmen (SNAC-K).25 calculated. Observed/expected ratios (O/E ratios) were calculated by dividing disease prevalence in the cluster by disease prevalence in the overall population. As the Patient and public involvement membership of each individual to any of the clusters was Patients were not involved in the study based on anony- given by a membership degree factor, and not as a binary mised data. variable, the observed disease prevalence (O) in a cluster was computed as the sum of the disease membership degree factors corresponding to all individuals suffering Results the disease. Exclusivity, defined as the proportion of In this study, 916 619 individuals were included (women: patients with the disease included in the cluster over the 57.7%; mean age: 75.4 (SD: 7.4)), and 853 085 (93.1%) of total number of patients with the disease, was also calcu- them met multimorbidity criteria (figure 1). lated. Further details on how these ratios were computed Participants’ characteristics are summarised in table 1. using the membership factors are given in online supple- Statistically significant differences were present between mentary additional file 1. A disease was considered the multimorbidity and non-multimorbidity groups for to be part of a multimorbidity cluster when O/E ratio all the variables included in the analysis (table 1). was ≥2 or exclusivity value ≥25%.24 Clusters names were Among the 60 SNAC-K chronic diseases, the most also defined taking into account the dominant gender prevalent were: hypertension (71.0%), dyslipidaemia or age in the cluster compared with the overall sample (50.9%), osteoarthritis and other degenerative joint distribution. diseases (32.8%), obesity (28.7%), diabetes (25.1%) and We conducted a sensitivity analysis by modifying the anaemia (18.3%) (table 2). prevalence threshold for disease inclusion in the cluster Eight multimorbidity patterns were identified using analysis. For chronic diseases we considered as alterna- fuzzy c-means algorithm with fuzziness parameter of tives no filtering, and ≥1% and ≥2% filters among the m=1.1, after computing different validation indices to included population. In order to conform to the Karl- obtain the optimal number of clusters (online supple- is-Saporta-Spinaki rule, a different number of dimensions mentary additional file 1). This number was the same for of the transformed data set were retained to construct the the three different prevalence thresholds: no filtering, clusters for every prevalence cut-off: 13 dimensions for and ≥1% and ≥2% filters. The cluster formed by the most the 2% prevalence, 14 dimensions for the 1% prevalence prevalent diseases was designated non-specified, young- and 17 dimensions with no filtering. The content of each est-old dominant (O/E ratio <2 and exclusivity <20). The cluster was compared across filtering approaches in terms remaining seven clusters were specific: nervous and diges- of diseases associated with that cluster, characteristics of tive; respiratory, circulatory and nervous; circulatory and diges- the included population and cluster size. Clinical evalua- tive; mental, nervous and digestive, female dominant; mental, tion of the consistency and significance of these solutions digestive and blood, female oldest-old dominant; nervous, was also conducted. musculoskeletal and circulatory, female dominant; and genito- The analyses were carried out using R V.3.3.1 (R Foun- urinary, mental and musculoskeletal, male dominant (table 3). dation for Statistical Computing, Vienna, Austria). The Table 3 shows the results, considering a 2% prevalence significance level was set at 0.05. filter, for each pattern based on the 15 diseases with the higher O/E ratios. Violán C, et al. BMJ Open 2019;9:e029594. doi:10.1136/bmjopen-2019-029594 3 Open access BMJ Open: first published as 10.1136/bmjopen-2019-029594 on 30 August 2019. Downloaded from http://bmjopen.bmj.com/ on August 2, 2021 by guest. Protected by copyright. Table 1 Characteristics of study participants aged 65–94 years stratified by multimorbidity and non-multimorbidity (n=916 619, Catalonia, 2012) Multimorbidity Non-multimorbidity All Variables* (n=853 085) (n=63 534) (n=916 619) Sex, women, n (%) 496 294 (58.2) 32 837 (51.7) 529 131 (57.7) Age, mean (SD) 75.6 (7.4) 73.2 (7.3) 75.4 (7.4) Age (categories), n (%)  (65, 70) 225 514 (26.4) 26 664 (42.0) 252 178 (27.5)  (70, 80) 370 356 (43.4) 24 230 (38.1) 394 586 (43.0)  (80, 90) 224 143 (26.3) 10 601 (16.7) 234 744 (25.6)  ≥90 33 072 (3.9) 2039 (3.2) 35 111 (3.8) MEDEA index†  Q1 130 894 (16.5) 13 897 (23.4) 144 791 (17.0)  Q2 126 537 (16.0) 9894 (16.6) 136 431 (16.0)  Q3 129 246 (16.3) 8976 (15.1) 138 222 (16.2)  Q4 125 322 (15.8) 7666 (12.9) 132 988 (15.6)  Q5 110 916 (14.0) 5967 (10.0) 116 883 (13.7)  Rural 169 190 (21.4) 13 059 (22.0) 182 249 (21.4) Number of chronic diseases, median (IQR) 6.0 (4.0; 8.0) 1.0 (0.0; 1.0) 6.0 (4.0; 8.0) Number of chronic diseases (categories), n (%)  0 0 (0.0) 25 380 (39.9) 25 380 (2.8)  1 0 (0.0) 38 154 (60.1) 38 154 (4.2)  (2, 5) 268 836 (31.5) 0 (0.0) 268 836 (29.3)  (5, 10) 463 709 (54.4) 0 (0.0) 463 709 (50.6)  ≥10 120 540 (14.1) 0 (0.0) 120 540 (13.2) Number of drugs, median (IQR) 5.0 (3.0; 8.0) 0.0 (0.0; 1.0) 5.0 (2.0; 8.0) Number of drugs (categories)  0 72 557 (8.5) 40 811 (64.2) 113 368 (12.4)  1 48 704 (5.7) 8378 (13.2) 57 082 (6.2)  (2, 5) 247 095 (29.0) 11 572 (18.2) 258 667 (28.2)  (5, 10) 360 030 (42.2) 2651 (4.2) 362 681 (39.6)  ≥10 124 699 (14.6) 122 (0.2) 124 821 (13.6) Number of visits, median (IQR) 10.0 (6.0; 17.0) 1.0 (0.0; 4.0) 9.0 (5.0; 16.0) Number of visits 2012 (categories), n (%)  0 24 543 (2.9) 23 402 (36.8) 47 945 (5.2)  1 24 281 (2.8) 9603 (15.1) 33 884 (3.7)  (2, 5) 114 198 (13.4) 16 241 (25.6) 130 439 (14.2)  (5, 10) 239 181 (28.0) 10 168 (16.0) 249 349 (27.2)  ≥10 450 882 (52.9) 4120 (6.5) 455 002 (49.6) *All comparisons between variables in multimorbidity and non-multimorbidity showed p<0.001. †MEDEA index goes from 1 (least deprived) to 5 (most deprived), in this variable n=851 564. Women were more represented than men in almost all The highest O/E ratio and exclusivity value were clusters, from 52.7% for respiratory, circulatory and neuro- observed in nervous and digestive for Parkinson, parkin- logical to 83.6% for mental, nervous and digestive, female sonism and other neurological diseases (17.0% and dominant. The exception was genitourinary, mental and 74.3%, and 15.9% and 69.4%, respectively). The lowest musculoskeletal, male dominant in which men made up values were found in non-specified, youngest-old dominant. 90.9% due to the presence of male reproductive system Clusters 1–3 presented the highest median number of diseases (table 4). visits with circulatory and digestive being associated with the 4 Violán C, et al. BMJ Open 2019;9:e029594. doi:10.1136/bmjopen-2019-029594 Open access BMJ Open: first published as 10.1136/bmjopen-2019-029594 on 30 August 2019. Downloaded from http://bmjopen.bmj.com/ on August 2, 2021 by guest. Protected by copyright. Table 2 Prevalence of the 60 chronic diseases included in the study in individuals aged 65–94 years (n=916 619, Catalonia, 2012) All diseases Rank Chronic conditions Frequency Percentage (%) included 1% 2% 1 Hypertension 650 899 71 2 Dyslipidaemia 466 585 50.9 3 Osteoarthritis and other degenerative joint diseases 300 803 32.8 4 Obesity 262 888 28.7 5 Diabetes 230 460 25.1 6 Anaemia 167 577 18.3 7 Cataract and other lens diseases 156 622 17.1 8 Chronic kidney diseases 153 756 16.8 9 Prostate diseases 153 635 16.8 10 Osteoporosis 151 847 16.6 11 Depression and mood diseases 148 751 16.2 12 Solid neoplasms 137 045 15 13 Colitis and related diseases 131 512 14.4 14 Venous and lymphatic diseases 126 997 13.9 15 Other musculoskeletal and joint diseases 124 765 13.6 16 Dorsopathies 124 603 13.6 17 Neurotic, stress-related and somatoform diseases 123 395 13.5 18 COPD, emphysema, chronic bronchitis 109 603 12 19 Ischaemic heart disease 95 434 10.4 20 Deafness, hearing impairment 90 261 9.9 21 Sleep disorders 88 739 9.7 22 Thyroid diseases 88 445 9.7 23 Other genitourinary diseases 85 468 9.3 24 Cerebrovascular disease 80 264 8.8 25 Atrial fibrillation 80 247 8.8 26 Oesophagus, stomach and duodenum diseases 80 043 8.7 27 Heart failure 74 077 8.1 28 Other eye diseases 68 939 7.5 29 Glaucoma 66 162 7.2 30 Inflammatory arthropathies 62 450 6.8 31 Dementia 59 213 6.5 32 Cardiac valve diseases 52 100 5.7 33 Peripheral neuropathy 49 127 5.4 34 Other psychiatric and behavioural diseases 46 841 5.1 35 Asthma 43 663 4.8 36 Allergy 40 394 4.4 37 Autoimmune diseases 39 350 4.3 38 Ear, nose, throat diseases 38 752 4.2 39 Peripheral vascular disease 30 674 3.4 40 Other neurological diseases 28 541 3.1 41 Chronic pancreas, biliary tract and gall bladder 27 321 3 diseases 42 Migraine and facial pain syndromes 25 999 2.8 43 Bradycardias and conduction diseases 25 476 2.8 44 Chronic liver diseases 22 633 2.5 Continued Violán C, et al. BMJ Open 2019;9:e029594. doi:10.1136/bmjopen-2019-029594 5 Open access BMJ Open: first published as 10.1136/bmjopen-2019-029594 on 30 August 2019. Downloaded from http://bmjopen.bmj.com/ on August 2, 2021 by guest. Protected by copyright. Table 2 Continued All diseases Rank Chronic conditions Frequency Percentage (%) included 1% 2% 45 Other digestive diseases 22 022 2.4 46 Parkinson and parkinsonism 20 833 2.3 47 Other metabolic diseases 18 997 2.1 48 Other cardiovascular diseases 16 833 1.8 49 Other skin diseases 15 363 1.7 50 Chronic ulcer of the skin 13 869 1.5 51 Blood and blood-forming organ diseases 13 575 1.5 52 Other respiratory diseases 9974 1.1 53 Epilepsy 8981 1 54 Haematological neoplasms 8174 0.9 55 Chronic infectious diseases 6647 0.7 56 Inflammatory bowel diseases 5549 0.6 57 Schizophrenia and delusional diseases 4792 0.5 58 Blindness, visual impairment 4772 0.5 59 Multiple sclerosis 576 0.1 60 Chromosomal abnormalities 77 0 In the last three columns, list of diseases was included by prevalence cut-off (All, 1%, 2%). COPD, chronic obstructive pulmonary disease. greatest number of visits over a 1 year period (median 18 In this pattern diseases did not exhibit an association visits), and the non-specified, youngest-old dominant pattern higher than chance because values of the O/E ratio and presenting the lowest median number of visits which was exclusivity were less than 2% and 20%, respectively. This equal to 5 (table 4). Online supplementary additional file suggests that such patients during their lives could change 2 shows tables of variables characterising each cluster in group. Two clusters presenting gender dominance were baseline study for 1% and for no prevalence cut-off points. observed: nervous, musculoskeletal and circulatory, female Multimorbidity patterns varied according to require- dominant was predominately made up of women >70 ments for minimal prevalence of selected conditions years, while genitourinary, mental and musculoskeletal, male in the population. As an example, figure 2 depicts the dominant was mostly formed of men of the same age. composition of cluster 1 according to prevalence levels of Such patterns represent 61% of the elderly participants disease, and the other clusters are shown in online supple- included in the study. The rest had fewer individuals and mentary additional file 3. Disease prevalence varied more some diseases were over-represented such as Parkinson greatly in the less populated patterns (eg, non-specified, and parkinsonism in nervous and digestive, and asthma in youngest-old dominant) (online supplementary additional respiratory, circulatory and nervous. file 3). Nevertheless, there was a group that remained in We observed that some diseases with O/E ratios ≥2 some clusters across all prevalence levels, for instance, were consistently associated with each other as part of some in neurological and digestive (Parkinson and parkin- the same clusters (for instance, nervous and digestive; respi- sonism, other neurological diseases, chronic liver diseases, ratory, circulatory and nervous; circulatory and digestive; and chronic pancreas, biliary tract and gall bladder diseases) mental, nervous and digestive, female dominant) regardless of formed part of the cluster regardless of changes in cut-off the prevalence threshold that had been set. They can be prevalence (online supplementary additional file 3). The considered core components of those clusters. Further selected level of prevalence resulted in changes in O/E research is needed to establish the role of these condi- ratios, with some of them doubling their values. tions from a longitudinal perspective. Comparison with the literature Discussion Comparison with other studies is hindered by variations The soft clustering method we employed identified eight in methods, data sources and structures, populations and multimorbidity patterns, regardless of the prevalence diseases studied. Nevertheless, there are similarities with selected. The non-specified, youngest-old dominant cluster other authors. The non-specified pattern is the one most included the largest number of individuals and those replicated in the literature, for example, Prados-Torres et who presented the smallest multimorbidity prevalence. al who employed an exploratory factor analysis25 and our 6 Violán C, et al. BMJ Open 2019;9:e029594. doi:10.1136/bmjopen-2019-029594 Open access BMJ Open: first published as 10.1136/bmjopen-2019-029594 on 30 August 2019. Downloaded from http://bmjopen.bmj.com/ on August 2, 2021 by guest. Protected by copyright. Table 3 Most frequent 15 diseases found in multimorbidity patterns in individuals aged 65–94 years (n=916 619, Catalonia, 2012) O/E O/E Pattern Disease O ratio EX Pattern Disease O ratio EX 1 Parkinson and 38.7 17 74.3 2 Asthma 34.5 7.2 40 Nervous and digestive parkinsonism Respiratory, circulatory and (n=40 037) Other neurological 49.5 15.9 69.4 nervous Peripheral vascular 13.9 4.2 22.9 diseases (n=50 639) disease Chronic liver diseases 13.2 5.4 23.4 Parkinson and 8.5 3.8 20.8 parkinsonism Chronic pancreas, 7.9 2.7 11.6 Other neurological 11.7 3.7 20.7 biliary tract and gall diseases bladder diseases Dementia 14.7 2.3 9.9 COPD, emphysema, 31 2.6 14.3 chronic bronchitis Other digestive 4.8 2 8.7 Allergy 10.8 2.4 13.5 diseases Cerebrovascular 16.9 1.9 8.4 Heart failure 16.6 2 11.3 disease Colitis and related 24.1 1.7 7.3 Ischaemic heart 21.1 2 11.2 diseases disease Other metabolic 3.4 1.7 7.2 Other eye diseases 14 1.9 10.3 diseases Depression and mood 25 1.5 6.7 Autoimmune diseases 7.2 1.7 9.3 diseases Anaemia 26.1 1.4 6.2 Other psychiatric and 8.5 1.7 9.2 behavioural diseases Oesophagus, 11.3 1.3 5.6 Ear, nose, throat 7.1 1.7 9.2 stomach and diseases duodenum diseases Sleep disorders 12.4 1.3 5.6 Anaemia 30.4 1.7 9.2 Other eye diseases 9.6 1.3 5.6 Peripheral neuropathy 8.8 1.6 9.1 Dorsopathies 17 1.2 5.4 Cerebrovascular 14.3 1.6 9 disease 3 Heart failure 51.4 6.4 46.9 4 Neurotic, stress- 64.9 4.8 49.7 Circulatory and Mental, nervous and related and digestive digestive, female dominant somatoform diseases (n=67 492) (n=94 453) Cardiac valve 34.2 6 44.3 Depression and mood 66.4 4.1 42.1 diseases diseases Atrial fibrillation 47.3 5.4 39.8 Migraine and facial 8.2 2.9 29.6 pain syndromes Bradycardias and 13.5 4.9 35.9 Sleep disorders 19 2 20.2 conduction diseases Ischaemic heart 33.7 3.2 23.8 Oesophagus, 14.9 1.7 17.6 disease stomach and duodenum diseases Chronic pancreas, 8 2.7 19.7 Osteoporosis 28 1.7 17.4 biliary tract and gall bladder diseases Chronic liver diseases 6.1 2.5 18.2 Thyroid diseases 16 1.7 17.1 Chronic kidney 35.9 2.1 15.8 Colitis and related 23.7 1.7 17 diseases diseases Anaemia 38.6 2.1 15.5 Other genitourinary 14.4 1.5 15.9 diseases Cerebrovascular 18.3 2.1 15.4 Ear, nose, throat 6.2 1.5 15.2 disease diseases COPD, emphysema, 23.6 2 14.5 Venous and lymphatic 19.9 1.4 14.8 chronic bronchitis diseases Other digestive 4.6 1.9 14 Allergy 6.1 1.4 14.3 diseases Continued Violán C, et al. BMJ Open 2019;9:e029594. doi:10.1136/bmjopen-2019-029594 7 Open access BMJ Open: first published as 10.1136/bmjopen-2019-029594 on 30 August 2019. Downloaded from http://bmjopen.bmj.com/ on August 2, 2021 by guest. Protected by copyright. Table 3 Continued O/E O/E Pattern Disease O ratio EX Pattern Disease O ratio EX Peripheral vascular 6.1 1.8 13.3 Osteoarthritis and 45 1.4 14.1 disease other degenerative joint diseases Other metabolic 3.2 1.5 11.3 Dorsopathies 18 1.3 13.7 diseases Dementia 9.5 1.5 10.9 Cardiac valve 7.4 1.3 13.5 diseases 5 Dementia 21.8 3.4 39.4 6 Peripheral neuropathy 12.4 2.3 36.6 Mental, digestive and Nervous, musculoskeletal Other digestive 5.8 2.4 28.1 Other musculoskeletal 26 1.9 30.2 blood, female oldest- and circulatory, female diseases and joint diseases old dominant dominant (n=106 845) Anaemia 38.5 2.1 24.6 (n=145 074) Venous and lymphatic 26.4 1.9 30.2 diseases Chronic kidney 33.3 2 23.1 Dorsopathies 25.3 1.9 29.4 diseases Colitis and related 26.2 1.8 21.3 Obesity 51 1.8 28.2 diseases Cerebrovascular 14.8 1.7 19.7 Other genitourinary 16 1.7 27.2 disease diseases Osteoporosis 26 1.6 18.3 Osteoarthritis and 55 1.7 26.5 other degenerative joint diseases Cataract and other 25.9 1.5 17.7 Osteoporosis 24.8 1.5 23.7 lens diseases Deafness, hearing 14 1.4 16.5 Other eye diseases 10.7 1.4 22.4 impairment Venous and lymphatic 19.5 1.4 16.4 Cataract and other 22.5 1.3 20.8 diseases lens diseases Osteoarthritis and 45.5 1.4 16.2 Thyroid diseases 12.6 1.3 20.7 other degenerative joint diseases Depression and mood 22.5 1.4 16.1 Glaucoma 9.2 1.3 20.1 diseases Other genitourinary 12.3 1.3 15.4 Diabetes 31.3 1.2 19.7 diseases Other eye diseases 9.9 1.3 15.4 Ear, nose, throat 5.2 1.2 19.5 diseases Sleep disorders 12.4 1.3 14.9 Dyslipidaemia 62.7 1.2 19.5 7 Prostate diseases 54.7 3.3 61.8 8 Dyslipidaemia 38.4 0.8 19.6 Genitourinary, mental Non-specified, youngest- Other psychiatric and 11.1 2.2 41.2 Thyroid diseases 7.3 0.8 19.6 and musculoskeletal, old dominant behavioural diseases male dominant (n=238 333) (n=173 746) Inflammatory 12.4 1.8 34.5 Osteoporosis 12.2 0.7 19.2 arthropathies COPD, emphysema, 20.5 1.7 32.5 Hypertension 47.6 0.7 17.4 chronic bronchitis Solid neoplasms 21.8 1.5 27.7 Glaucoma 4.4 0.6 16 Peripheral vascular 4.7 1.4 26.7 Solid neoplasms 9.1 0.6 15.7 disease Ischaemic heart 13.7 1.3 25 Migraine and facial 1.7 0.6 15.7 disease pain syndromes Diabetes 31.8 1.3 24 Autoimmune 2.2 0.5 13.4 diseases Ear, nose, throat 5.3 1.3 23.7 Other metabolic 1.1 0.5 13.3 diseases diseases Deafness, hearing 11.6 1.2 22.3 Allergy 2.2 0.5 13 impairment Allergy 4.8 1.1 20.5 Chronic liver diseases 1.2 0.5 12.8 Continued 8 Violán C, et al. BMJ Open 2019;9:e029594. doi:10.1136/bmjopen-2019-029594 Open access BMJ Open: first published as 10.1136/bmjopen-2019-029594 on 30 August 2019. Downloaded from http://bmjopen.bmj.com/ on August 2, 2021 by guest. Protected by copyright. Table 3 Continued O/E O/E Pattern Disease O ratio EX Pattern Disease O ratio EX Hypertension 75.8 1.1 20.2 Other genitourinary 4.5 0.5 12.7 diseases Glaucoma 7.5 1 19.6 Oesophagus, 4.1 0.5 12.2 stomach and duodenum diseases Autoimmune diseases 4.4 1 19.4 Other psychiatric and 2.4 0.5 12 behavioural diseases Obesity 29 1 19.2 Diabetes 10.8 0.4 11.2 COPD, chronic obstructive pulmonary disease; EX, exclusivity; O, disease prevalence in the cluster; O/E ratio, observed/expected ratio. group with k-means.24 Specifically, although the age range Strengths and limitations and the exclusivity threshold in our previous study were A major strength of this study is that it has employed a different, the hard clustering method provided clusters large, high-quality database made up of primary care that overlap with some of the patterns obtained in this records representative of the Catalan population aged ≥65 study, since both clustering results were predominantly years.18 Patterns of multimorbidity have been studied defined by the O/E ratio (≥2) criteria. However, the soft based on the whole eligible sample. This approach is approach allows a more flexible distribution of the indi- epidemiologically robust as the prevalence of diseases has vidual and diseases. been estimated on the whole sample rather than limited Recent research has provided support for physiopatho- to patients with multimorbidity.2 Another strength is that logical and genetic associations that explain the observed individuals rather than diseases have been considered as multimorbidity patterns. For instance, neurological and the unit of analysis.8 24 Such an approach permits a more digestive included chronic liver disease which has been realistic and rational monitoring of participants than linked to Parkinson through the accumulation of toxic cohort studies in order to analyse multimorbidity patterns substances in the brain (ammonia and manganese) along time. Moreover, the use of different prevalence and neuroinflammation.26 A higher risk of Parkinson cut-offs to obtain multimorbidity patterns has allowed the among patients with chronic hepatitis C virus has also identification of nuclear diseases. We selected the higher been reported (OR: 1.35),27 in addition to associations prevalence (2%) because the patterns obtained had more between digestive diseases and neurodegenerative ones clinical representativeness. The inclusion of all the poten- (eg, Parkinson and Alzheimer) through the microbi- tial diagnoses may have signified a greater complexity that ome-gut-brain axis.27 A possible link between micro- would have hindered both the interpretation of findings biota and digestive diseases such as chronic pancreatitis and comparison with other studies. and pancreatic cancer has also been suggested.28 29 For Compared with hierarchical clustering, fuzzy c-means the respiratory, circulatory and neurological cluster there is cluster analysis is less susceptible to: outliers in the data, evidence of an association between chronic bronchial choice of distance measure and the inclusion of inappro- pathology, particularly asthma and chronic obstructive priate or irrelevant variables.38 Nevertheless, some disad- pulmonary disease (COPD), and the risk of cardiovas- vantages of the method are that different solutions for cular events.30 Longitudinal studies have observed an each set of seed points can occur and there is no guarantee increased risk of developing Parkinson among individuals of optimal clustering.11 To minimise this shortcoming, we suffering from asthma and/or COPD.31 32 The association carried out 100 cluster realisations with different seeds between asthma and allergy is known, and its coexis- to finally use the average result of all of them. In addi- tence defines a specific phenotype. For the circulatory and tion, the method is not efficient when a large number digestive cluster, non-alcoholic fatty liver disease has been of potential cluster solutions are to be considered.38 To associated with the development of atrial fibrillation,33 address this limitation, we computed the optimal number and hepatitis C infection with an increase in the risk of of clusters using analytical indexes (online supplemen- developing cardiovascular and cerebrovascular events.34 tary additional file 1). In addition, anaemia has been associated with advanced Other limitations need to be taken into account. The stages of chronic renal diseases and erythropoietin defi- dimensional reduction method performed in this work to ciency.35 Iron deficiency anaemia has been associated reduce data complexity was PCAmix. Such methods can with an increased risk of stroke36 through thromboem- produce low percentages of variation on principal axes bolic phenomena secondary to reactive thrombocytosis. and make it difficult to choose the number of dimensions Chronic kidney disease produces auricle injuries (dila- to retain. In order to decide on the most suitable number tation, fibrosis) and systemic inflammation, both of of dimensions we applied the Karlis-Saporta-Spinaki rule27 which can favour the onset and maintenance of atrial which resulted in a 13-dimensional space for the 2% prev- fibrillation.37 alence cut-off. Furthermore, the feasibility of developing Violán C, et al. BMJ Open 2019;9:e029594. doi:10.1136/bmjopen-2019-029594 9 10 Table 4 Variables characterising each cluster in baseline study for 2% prevalence cut-off point (n=916 619) 5 4 Mental, 6 7 8 2 Mental, nervous digestive and Nervous, Genitourinary, Non- 1 Respiratory, 3 and digestive, blood, female musculoskeletal mental and specified, Nervous and circulatory and Circulatory and female oldest-old and circulatory, musculoskeletal, youngest-old Open access digestive nervous digestive dominant dominant female dominant male dominant dominant All Number of 40 037 50 639 67 492 94 453 106 845 145 074 173 746 238 333 916 619 people Multimorbidity, 39 776 (99.3) 50 513 (99.8) 67 443 (99.9) 94 442 (100.0) 106 696 (99.9) 144 869 (99.9) 171 983 (99.0) 177 363 (74.4) 853 085 (93.1) n (%) Polypharmacy, 28 484 (71.1) 38 869 (76.8) 54 658 (81.0) 64 154 (67.9) 71 830 (67.2) 86 317 (59.5) 90 603 (52.1) 52 588 (22.1) 487 502 (53.1) n (%) Women, n (%) 22 628 (56.5) 26 690 (52.7) 38 023 (56.3) 78 922 (83.6) 85 735 (80.2) 113 629 (78.3) 15 730 (9.1) 147 773 (62.0) 529 131 (57.7) Men, n (%) 17 409 (43.5) 23 949 (47.3) 29 469 (43.7) 15 531 (16.4) 21 110 (19.8) 31 445 (21.7) 158 016 (90.9) 90 560 (38.0) 387 488 (42.3) Age (categories),                   n (%)  (65, 70) 7188 (18.0) 10 400 (20.5) 7233 (10.7) 28 305 (30.0) 12 036 (11.3) 38 829 (26.8) 52 003 (29.9) 96 184 (40.4) 252 178 (27.5)  (70, 80) 17 804 (44.5) 22 743 (44.9) 24 724 (36.6) 40 577 (43.0) 33 624 (31.5) 70 643 (48.7) 84 037 (48.4) 100 435 (42.1) 394 586 (43.0)  (80, 90) 13 460 (33.6) 15 568 (30.7) 29 908 (44.3) 22 638 (24.0) 48 453 (45.3) 32 714 (22.6) 34 785 (20.0) 37 217 (15.6) 234 744 (25.6)  (90, 99) 1587 (4.0) 1927 (3.8) 5628 (8.3) 2934 (3.1) 12 732 (11.9) 2888 (2.0) 2920 (1.7) 4497 (1.9) 35 111 (3.8) MEDEA index*                    R 7831 (21.8) 9300 (20.2) 13 718 (23.2) 17 266 (19.7) 22 183 (23.0) 27 401 (20.0) 35 145 (21.5) 49 405 (21.9) 182 249 (21.4)  U1 6010 (16.7) 6890 (15.0) 9537 (16.1) 15 027 (17.2) 16 556 (17.2) 19 599 (14.3) 25 656 (15.7) 45 516 (20.2) 144 791 (17.0)  U2 5690 (15.8) 7134 (15.5) 9140 (15.4) 14 335 (16.4) 15 272 (15.8) 21 379 (15.6) 25 951 (15.9) 37 530 (16.6) 136 431 (16.0)  U3 5941 (16.5) 7520 (16.4) 9187 (15.5) 14 223 (16.3) 15 421 (16.0) 23 261 (16.9) 26 908 (16.5) 35 761 (15.8) 138 222 (16.2)  U4 5540 (15.4) 7686 (16.7) 9016 (15.2) 14 012 (16.0) 14 272 (14.8) 23 780 (17.3) 26 526 (16.2) 32 157 (14.2) 132 988 (15.6)  U5 4982 (13.8) 7421 (16.2) 8638 (14.6) 12 652 (14.5) 12 699 (13.2) 21 923 (16.0) 23 064 (14.1) 25 506 (11.3) 116 883 (13.7) Number of 8.0 (6.0; 10.0) 8.0 (6.0; 10.0) 8.0 (7.0; 11.0) 7.0 (6.0; 9.0) 7.0 (5.0; 9.0) 6.0 (5.0; 8.0) 5.0 (4.0; 7.0) 3.0 (3.0; 4.0) 6.0 (4.0; 8.0) chronic diseases, median (IQR) Number of                   chronic diseases (categories), n (%)  0 0 (0.0) 0 (0.0) 0 (0.0) 0 (0.0) 1 (0.0) 0 (0.0) 235 (0.1) 25 144 (10.5) 25 380 (2.8)  1 262 (0.7) 125 (0.2) 49 (0.1) 11.0 (0.0) 149 (0.1) 204 (0.1) 1528 (0.9) 35 826 (15.0) 38 154 (4.2)  (2, 5) 5409 (13.5) 4507 (8.9) 4275 (6.3) 8781 (9.3) 14 601 (13.7) 22 400 (15.4) 57 561 (33.1) 151 302 (63.5) 268 836 (29.3)  (5, 10) 23 502 (58.7) 30 257 (59.8) 37 910 (56.2) 62 490 (66.2) 73 427 (68.7) 105 620 (72.8) 104 915 (60.4) 25 588 (10.7) 463 709 (50.6) Continued Violán C, et al. BMJ Open 2019;9:e029594. doi:10.1136/bmjopen-2019-029594 BMJ Open: first published as 10.1136/bmjopen-2019-029594 on 30 August 2019. Downloaded from http://bmjopen.bmj.com/ on August 2, 2021 by guest. Protected by copyright. Table 4 Continued 5 4 Mental, 6 7 8 2 Mental, nervous digestive and Nervous, Genitourinary, Non- 1 Respiratory, 3 and digestive, blood, female musculoskeletal mental and specified, Nervous and circulatory and Circulatory and female oldest-old and circulatory, musculoskeletal, youngest-old digestive nervous digestive dominant dominant female dominant male dominant dominant All  ≥10 10 864 (27.1) 15 749 (31.1) 25 259 (37.4) 231 715 (24.5) 18 668 (17.5) 16 850 (11.6) 9506 (5.5) 473 (0.2) 120 540 (13.2) Number of drugs, 7.0 (4.0; 9.0) 7.0 (5.0; 10.0) 8.0 (5.0; 11.0) 6.0 (4.0; 9.0) 6.0 (4.0; 9.0) 5.0 (3.0; 8.0) 5.0 (3.0; 7.0) 2.0 (0.0; 4.0) 5.0 (2.0; 8.0) median (IQR) Number of drugs                   (categories)  0 2576 (6.4) 2491 (4.9) 3349 (5.0) 5636 (6.0) 7037 (6.6) 8330 (5.7) 13 389 (7.7) 70 561 (29.6) 113 368 (12.4) Violán C, et al. BMJ Open 2019;9:e029594. doi:10.1136/bmjopen-2019-029594  1 1212 (3.0) 1072 (2.1) 1015 (1.5) 2939 (3.1) 3390 (3.2) 6772 (4.7) 11 440 (6.6) 29 242 (12.3) 57 082 (6.2)  (2, 5) 7766 (19.4) 8207 (16.2) 8471 (12.6) 21 725 (23.0) 24 587 (23.0) 43 656 (30.1) 58 314 (33.6) 85 942 (36.1) 258 667 (28.2)  (5, 10) 18 510 (46.2) 23 597 (46.6) 31 850 (47.2) 46 022 (48.7) 52 653 (49.3) 68 193 (47.0) 73 694 (42.4) 48 161 (20.2) 362 681 (39.6)  ≥10 9973 (24.9) 15 272 (30.2) 22 808 (33.8) 18 132 (19.2) 19 177 (17.9) 18 123 (12.5) 16 909 (9.7) 4427 (1.9) 124 821 (13.6) Number of visits 12.0 (7.0; 20.0) 14.0 (8.0; 22.0) 18.0 (9.0; 30.0) 11.0 (6.0; 19.0) 12.0 (7.0; 19.0) 11.0 (7.0; 17.0) 9.0 (5.0; 15.0) 5.0 (2.0; 9.0) 9.0 (5.0; 16.0) 2012, median (IQR) Number of visits                   2012 (categories), n (%)  0 976 (2.4) 871 (1.7) 1143 (1.7) 2219 (2.3) 2515 (2.4) 2410.3 (1.7) 4137 (2.4) 33 673 (14.1) 47 945 (5.2)  1 874 (2.2) 754 (1.5) 929 (1.4) 2055 (2.2) 2238 (2.1) 2412.4 (1.7) 4685 (2.7) 19 938 (8.4) 33 884 (3.7)  (2, 5) 4000 (10.0) 3918 (7.7) 4329 (6.4) 10 589 (11.2) 11 018 (10.3) 14 943.7 (10.3) 24 319 (14.0) 57 322 (24.1) 130 439 (14.2)  (5, 10) 9158 (22.9) 10 774 (21.3) 10 883 (16.1) 24 504 (25.9) 27 003 (25.3) 42 180.7 (29.1) 54 212 (31.2) 70 634 (29.6) 249 349 (27.2)  ≥10 25 030 (62.5) 34 322 (67.8) 50 209 (74.4) 55 085 (58.3) 64 071 (60.0) 83 126.5 (57.3) 86 393 (49.7) 56 766 (23.8) 455 002 (49.6) For the sake of simplicity, all numbers in the table were rounded to its closest natural number. *MEDEA index goes from 1 (least deprived) to 5 (most deprived), in this variable n=851 564. Open access 11 BMJ Open: first published as 10.1136/bmjopen-2019-029594 on 30 August 2019. Downloaded from http://bmjopen.bmj.com/ on August 2, 2021 by guest. Protected by copyright. Open access BMJ Open: first published as 10.1136/bmjopen-2019-029594 on 30 August 2019. Downloaded from http://bmjopen.bmj.com/ on August 2, 2021 by guest. Protected by copyright. Figure 2 Composition of cluster 1 (nervous and digestive) in individuals aged 65–94 years according to disease levels of prevalence (n=916 619, Catalonia, 2012). COPD, chronic obstructive pulmonary disease; O/E ratio, observed/expected ratio. clinical practice guidelines in accordance with these services, polypharmacy and increased mortality), to a patterns might prove difficult due to the dimension of multimorbidity orientation.39 the diseases included in each pattern. Nonetheless, new Further investigation on this topic is called for with clinical practice guidelines should consider the diseases particular focus on five major issues. First, the genetic that are over-represented (O/E ratio ≥2). study of these patterns will help the identification of risk subgroups. Second, research is needed on the life style Implications for practice, policy and research and environmental factors (diet, physical exercise, toxics) Soft clustering methods offer a new methodological associated with such patterns. Third, longitudinal studies approach to understanding the relationships between should be performed to establish the onset order of the specific diseases in individuals. This is an essential step core diseases. Fourth, alternative approaches to handle in improving the care of patients and health systems. covariates in cluster analysis should be addressed in future Analysing multimorbidity patterns permits the identi- analysis plan. Recently, a new method that allows the fication of patient subgroups with different associated covariates to be incorporated into the membership factor diseases. Our analysis focuses on groups of patients as to model individual probabilities of cluster membership opposed to diseases. In this case, a disease is present has been proposed.40 And fifth, the characteristics of the in all patterns (clusters), but in different degrees. In diseases in the same cluster and their potential implica- this context, the O/E ratios are used to measure which tion on the quality of primary care should be ascertained diseases are over-represented in each cluster and to lead in greater detail. the clinical practice guidelines. The inclusion of varying Our findings suggest non-hierarchical cluster analysis cut-off points (prevalence filters) of the diseases that identified multimorbidity patterns and phenotypes of form the multimorbidity patterns allowed us to identify certain subgroups of patients that were more consistent common nuclear diseases that remained independent with clinical practice. from the prevalence that build such patterns. It is noteworthy that 60% of the population ≥65 years Author affiliations 1 was included in multimorbidity patterns made up of the Fundació Institut Universitari per a la recerca a l'Atenció Primària de Salut Jordi most prevalent diseases. The rest of the population was Gol i Gurina (IDIAPJGol), Barcelona, Spain 2 Universitat Autònoma de Barcelona, Bellaterra (Cerdanyola del Vallès), Barcelona, grouped into five more specific patterns which permitted Spain their better management. 3 Department of Psychiatry, Vic University Hospital. Francesc Pla el Vigatà, 1, Vic, While clinical guidelines are currently aimed at Barcelona, Spain covering the management of the diseases found in the 4 Department of Basic and Methodological Sciences, Faculty of Health Sciences and non-specified, youngest-old dominant cluster, there is a lack Welfare. University of Vic- Central University of Catalonia (UVic-UCC), Vic, Barcelona, of information regarding the associated diseases in the Spain 5 Signal Theory and Communications Department, Universitat Politecnica de other patterns. The challenge will be to refocus health- Catalunya, Barcelona, Spain care policy from that based on individual diseases, with 6 Hospital Universitari de Bellvitge, Barcelona, Spain the accompanying consequences (increased risk of 7 Health Services and Policy Research Group, Academic Collaboration for Primary functional decline, poorer quality of life, greater use of Care, University of Exeter Medical School, Exeter, UK 12 Violán C, et al. BMJ Open 2019;9:e029594. doi:10.1136/bmjopen-2019-029594 Open access BMJ Open: first published as 10.1136/bmjopen-2019-029594 on 30 August 2019. Downloaded from http://bmjopen.bmj.com/ on August 2, 2021 by guest. Protected by copyright. Contributors All authors contributed to the design of the study, revised the article polypharmacy: a systematic guideline review and expert consensus. and approved the final version. CVF, ARL and SFB obtained the funding. CVF, QFB J Intern Med 2018;2. [Epub ahead of print]. and SFB drafted the article. CVF, QFB, SFB, MGC, MCB, FF, JMV and ARL contributed 10. Palmer K, Marengoni A, Forjaz MJ, et al. Multimorbidity care model: to the analysis and interpretation of data. CVF, QFB and SFB wrote the first draft, recommendations from the consensus meeting of the joint action on chronic diseases and promoting healthy ageing across the life cycle and all authors contributed ideas, interpreted the findings and reviewed rough (JA-CHRODIS). Health Policy 2018;122:4–11. drafts of the manuscript. 11. Wolfram. Fuzzy Clustering [Internet]. Available: https://​reference.​ Funding This work was supported by a research grant from the Carlos III wolfram.​com/​legacy/​applications/​fuzzylogic/​Manual/​12.​html Institute of Health, Ministry of Economy and Competitiveness (Spain), awarded on 12. MathWorks. Fuzzy Clustering [Internet]. Available: https://www.​ mathworks.​com/​help/​fuzzy/​fuzzy-​clustering.​html the 2016 call under the Health Strategy Action 2013–2016, within the National 13. France EF, Wyke S, Gunn JM, et al. Multimorbidity in primary care: Research Program oriented to Societal Challenges, within the Technical, Scientific a systematic review of prospective cohort studies. Br J Gen Pract and Innovation Research National Plan 2013–2016 (grant number PI16/00639), 2012;62:e297–307. cofunded with European Union ERDF funds (European Regional Development Fund) 14. SK N, Tawiah R, Sawyer M, et al. Patterns of multimorbid health and Department of Health of the Catalan Government, in the call corresponding to conditions: a systematic review of analytical methods and 2017 for the granting of subsidies from the Strategic Plan for Research in Health comparison analysis. Int J Epidemiol 2018;47:1687–704. (Pla Estratègic de Recerca i Innovació en Salut, PERIS) 2016–2020, modality 15. Violán C, Foguet-Boreu Q, Roso-Llorach A, et al. Burden of research oriented to primary care (grant number SLT002/16/00058) and from the multimorbidity, socioeconomic status and use of health services across stages of life in urban areas: a cross-sectional study. BMC Catalan Government (grant number AGAUR 2017 SGR 578). Public Health 2014;14:530. Disclaimer The views expressed in this publication are those of the author(s) and 16. Willadsen TG, Bebe A, Køster-Rasmussen R, et al. The role not necessarily those of the National Health Service, the National Institute for Health of diseases, risk factors and symptoms in the definition of Research or the National Department of Health. multimorbidity – a systematic review. Scand J Prim Health Care 2016;34:112–21. Competing interests None declared. 17. Xu X, Mishra GD, Jones M. Evidence on multimorbidity from definition to intervention: an overview of systematic reviews. Ageing Patient consent for publication Not required. Res Rev 2017;37:53–68. Ethics approval The protocol of the study was approved by the Committee on 18. Del Mar García-Gil M, Hermosilla E, Prieto-Alhambra D, et al. the Ethics of Clinical Research, Fundació Institut Universitari per a la recerca a Construction and validation of a scoring system for the selection l'Atenció Primària de Salut Jordi Gol i Gurina (IDIAPJGol) (P16/151). All data were of high-quality data in a Spanish population primary care database (SIDIAP). Inform Prim Care 2012;19:135–45. anonymised and the confidentiality of EHR was respected at all times in accordance 19. Calderón-Larrañaga A, Vetrano DL, Onder G, et al. Assessing with national and international law. and measuring chronic multimorbidity in the older population: a Provenance and peer review Not commissioned; externally peer reviewed. proposal for its Operationalization. J Gerontol A Biol Sci Med Sci 2017;72:1417–23. Data availability statement All data relevant to the study are included in the 20. Domínguez-Berjón MF, Borrell C, Cano-Serral G, et al. Constructing article or uploaded as supplementary information. a deprivation index based on census data in large Spanish cities (the MEDEA project)]. Gac Sanit 2008;22:179–87. Open access This is an open access article distributed in accordance with the 21. Karlis D, Saporta G, Spinakis A. A simple rule for the selection Creative Commons Attribution Non Commercial (CC BY-NC 4.0) license, which of principal components. Commun Stat Theory Methods permits others to distribute, remix, adapt, build upon this work non-commercially, 2003;32:643–66. and license their derivative works on different terms, provided the original work is 22. Bezdek JC, Ehrlich R, Full W. FCM: the fuzzy C-Means clustering properly cited, appropriate credit is given, any changes made indicated, and the use algorithm. Comput Geosci 1984;10:191–203. is non-commercial. See: http://​creativecommons.​org/​licenses/​by-​nc/​4.​0/. 23. Bora DJ, Gupta DAK, Kumar Gupta A. A comparative study between fuzzy clustering algorithm and hard clustering algorithm. IJCTT 2014;10:108–13. 24. Violán C, Roso-Llorach A, Foguet-Boreu Q, et al. Multimorbidity patterns with k-means nonhierarchical cluster analysis. BMC Fam Pract 2018;19:108. References 25. Prados-Torres A, Poblador-Plou B, Calderón-Larrañaga A, et al. 1. Valderas JM, Starfield B, Sibbald B, et al. Defining comorbidity: Multimorbidity patterns in primary care: interactions among chronic implications for understanding health and health services. Ann Fam diseases using factor analysis. PLoS One 2012;7:e32190. Med 2009;7:357–63. 26. Shin HW, Park HK. Recent updates on acquired hepatocerebral 2. Violan C, Foguet-Boreu Q, Flores-Mateo G, et al. Prevalence, degeneration. Tremor Other Hyperkinet Mov 2017;7. determinants and patterns of multimorbidity in primary care: 27. Wijarnpreecha K, Chesdachai S, Jaruvongvanich V, et al. Hepatitis C a systematic review of observational studies. PLoS One virus infection and risk of Parkinson’s disease: A systematic review 2014;9:e102149. and meta-analysis. Eur J Gastroenterol Hepatol 2018;30:9–13. 3. Salive ME. Multimorbidity in older adults. Epidemiol Rev 28. Westfall S, Lomis N, Kahouli I, et al. Microbiome, probiotics and 2013;35:75–83. neurodegenerative diseases: deciphering the gut brain axis. Cell. 4. Barnett K, Mercer SW, Norbury M, et al. Epidemiology of Mol. Life Sci. 2017;74:3769–87. multimorbidity and implications for health care, research, 29. Memba R, Duggan SN, Ni Chonchubhair HM, et al. The potential and medical education: a cross-sectional study. The Lancet role of gut microbiota in pancreatic disease: a systematic review. 2012;380:37–43. Pancreatology 2017;17:867–74. 5. Vos T, Barber RM, Bell B, et al. Global, regional, and national 30. Xu M, Xu J, Yang X. Asthma and risk of cardiovascular disease or incidence, prevalence, and years lived with disability for 301 acute all-cause mortality: a meta-analysis. Ann Saudi Med 2017;37:99–105. and chronic diseases and injuries in 188 countries, 1990–2013: a 31. Cheng C-M, Wu Y-H, Tsai S-J, et al. Risk of developing Parkinson's systematic analysis for the global burden of disease study 2013. The disease among patients with asthma: a nationwide longitudinal Lancet 2015;386:743–800. study. Allergy 2015;70:1605–12. 6. Gruneir A, Bronskill SE, Maxwell CJ, et al. The association between 32. CH L, Chen WC, Liao WC, et al. The association between chronic multimorbidity and hospitalization is modified by individual obstructive pulmonary disease and Parkinson’s disease: A demographics and physician continuity of care: a retrospective nationwide population-based retrospective cohort study. Qjm cohort study. BMC Health Serv Res 2016;16:154. 2015;108:39–45. 7. Rocca WA, Boyd CM, Grossardt BR, et al. Prevalence of 33. Wijarnpreecha K, Boonpheng B, Thongprayoon C, et al. The multimorbidity in a geographically defined American population: association between non-alcoholic fatty liver disease and atrial patterns by age, sex, and race/ethnicity. Mayo Clin Proc fibrillation: a meta-analysis. Clin Res Hepatol Gastroenterol 2014;89:1336–49. 2017;41:525–32. 8. Prados-Torres A, Calderón-Larrañaga A, Hancco-Saavedra J, et al. 34. Ambrosino P, Lupoli R, Di Minno A, et al. The risk of coronary artery Multimorbidity patterns: a systematic review. J Clin Epidemiol disease and cerebrovascular disease in patients with hepatitis C: a 2014;67:254–66. systematic review and meta-analysis. Int J Cardiol 2016;221:746–54. 9. Muth C, Blom JW, Smith SM, et al. Evidence supporting the 35. Kepez A, Mutlu B, Degertekin M, et al. Association between left best clinical management of patients with multimorbidity and ventricular dysfunction, anemia, and chronic renal failure. Analysis of Violán C, et al. BMJ Open 2019;9:e029594. doi:10.1136/bmjopen-2019-029594 13 Open access BMJ Open: first published as 10.1136/bmjopen-2019-029594 on 30 August 2019. Downloaded from http://bmjopen.bmj.com/ on August 2, 2021 by guest. Protected by copyright. the heart failure prevalence and predictors in turkey (happy) cohort. 38. Badsha MB, Mollah MNH, Jahan N, et al. Robust complementary Herz 2015;40:616–23. hierarchical clustering for gene expression data analysis by 36. Chang Y-L, Hung S-H, Ling W, et al. Association between ischemic β-divergence. J Biosci Bioeng 2013;116:397–407. stroke and iron-deficiency anemia: a population-based study. PLoS 39. Yarnall AJ, Sayer AA, Clegg A, et al. New horizons in multimorbidity One 2013;8:e82952. in older adults. Age Ageing 2017;46:882–8. 37. Turakhia MP, Blankestijn PJ, Carrero J-J, et al. Chronic kidney 40. SKN, TawiahR, McLachlanGJ. Unsupervised pattern recognition disease and arrhythmias: conclusions from a kidney disease: of mixed data structures with numerical and categorical features improving global outcomes (KDIGO) controversies conference. Eur using a mixture regression modelling framework. Patter Recognition Heart J 2018;39:2314–25. 2019;88:261–71. 14 Violán C, et al. BMJ Open 2019;9:e029594. doi:10.1136/bmjopen-2019-029594