ARTICLE https://doi.org/10.1038/s41467-020-16780-x OPEN Twelve-year clinical trajectories of multimorbidity in a population of older adults Davide L. Vetrano 1,2,8 ✉, Albert Roso-Llorach 3,4,8, Sergio Fernández3,4, Marina Guisado-Clavero3,4, Concepción Violán3,4, Graziano Onder5, Laura Fratiglioni1,6, Amaia Calderón-Larrañaga1,9 & Alessandra Marengoni1,7,9 1234567890():,; Multimorbidity—the co-occurrence of multiple diseases—is associated to poor prognosis, but the scarce knowledge of its development over time hampers the effectiveness of clinical interventions. Here we identify multimorbidity clusters, trace their evolution in older adults, and detect the clinical trajectories and mortality of single individuals as they move among clusters over 12 years. By means of a fuzzy c-means cluster algorithm, we group 2931 people ≥60 years in five clinically meaningful multimorbidity clusters (52%). The remaining 48% are part of an unspecific cluster (i.e. none of the diseases are overrepresented), which greatly fuels other clusters at follow-ups. Clusters contribute differentially to the longitudinal development of other clusters and to mortality. We report that multimorbidity clusters and their trajectories may help identifying homogeneous groups of people with similar needs and prognosis, and assisting clinicians and health care systems in the personalization of clinical interventions and preventive strategies. 1 Aging Research Center, Department of Neurobiology, Care Sciences and Society, Karolinska Institutet and Stockholm University, Tomtebodavägen 18A, SE- 171 65 Solna, Stockholm, Sweden. 2 Centro Medicina dell’Invecchiamento, Fondazione Policlinico Universitario “A. Gemelli” IRCCS, and Università Cattolica del Sacro Cuore, L.go Francesco Vito 1, 00100 Rome, Italy. 3 Fundació Institut Universitari per a la recerca a l’Atenció Primària de Salut Jordi Gol i Gurina (IDIAPJGol), Gran Via 587 àtic, Barcelona, Spain. 4 Universitat Autònoma de Barcelona, Campus de la UAB, Plaça Cívica, 08193 Bellaterra (Cerdanyola del Vallès), Spain. 5 Department of Cardiovascular, Endocrine-Metabolic Diseases and Aging, Istituto Superiore di Sanità, Via Giano della Bella 34, 00161 Rome, Italy. 6 Stockholm Gerontology Research Center, Sveavägen 155, 11346 Stockholm, Sweden. 7 Department of Clinical and Experimental Sciences, University of Brescia, Piazza Mercato 15, 25121 Brescia, Italy. 8These authors contributed equally: Davide L Vetrano, Albert Roso-Llorach. 9These authors jointly supervised this work: Amaia Calderón-Larrañaga, Alessandra Marengoni. ✉email: davide.vetrano@ki.se NATURE COMMUNICATIONS | (2020)11:3223 | https://doi.org/10.1038/s41467-020-16780-x | www.nature.com/naturecommunications 1 ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-020-16780-x A s people age they tend to develop multiple chronic dis- ongoing population-based study started in 2001 and involving eases; the term multimorbidity identifies this condition1. 3363 individuals aged ≥60 years from a central area in Stockholm, After 60 years of age, 55–98% of people are affected by Sweden. From the whole sample, 432 participants with <2 chronic two or more chronic diseases, and patients with multimorbidity disease have been excluded (i.e., those without multimorbidity). account for up to 80% of consultations with general practitioners Those excluded were younger, reported a higher level of educa- and virtually all consultations with geriatricians2,3. Co-occurring tion, and were more often male than those included in the study diseases interact with each other, increasing the risk of negative (p for t test < 0.001). At baseline, study participants’ mean age was events beyond the sum of the risk of each disease4. Multi- 76.1 ± 11.0 [standard deviation] and 1951 (66.6%) were female. morbidity triggers complex pharmacological regimes, increases Over the 12 years, 1290 (44%) deaths occurred (28% within the the use of health care resources, and reduces the quality and first 6 years and 16% between 6 and 12 years). Moreover, 625 length of life1,4–6. It challenges physicians, who are usually (22%) individuals dropped out (14% within the first 6 years and trained to consider only a limited number of interactions between 8% between 6 and 12 years). At each follow-up, we performed a diseases and between diseases and drugs, and it puts pressure on dimensionality reduction (i.e., multiple correspondence analysis) health care systems, which struggle to offer older adults with to obtain the input data for participants’ clustering. A fuzzy c- multimorbidity comprehensive assessment, effective treatments, means cluster analysis with optimal a fuzziness parameter at m = and integrated care paths6–10. Moreover, because older people 1.1 (which outperformed other m values; see “Methods”) was with multimorbidity are usually excluded from randomized employed to identify clusters of individuals based on their clinical trials, there are few clear recommendations about how to underlying patterns of multimorbidity. Using an observed/ provide health care for older adults with multimorbidity. Com- expected ratio ≥2 (O/E ratio; i.e., the ratio between the prevalence plexity is thus translated into frustrating uncertainty and pow- of a given condition in a cluster and its prevalence in the whole erlessness and affects the quality of care at every level of the sample) and an exclusivity ≥25% (i.e., the proportion of indivi- health care process9. duals with a given condition in the whole sample that belong to a Both clinical experience and epidemiological studies suggest cluster) for each disease, five clusters of people were identified at that diseases cluster in the same person according to specific baseline: those with psychiatric and respiratory diseases (5.4%), patterns5,11. Several clusters of diseases have been identified with heart diseases (9.3%), respiratory and musculoskeletal diseases some consistency across studies; however, there are a number of (15.7%), cognitive and sensory impairment (10.6%), and eye dis- discrepancies in study findings12. A systematic review by Prados- eases and cancer (10.7%). Solutions were evaluated based on their Torres et al. identified 97 clusters of multimorbidity, and the clinical consistency and significance criteria (Supplementary findings of most of the reviewed studies suggested three clusters Figs. 1–15). Half of the people (48.7%) were grouped in an of multimorbidity: cardiometabolic, mental health, and muscu- additional unspecific cluster, as they were affected by prevalent loskeletal. At the same time, the studies in the review identified diseases but whose occurrence did not exceed the expected. many unexplained heterogeneous clusters12. In addition to Similarly, five clusters (plus the unspecific one) were identified at between-study methodological differences, one of the explana- 6 and 12 years. At follow-ups, those diseases characterizing the tions for this finding may lie in the dynamic nature of disease baseline clusters were regrouped into different multimorbidity clusters, which is not accounted for in cross-sectional studies. clusters. The clinical characteristics of the clusters are reported in These clusters evolve overtime, and mortality selection plays an Supplementary Table 1. important role in shaping the observed population13. Capturing such dynamism is the only way to better understand the natural history of multimorbidity and to shed light on previously unex- Individuals had different demographic, clinical and functional plained findings. profiles across the clusters. Descriptive analyses were carried out Most previous studies in this field have focused on clusters to characterize the six clusters of individuals with multimorbidity. from the viewpoint of disease analyses rather than the analysis of At baseline, participants in the cognitive and sensory diseases, the groups of individuals12,14. Focusing on people is in keeping with eye diseases and cancer, and the heart diseases clusters were the the principle of patient-centered care and can provide informa- oldest. Participants in the heart diseases, the eye diseases and tion that facilitates the move toward personalized medicine15. A cancer, and the psychiatric and respiratory diseases clusters pre- better understanding of older adults’ transitions among multi- sented the greatest number of chronic diseases (mean number: morbidity clusters overtime may help detect homogeneous 7.7 ± 2.4 [standard deviation], 6.0 ± 2.0, and 5.7 ± 2.2, respec- groups of individuals who may benefit from similar preventive tively). Participants in the heart diseases and psychiatric and (secondary and tertiary) interventions, treatment, and care. We respiratory diseases clusters and those in the cognitive and sensory therefore aimed to identify multimorbidity clusters in a impairment cluster used the highest number of drugs (mean population-based cohort of older adults, trace the evolution of the number: 7.7 ± 3.5, 6.2 ± 3.7, and 6.1 ± 3.4, respectively). More- clusters over 12 years, and follow the clinical trajectories of the over, individuals included in the heart diseases, the eye diseases individuals as they moved between clusters or to death over time. and cancer, and the cognitive and sensory impairment clusters We found that multimorbidity clusters change dynamically presented the highest prevalence of disability and slow walking overtime in older adults, following different clinical trajectories. speed. The cognitive and sensory impairment and the psychiatric Different clusters are also associated with different prognosis. and respiratory diseases cluster showed the lowest Mini-Mental Multimorbidity trajectories may help identifying homogeneous State Examination (MMSE) scores. The unspecific cluster was groups of people with similar needs and prognosis, and assisting characterized by the lowest mean age and the lowest number of clinicians and health care systems in the personalization of clin- chronic diseases and drugs. This group had the lowest prevalence ical interventions and preventive strategies. of disability and the highest walking speed, yet it had a high prevalence of hypertension, diabetes, dyslipidemia, and obesity. Such conditions were frequent also among participants in the Results heart diseases and the eye diseases and cancer clusters. Six clusters of individuals with multimorbidity were identified. At follow-ups, in spite of varied clustering, a similar clinical The present study is based on data from the Swedish National distribution was observed for the different types of disorders. Study on Aging and Care in Kungsholmen (SNAC-K), an That is, people in clusters characterized by cardiovascular, 2 NATURE COMMUNICATIONS | (2020)11:3223 | https://doi.org/10.1038/s41467-020-16780-x | www.nature.com/naturecommunications NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-020-16780-x ARTICLE 3000 Psychiatric and resp. Heart and vasc. diseases Vascular diseases diseases Heart diseases and Cardiometabolic diseases cognitive impairment Respiratory diseases Heart diseases Neuropsychiatric and Neuropsychiatric diseases respiratory diseases Eye and MSK diseases Eye diseases and cancer Eye diseases MSK, respiratory and Unspecific Cognitive and sensory immune diseases 2000 impairment Number of participants Respiratory and MSK diseases Unspecific Death 1000 Death Unspecific Dropout Dropout 0 Baseline 6 years 12 years Fig. 1 Evolution of multimorbidity clusters and clinical trajectories of older adults with multimorbidity over 12 years. The height of the boxes and the thickness of the stripes are proportional to the amount of people belonging to the cluster and moving from the cluster, respectively. MSK musculoskeletal. To note, for this analysis participants were assigned to the cluster they were more likely to belong in order to investigate the most likely trajectories. neuropsychiatric, and respiratory diseases showed the highest sex, and education, taking the unspecific cluster as the reference number of diseases and drugs and the highest levels of functional group. As shown in Table 1, at baseline the heart diseases (OR impairment. 3.07; 95% CI 2.26–4.19), the cognitive and sensory impairment (OR 6.00; 95% CI 4.21–8.54), and the psychiatric and respiratory Patterns of transitions between clusters can be identified over diseases (OR 1.60; 95% CI 1.02–2.51) clusters were significantly time. Upon assigning the individuals into the cluster they were associated with a higher six-year mortality, compared with the more likely to belong to, we described their trajectories as they people in the unspecific cluster. These clusters accounted for 51% moved between clusters or to death over time. Figure 1 depicts of deaths. At first follow-up, the heart and vascular diseases (OR the longitudinal evolution of multimorbidity clusters over 12 3.78; 95% CI 2.13–6.70), the heart diseases and cognitive years and includes both the change overtime of disease patterns impairment (OR 3.73; 95% CI 2.41–5.79), and neuropsychiatric (the diseases that characterize a specific cluster of individuals) and respiratory diseases (OR 4.30; 95% CI 2.95–6.27) clusters had and the migration of participants from one cluster to another. the highest OR for 6-year mortality, compared with the group of The height of the boxes and the thickness of the stripes in the people in the unspecific cluster. These clusters accounted for 57% figure are proportional to the amounts of people in the cluster of deaths in the following 6 years. and moving out from the cluster, respectively. In order to better characterize such transitions, we report in Discussion Figs. 2 and 3 the proportion of participants that were part of the Tracing the evolution of multimorbidity clusters and the clinical 6-year and 12-year follow-ups clusters and that moved from trajectories of older adults with multimorbidity overtime led to multimorbidity clusters detected at an earlier wave. The two major findings. The first was a high heterogeneity in the percentages of participants moving from baseline and 6-year multimorbidity clustering at baseline. Only half of the partici- clusters, to 6-year and 12-year clusters, respectively, are reported pants could be grouped into a well-characterized cluster: psy- in Supplementary Tables 4–7. During both first and second chiatric and respiratory diseases, heart diseases, respiratory and follow-up periods, the main shifts among clusters involved musculoskeletal diseases, cognitive and sensory impairment, and participants in the unspecific cluster, who moved primarily to eye diseases and cancer. The other half of the participants were clusters characterized by cardiovascular, eye, respiratory, and sorted into an unspecific cluster and were characterized by having musculoskeletal diseases. For example, persons in the unspecific a younger age, lower numbers of co-occurring diseases and drugs, group at baseline moved and represented 48.7%, 45.0%, and good functional and cognitive abilities, and a high percentage of 38.8% of the 6-year follow-up heart and vascular diseases, cardiovascular risk factors. The second major finding was a highly musculoskeletal, respiratory and immune diseases, and eye diseases dynamic evolution of multimorbidity clusters at both 6 and 12 clusters, respectively. Similarly, persons belonging to the years. Over 12 years, changes in cluster composition, participants’ unspecific group at the 6-year follow-up moved and represented transitions from one cluster to another, and participant mortality 49.5%, 49.1%, and 20.6% of the 12-year follow up cardiometabolic generated a dynamic but well-defined clinical picture. The first diseases, eye and musculoskeletal diseases, and vascular diseases remarkable trajectory involved the group of people part of the clusters, respectively. unspecific cluster at baseline. The number of participants grouped in this cluster halved at the 6- and 12-year follow-ups as the Different multimorbidity clusters confer different mortality majority transitioned toward the specific multimorbidity clusters risks. The association between the multimorbidity clusters and identified at follow-ups. Given the young age and less complex mortality was tested in logistic regression models adjusted by age, clinical picture of these individuals, they may be considered an NATURE COMMUNICATIONS | (2020)11:3223 | https://doi.org/10.1038/s41467-020-16780-x | www.nature.com/naturecommunications 3 ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-020-16780-x Psychiatric and resp. diseases 2.6 4.6 19.2 4.4 4.6 2.4 4.3 6.2 (0.0–14.3) (0.0–13.1) (13.2–25.5) (0.0–11.5) (0.0–11.6) (0.1–4.8) (0.9–7.9) (1.5–11.1) Heart diseases 19.7 41.4 2.6 1.8 1.3 0.5 18.0 8.7 (9.2–31.4) (33.6–50.0) (0.0–8.9) (0.0–8.9) (0.0–8.3) (0.0–3.0) (14.5–21.6) (4.0–13.6) Eye diseases and cancer 9.2 14.5 10.2 34.4 2.5 2.2 12.9 10.7 From baseline (0.0–20.9) (6.6–23.0) (4.2–16.4) (27.8–41.5) (0.0–9.5) (0.0–4.7) (9.5–16.5) (6.0–15.6) Cognitive and sensory impairment 3.9 2.0 12.5 4.0 0.8 0.5 28.7 4.7 (0.0–15.6) (0.0–10.5) (6.4–18.7) (0.0–11.1) (0.0–7.8) (0.0–3.0) (25.2–32.3) (0.0–9.6) Respiratory and MSK diseases 15.8 8.6 22.6 16.7 45.8 9.0 11.2 16.2 (5.3–27.5) (0.7–17.1) (16.6–28.9) (10.1–23.9) (39.5–52.8) (6.7–11.4) (7.7–14.8) (11.4–21.1) Unspecific 48.7 28.9 32.8 38.8 45.0 85.4 25.0 53.5 (38.2–60.4) (21.1–37–5) (26.8–39.1) (32.2–45.9) (38.7–52.0) (83.1–87.8) (21.5–28.6) (48.8–58.4) es ir. s es es fic th ut se pa ea po ci as as as ea pe im D ro se se se is ns D di di di ve .d U ar e e iti sp Ey un gn ul re sc m co d im va an d an d d an h. an es yc p. rt as ps ea ea se e, H ,r di tiv SK rt ni M ea og H C At 6 years Fig. 2 Contribution of the baseline multimorbidity clusters to the 6-year follow-up clusters. Numbers indicate the percentage (%) of people belonging to the 6-year follow-up clusters that moved from the baseline clusters. To note, for this analysis participants were assigned to the cluster they were more likely to belong. at-risk population for developing more complex multimorbidity data from a large sample of general practice showed that the more and as such potentially susceptible to preventive intervention. The diseases present at baseline, the higher the cumulative incidence second relevant trajectory was the high mortality of individuals in rates of one or more new diseases (up to 47% at the 3-year follow- clusters characterized by cardiovascular and neuropsychiatric up and 76% at the 5-year follow-up)20. Another study of a diseases, which, despite representing only 25%, 28%, and 29% of population-wide registry of more than six million patients in the participants at baseline, 6 years, and 12 years, respectively, Denmark showed more than a thousand significant longitudinal accounted for 51% and 57% of deaths during the first and second disease trajectories and some major multimorbidity clusters follow-up periods, respectively. characterized by diseases of the prostate, chronic obstructive Increasingly, studies are analyzing clusters of multimorbidity pulmonary disease, cerebrovascular disease, cardiovascular dis- across different populations, settings, and countries, but most ease, and diabetes mellitus. The study had the limitation of data studies have had a cross-sectional design or focused on the pro- drawn retrospectively from a hospital registry of primary and gression of co-morbidities of index diseases12,16,17. There is secondary diagnostic codes. Further, both chronic and acute scanty evidence of how clusters of multimorbidity change over- diseases were included21, making the findings difficult to compare time. The comparison is also limited by the fact that previous with ours. Finally, in an Australian study more than 13,000 studies have used primary care, hospital-based registries or self- middle-aged women with no history of diabetes, heart disease, or reported diagnoses, included only middle-aged people, or exam- stroke at baseline were followed for 20 years in order to evaluate ined both acute and chronic conditions. A study from Spain that the longitudinal progression of the three conditions. Over 20 used a similar analytical strategy on large data from electronic years, 18% of the women progressed to at least one condition, and primary health care records identified six clusters of multi- 16.8% had two or three of these conditions; moreover, the onset morbidity: musculoskeletal, endocrine-metabolic, digestive/ of stroke was more strongly associated with an increased risk of respiratory, neuropsychiatric, cardiovascular, and an unspecific progressing to the other two diseases. This is in contrast with group. These clusters exhibited less variation during the 6 years of what we observed in our study, which showed an opposite follow-up than the patterns identified in our study, which could transition, from cardiovascular risk factors (e.g., diabetes) to overt be explained by our longer follow-up period18. The use of elec- cardiovascular and neuropsychiatric diseases. In the same Aus- tronic health records may also have led to an under detection of tralian study, social inequality, obesity, hypertension, physical less severe diseases and multimorbidity19. A study from the inactivity, smoking, and other chronic conditions were sig- Netherlands focused on six cardiovascular conditions. Clinical nificantly associated with the three diseases independently but 4 NATURE COMMUNICATIONS | (2020)11:3223 | https://doi.org/10.1038/s41467-020-16780-x | www.nature.com/naturecommunications NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-020-16780-x ARTICLE Heart and vascular diseases 30.2 3.4 0.0 0.8 0.5 0.5 7.8 5.4 (19.0–43.3) (0.0–13.2) (0.0–8.5) (0.0–10.0) (0.0–7.8) (0.0–3.8) (3.4–12.5) (0.0–11.9) Heart diseases and cognitive impair. 7.9 26.7 1.7 3.3 1.5 0.0 19.5 6.3 (0.0–21.0) (18.1–36.5) (0.0–10.2) (0.0–12.5) (0.0–8.8) (0.0–3.3) (15.1–24.2) (0.0–12.8) Cognitive, psych. and resp. diseases 6.3 4.3 16.1 25.8 6.0 5.5 30.0 13.0 From 6 years (0.0–19.4) (0.0–14.1) (7.6–24.6) (16.7–35.0) (0.0–13.3) (2.3–8.9) (25.6–34.7) (6.7–19.5) Eye diseases 15.9 10.3 4.2 35.0 26.0 1.0 16.6 10.3 (4.8–29.0) (1.7–20.1) (0.0–12.8) (25.8–44.2) (19.0–33.3) (0.0–4.3) (12.2–21.3) (4.0–16.9) MSK, reap. and immune diseases 19.0 6.0 61.9 15.0 16.5 8.5 6.7 13.0 (7.9–32.1) (0.0–15.8) (53.4–70.4) (5.8–24.2) (9.5–23.8) (5.3–11.9) (2.3–11.4) (6.7–19.5) Unspecific 20.6 49.1 16.1 20.0 49.5 84.5 19.5 52.0 (9.5–33.7) (40.5–58.9) (7.6–24.6) (10.8–29.2) (42.5–56.8) (81.2–87.8) (15.1–24.2) (45.7–58.6) es es es es es fic th ut ea po ci as as as as as pe D ro se se se se se ns D di di di di di U ar ic ry ric SK ol to ul at M ab sc ra hi d pi et Va yc an es om ps R e ro di Ey ar eu C N At 12 years Fig. 3 Contribution of the 6-year follow-up multimorbidity clusters to the 12-year follow-up clusters. Numbers indicate the percentage (%) of people belonging to the 12-year follow-up clusters that moved from the 6-year follow-up clusters. To note, for this analysis participants were assigned to the cluster they were more likely to belong. also with their co-occurrence. The study used self-reported may have evolved into, or been correctly classified as, cognitive diagnoses22. impairment and dementia, putting them in the cognitive Some diseases may not be as independent of each other as we impairment, psychiatric and respiratory diseases cluster. have previously thought. Biological, health-care related (e.g., Finally, the unspecific cluster deserves special attention. These pharmacological treatment), and psychosocial factors may participants were characterized by diseases that were not over- increase susceptibility to a specific disease or to diseases in general represented. However, despite their younger age and better in an individual1,23. Such factors can systematically drive diseases physical and mental fitness, they had a high prevalence of car- clustering beyond chance as well as their evolution to other diovascular and metabolic risk factors (diabetes, obesity, dyslipi- clusters over time. First, direct consequences may explain why a demia, and hypertension). At baseline, almost half of the sample large number of people in the heart diseases cluster at baseline was part of this group. These people contributed to 29–49% of the became part of the heart diseases and cognitive impairment cluster multimorbidity clusters at the 6-year follow-up and to 16–50% of at 6 years. Extensive scientific evidence supports the association the multimorbidity clusters at 12 years, especially to those char- between heart disease and cognitive decline through different acterized by cardiovascular, eye, respiratory, and musculoskeletal mechanisms such as emboli, ischemic events, small vessel disease, diseases. Despite it is now well established that cardiometabolic cerebral hypoperfusion, and hypoxia. Indeed, mixed dementia, conditions such as diabetes, obesity, dyslipidemia, and hyper- resulting from both cerebrovascular lesions and neurodegenera- tension are important risk factors for the development of several tion, accounts for the majority of dementia cases among very old cardiovascular diseases, less is known about the same risk factors, individuals24. Second, treatment consequences are another pos- and the risk of other chronic conditions26,27. A few individuals sible pathway when a disease occurs as the result of the phar- moved from a specific cluster to the unspecific cluster over time. macological or surgical treatment of another condition. For This may be explained by the fact that the progressive accrual of example, part of the neuropsychiatric and respiratory diseases new diseases and the mortality (or dropout) of participants cluster, an association that remained over the entire course of our included in any of the specific clusters changed the reciprocal study, may be linked to the steroid treatment of respiratory dis- relation among diseases in survivors—in terms of prevalence, O/E eases. Steroid treatment can often cause neurotic disorders and ratio and exclusivity—making some of the subjects no longer depression25. Third, overlapping symptomatology may result in classifiable into a specific cluster. diseases being misdiagnosed in an initial phase. This may have At least four out of ten participants died over the course of the occurred with some baseline psychiatric conditions in the psy- study. Both at baseline and at 6-year follow-up, individuals with chiatric and respiratory diseases cluster, which by 6 or 12 years multimorbidity patterns characterized by cardiovascular and NATURE COMMUNICATIONS | (2020)11:3223 | https://doi.org/10.1038/s41467-020-16780-x | www.nature.com/naturecommunications 5 ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-020-16780-x neuropsychiatric diseases had the highest mortality; with adjusted odds ratios ranging between 1.60 and 6.00 (taking people in the 1.33 (0.89–2.00) 4.30 (2.95–6.27) 1.06 (0.67–1.70) 3.78 (2.13–6.70) 3.73 (2.41–5.79) OR (95% CI)* unspecified cluster as the reference). Those clusters accounted for 6–12 years 51% of deaths during the first follow-up and for 57% of deaths mortality during the second follow-up. Notably, at 6 years there were two Ref. clusters characterized by cardiovascular diseases. Cardiovascular and neuropsychiatric diseases—the former including diseases such as heart failure and coronary diseases and the latter including diseases such as dementia and depression—are frequent and burdensome chronic conditions in older adults and are among the most important determinants of years of life spent with disability28. This is in line with a previous study from our Events/at risk 37/76 (49) 143/265 (54) 79/227 (35) 93/152 (61) 93/758 (12) 32/238 (13) group, showing that neuropsychiatric disease clusters, especially when combined with one or multiple cardiovascular diseases, have the highest impact on function decline in older persons5. Such findings were confirmed in other studies as well29–31. Indeed, the high mortality of people belonging to neuropsychia- tric and heart disease clusters was not surprising as those clusters had the highest functional disability and lowest walking speed both at baseline and at first follow-up. Similar findings were reported also in studies from Spain13 and from the United Kingdom4. The authors of the first report found that, compared MSK, respiratory, and immune Heart diseases and cognitive Heart and vascular diseases with those subjects part of the musculoskeletal cluster, women in Multimorbidity clusters at Table 1 Association between clusters and mortality during the first (0–6 years) and second (6–12 years) follow-up. the cardiovascular clusters had the highest risk of dying. In the Neuropsychiatric and latter, co-occurring cardiometabolic disorders, unlike single dis- orders, decreased survival in a multiplicative way. It can be Unspecific group respiratory dis. argued that not all diseases included in the cardiovascular or Eye diseases neuropsychiatric clusters transmit the same mortality risk. In fact, impairment diseases the nature of diseases, their impact at the organism level, and 6 years their severity may play major prognostic roles13. However, pre- vious studies conducted in the field of associative multimorbidity have shown that the group-specific effect of clusters of diseases remains regardless of the role played by single diseases5. The main strength of this study was the thorough clinical evaluation that underlay disease assessment. Each participant in SNAC-K undergoes a 5 h comprehensive assessment that follows 1.23 (0.90–1.68) 1.29 (0.96–1.74) 6.00 (4.21–8.54) 3.07 (2.26–4.19) a standard protocol and is carried out by a physician, a nurse, and 1.60 (1.02–2.51) OR (95% CI)* a psychologist. We then categorized diseases using a strict clini- 0–6 years cally driven method developed and tested by our group32. Fur- mortality thermore, the lack of missing information on disease status Ref. increases the internal validity of our study. Another major strength of this study was the statistical method, which allowed us To note, for this analysis participants were assigned to the cluster they were more likely to belong. to cluster people by their co-occurring diseases. We took advantage of the method to follow individuals overtime and track their trajectories. The fuzzy c-means cluster algorithm is the choice method for pattern recognition when clusters tend to overlap, which is often the case as older adults present high 91/456 (20) 105/305 (34) 233/306 (76) 35/159 (22) 146/277 (53) 203/1428 (14) prevalence of co-occurring conditions. In contrast to previous Events/at studies, each participant was assigned a probability of belonging risk (%) to a cluster without being forced to be part of it exclusively. Other strengths included the long follow-up time, the high number of OR odds ratio; CI confidence interval; MSK musculoskeletal. very old people, and the large age span of the participants (60–104 years). Moreover, including both mental and physical conditions in the analyses gave us the opportunity to investigate the interplay, potentially bidirectional, between mental health Asterisk adjusted for age, sex, and education. problems and chronic physical conditions. Several limitations of Respiratory and MSK diseases Cognitive and sensory impair. the present study should be mentioned. First, diseases were Multimorbidity clusters at Psychiatric and respiratory considered regardless their severity. Disease severity may indeed Eye diseases and cancer partially explain the clinical trajectories described in the present study. However, the interaction among different comorbidities Unspecific group still seems to play a major role—as it has been shown by us and Heart diseases others in previous studies—even when measures of disease severity are taken into account4,5,31,33. Moreover, in our opinion, baseline diseases independently from disease severity, the insights on the natural evolution of multimorbidity provided in this study are highly valuable and cover an important knowledge gap left by previous 6 NATURE COMMUNICATIONS | (2020)11:3223 | https://doi.org/10.1038/s41467-020-16780-x | www.nature.com/naturecommunications NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-020-16780-x ARTICLE cross-sectional studies. Further, there is evidence that the burden pathophysiology, development of diseases as a consequence of other of specific conditions changes depending on the overall multi- conditions or treatments, and symptomatic overlap among diseases morbidity status of one individual, making it difficult—especially underlie most of the trajectories identified. Although the ability to in older individuals—to ascertain the relevance of single disease discriminate among the potential mechanisms underlying the co- severity. Second, the dropout rate of participants (14% at 6 years occurrence of multiple chronic diseases needs further improvement, and 8% at 12 years) may have affected cluster definition. How- taking into account multimorbidity clusters, and their evolution ever, to the best of our knowledge, this is an exceptionally low overtime may enable better decisions for patients with multi- figure compared with studies of this type. Third, the dis- morbidity at every health care level and better tailoring of the target continuous follow-up carried out in SNAC-K—every 3 or 6 years population in future interventions. —may have affected disease detection and consequently the cluster analysis, especially among people who died or dropped out Methods during the observation period. Finally, the average high socio- Study population. We used longitudinal data from the population-based SNAC- economic status of participants in SNAC-K may potentially limit K39. The study population consists of adults ≥60 years living in the community or the generalizability of the findings. in institutions in the Kungsholmen district of Stockholm, Sweden. A random sample of 11 age cohorts born between 1892 and 1939 (the youngest and oldest age Over their life course, individuals develop multiple diseases. cohorts were oversampled) was invited to participate in the study. People who This challenges the current organization of medical care services agreed to participate were evaluated for the first time between 2001 and 2004. and the traditional research approach based on single diseases. Participants who were <78 years of age were then followed up every 6 years and Programs that bridge multiple clinical specialties and health care participants ≥78 years every three years. The present study is based on data col- lected at baseline, 6 years, and 12 years. At baseline, 3363 people were examined units should be developed to focus on single individuals, their (participation rate 73%). Overall, 432 participants were excluded because they did specific clinical profiles, and their specific clinical trajectories34. not have multimorbidity (≥2 chronic diseases) at baseline. The study was approved Knowing how diseases cluster together, and importantly, how the by the Regional Ethics Review Board in Stockholm. Participants in the study clinical status of people with multimorbidity can change over provided written informed consent. For participants with prevalent or incident subsequent years helps not only in understanding the complexity cognitive impairment, written informed consent was obtained from the next of kin. The present study was reported in keeping with the STrengthening the Reporting and dynamic evolution of multimorbidity clusters but also in of OBservational studies in Epidemiology recommendations. supporting clinicians who manage co-occurring chronic diseases and health policy makers who plan care resources use. The Chronic diseases. At each study wave, SNAC-K participants undergo an ~5 h-long findings from our study contribute in many ways. Firstly, they comprehensive clinical and functional assessment carried out by trained physi- help identify people at high risk of progressing to severe disease cians, nurses, and neuropsychologists. Physicians collect information on diagnoses clusters with worse prognosis. The people who could not be via physical examination, medical history, examination of medical charts, self- grouped in any specific cluster are at risk of cumulating further reported information, and/or proxy interviews. Clinical parameters, lab tests, drug information, and inpatient and outpatient care data are also used to identify spe- chronic disorders and increasing the severity of their multi- cific conditions. All diagnoses are coded in accordance with the International morbidity profile. However, 28% of the people in this group Classification of Diseases, 10th revision (ICD-10). In the current study we sorted remained relatively healthy during follow-ups. They had the the ICD-10 codes into 60 chronic disease categories in accordance with a clinically lowest numbers of co-occurring chronic diseases and drugs and a driven methodology (Tables S2 and S3)32. To avoid statistical noise and the resulting spurious findings in the models, we excluded diseases with a prevalence of better functional status than people in specific multimorbidity <2%. In SNAC-K at each study wave, drugs are coded in accordance with the clusters, providing a large time window for preventive interven- Anatomical Therapeutic Chemical classification. tion. Future studies should focus on promotion of healthy aging in this group of individuals. Our findings contribute secondly to Vital status and loss to follow-up. Information about vital status was derived the development of personalized medicine in multimorbidity as from death certificates provided by Statistics Sweden, the Swedish governmental our analysis is based on individuals and not diseases. There is statistics agency. Survival status was assessed throughout the follow-up period. solid evidence that persons who are affected by multimorbidity, Participants were considered lost to follow up if they or a proxy declined to participate, could not be contacted, had moved out of the study area, or canceled an face complex treatments, and require continuous monitoring far assessment. better from primary care with a patient-centered approach35. The strong transition we found from heart to brain diseases gives Other variables. Information on demographics (age, sex, and education) was impetus to efforts in primary care to treat and monitor patients collected during nurse interviews. We divided education into elementary, sec- affected by heart disease. Treatment adherence is very low among ondary, university, or higher. Level of disability was measured as the sum of the older people with multimorbidity and heart diseases in parti- basic and instrumental activities of daily living (ADL and IADL) a person was cular36. Thirdly, our findings support prognostic counseling for unable to perform independently40. The six ADLs were bathing, dressing, toileting, continence, transferring, and eating. The eight IADLs were grocery shopping, meal patients and caregivers, given the high mortality of people with preparation, housekeeping, doing laundry, managing money, using the telephone, co-occurring mental and cardiovascular disorders. Fourthly, our taking medications, and using public transportation. Walking speed (m/s) was findings encourage the planning of future randomized clinical assessed by asking participants to walk 6 m at their usual speed or 2.44 m if the trials toward the better management of multimorbidity. The 3D participant reported walking quite slowly. Speeds of <0.8 m/s were categorized as impaired41. Cognitive status was assessed by physicians using the MMSE, with a approach recently proposed by Salisbury et al. is an example of an score range of 30 at best to 0 at worst42. intervention that could have focused on those multimorbidity clusters that may most likely lead to negative health outcomes Statistical analysis. Sample characteristics at baseline, 6-year follow-up, and (neuropsychiatric and cardiovascular clusters)37. In this prag- 12-year follow-up were described for each multimorbidity cluster using weighted matic trial, the target population was selected based exclusively on means and proportions obtained by the membership matrix (see below). At each the number of diseases and did not take into account specific study wave, clusters of older adults who shared patterns of multimorbidity were groups of diseases. This may explain why the intervention was independently identified using the fuzzy c-means cluster analysis algorithm, which belongs to the family of soft clustering algorithms. The algorithm estimates c cluster not able to improve participants’ quality of life38. centers (similar to k-means) but with fuzziness so that individuals may belong to In conclusion, clinical trajectories of older adults with multi- more than one cluster. The use of a fuzzy cluster analysis over a hard cluster morbidity are characterized by great dynamism and complexity but analysis helps to better handle the stochastic nature of some disease association, the can still be tracked over time. By analyzing data from a large potential noise stemming from the measurement (e.g., disease assessment), and the variance due to between-individual differences. Through this technique, we population-based study of people aged 60+ years, we were able to obtained clusters of individuals and a membership matrix that indicated the degree identify multimorbidity clusters, trace their evolution overtime, and of participation of each subject in each cluster. In a second step, to evaluate the follow individuals’ trajectories over 12 years. Shared risk factors and most likely clinical trajectories of the participants as they moved among clusters NATURE COMMUNICATIONS | (2020)11:3223 | https://doi.org/10.1038/s41467-020-16780-x | www.nature.com/naturecommunications 7 ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-020-16780-x over time, each individual was assigned to the cluster with the highest membership 2. Marengoni, A. et al. Aging with multimorbidity: a systematic review of the score at each time point. We used dimensionality reduction techniques (multiple literature. Ageing Res. Rev. 10, 430–439 (2011). correspondence analysis) to obtain the input data for clustering the participants. 3. Salisbury, C. et al. Epidemiology and impact of multimorbidity in primary The Karlis–Saporta–Spinaki rule was used to decide how many dimensions to care: a retrospective cohort study. Br. J. Gen. Pract. 61, e12–e21 (2011). retain43. The main parameters used during our cluster analysis were the number of 4. Di Angelantonio, E. et al. Association of cardiometabolic multimorbidity with clusters and a fuzziness parameter, denoted as “m”, which ranges from just above 1 mortality. Jama 314, 52–60 (2015). to infinity. High m values produce a fuzzy set of c clusters, so that individuals are 5. Vetrano, D. L. et al. Trajectories of functional decline in older adults with equally distributed across clusters, whereas lower ones generate non-overlapped neuropsychiatric and cardiovascular multimorbidity: a Swedish cohort study. clusters. In our study we checked m = 1.1, 1.2, 1.4, 1.5, 2, 4 over 1 to 20 cluster PLoS Med. 15, e1002503 (2018). combinations; the value m = 1.1 over performed the rest of values. Since clustering 6. Dumbreck, S. et al. Drug-disease and drug-drug interactions: systematic algorithms are unsupervised techniques, the model fitting the dataset is tradi- examination of recommendations in 12 UK national clinical guidelines. BMJ tionally computed through cost functions that depend on both the dataset and the 350, h949 (2015). clustering parameters and are denoted as validation indices. We computed dif- 7. Vetrano, D. L. et al. An international perspective on chronic multimorbidity: ferent validation indices to obtain the optimal number of clusters c and the optimal approaching the elephant in the room. J. Gerontol. A Biol. Sci. Med. Sci. 73, value of the fuzziness parameter m. Included parameters were: the Fukuyama index 1350–1356 (2018). (optimal when presenting low values), the Xie–Beni index (optimal when pre- 8. Farmer, C. et al. Clinical assessment and management of multimorbidity: senting low values), the Partition coefficient index (optimal when presenting high summary of NICE guidance. BMJ 354, i4843 (2016). values), the Partition entropy index (optimal when presenting low values), and the 9. Muth, C. et al. Evidence supporting the best clinical management of patients Calinski–Harabasz index (optimal when presenting high values; Supplementary with multimorbidity and polypharmacy: a systematic guideline review and Figs. 1–15)44. Given the stochastic nature of the clusters, we ran 100 independent expert consensus. J. Intern. Med. 285, 272–288 (2019). clustering repetitions to obtain the average final solution. We based our evaluation of the consistency and significance of the final solution on clinical criteria. To 10. Marengoni, A. & Onder, G. Guidelines, polypharmacy, and drug-drug cross-validate the model, we randomly split the individuals into two independent interactions in patients with multimorbidity. BMJ 350, h1059 (2015). data sets and compared their validation indices. Indices were computed and 11. Vetrano D. L., et al. Walking Speed Drives the Prognosis of Older Adults with averaged over 100 repetitions. To characterize the clusters of multimorbidity that Cardiovascular and Neuropsychiatric Multimorbidity. Am. J. Med. https://doi. corresponded to each cluster of individuals, we calculated the frequency of chronic org/10.1016/j.amjmed.2019.05.005. (2019). diseases in each cluster. Observed/expected ratios (O/E-ratios) were calculated by 12. Prados-Torres, A. et al. Multimorbidity patterns: a systematic review. J. Clin. dividing the prevalence of a given disease within a cluster by its prevalence in the Epidemiol. 67, 254–266 (2014). overall population. The exclusivity of different diseases, defined as the fraction of 13. Ibarra-Castillo, C. et al. Survival in relation to multimorbidity patterns in participants with the disease in the cluster over the total number of participants older adults in primary care in Barcelona, Spain (2010–2014): a longitudinal with the disease, was also calculated. We considered a disease to be associated with study based on electronic health records. J Epidemiol Community Health 72, a given cluster of individuals when the O/E ratio was ≥2 or the exclusivity was 185–192 (2018). ≥25%18. Such criteria were used to name multimorbidity clusters after the diseases 14. Westergaard, D. et al. Population-wide analysis of differences in disease that mostly characterized them. To note, due to the dynamism of the phenomenon, progression patterns in men and women. Nat. Commun. 10, 666 (2019). the names of the clusters change overtime, reflecting the evolving combinations of 15. Bierman, A. S. & Tinetti, M. E. Precision medicine to precision care: managing diseases that characterize them at each time point. Shifts between clusters were multimorbidity. Lancet 388, 2721–2723 (2016). computed by cross-tabulating individuals between each wave (baseline to 6-year 16. Goodman, R. A. et al. Multimorbidity patterns in the United States: follow-up and 6-year to 12-year follow-up) after assigning them individuals to the implications for research and clinical practice. J. Gerontol. A Biol. Sci. Med. Sci. cluster where they were more likely to belong. In this way, we analyzed the most 71, 215–220 (2016). likely individual trajectories. Frequencies (percentages) of participants who chan- 17. Petri, H., Maldonato, D. & Robinson, N. J. Data-driven identification of co- ged from one cluster to another were computed to assess the overlap between morbidities associated with rheumatoid arthritis in a large US health plan waves. Both column percentages and row percentages are provided in Supple- claims database. BMC Musculoskelet. Disord. 11, 247 (2010). mentary Tables. Mortality and dropout status were considered as fixed clusters in 18. Guisado-Clavero, M. et al. Multimorbidity patterns in the elderly: a both 6-year and 12-year follow-ups. Logistic regression models adjusted by age, sex prospective cohort study with cluster analysis. BMC Geriatr. 18, 16 (2018). and education were fitted to estimate the association between clusters and mor- 19. Rizzuto, D. et al. Detection of Dementia cases in two Swedish health registers: tality, using the unspecific cluster as the reference group. Also in this case, parti- a validation study. J. Alzheimers Dis. 61, 1301–1310 (2018). cipants were assigned to the cluster where they were more likely to belong. Odd 20. Lappenschaar, M. et al. Multilevel temporal Bayesian networks can model ratios (OR) and 95% confidence intervals (CI) were adjusted for age, sex, and longitudinal change in multimorbidity. J. Clin. Epidemiol. 66, 1405–1416 education. All comparisons were adjusted for multiplicity. Pairwise comparison of (2013). p values, corrected for multiple comparisons, was calculated. Tukey method were 21. Jensen, A. B. et al. Temporal disease trajectories condensed from population- used when the explanatory variable was normal-distributed or Benjamini and wide registry data covering 6.2 million patients. Nat. Commun. 5, 4022 (2014). Hochberg method otherwise45. The significance level was set at p = 0.05. Although 22. Xu X, Mishra GD, Dobson AJ, & Jones M. Progression of diabetes, heart the overall number of significant tests between clusters at each follow-up remained disease, and stroke multimorbidity in middle-aged women: A 20-year cohort stable at each follow-up, the number of highly significant pairwise statistical test study. PLoS Med. 15, e1002516 (2018). (i.e., p < 0.001) decreased from 60.0 to 36.7%. Statistical analyses were performed 23. van den Akker, M. et al. Multimorbidity in general practice: prevalence, using R 3.5.1 and Stata 15. Codes are available on demand. incidence, and determinants of co-occurring chronic and recurrent diseases. J. Clin. Epidemiol. 51, 367–375 (1998). Reporting summary. Further information on research design is available in 24. Qiu, C. & Fratiglioni, L. A major role for cardiovascular burden in age-related the Nature Research Reporting Summary linked to this article. cognitive decline. Nat. Rev. Cardiol. 12, 267–277 (2015). 25. Herbert, J. Cortisol and depression: three questions for psychiatry. Psychol. Data availability Med. 43, 449–469 (2013). The source data underlying all the figures and tables (including supplementary ones) is 26. Marseglia, A. et al. Prediabetes and diabetes accelerate cognitive decline and represented by the SNAC-K project, a population-based study on aging and dementia predict microvascular lesions: a population-based cohort study. Alzheimers (http://www.snac-k.se/). Access to these original data is available to the research Dement. https://doi.org/10.1016/j.jalz.2018.06.3060 (2018). community upon approval by the SNAC-K data management and maintenance 27. Liu, Y. et al. Hypertension-induced cerebral small vessel disease leading to committee. Applications for accessing these data can be submitted to Maria Wahlberg cognitive impairment. Chin. Med. J. 131, 615–619 (2018). (Maria.Wahlberg@ki.se) at the Aging Research Center, Karolinska Institutet. 28. Rizzuto, D. et al. Effect of chronic diseases and multimorbidity on survival and functioning in elderly adults. J. Am. Geriatr. Soc. 65, 1056–1060 (2017). 29. Quinones, A. R., Markwardt, S. & Botoseneanu, A. Multimorbidity Received: 14 February 2019; Accepted: 22 May 2020; combinations and disability in older adults. J. Gerontol. A Biol. Sci. Med. Sci. 71, 823–830 (2016). 30. Quinones, A. R. et al. Prospective disability in different combinations of somatic and mental multimorbidity. J. Gerontol. A Biol. Sci. Med. Sci. 73, 204–210 (2018). 31. Jindai, K. et al. Multimorbidity and functional limitations among adults 65 or References older, NHANES 2005-2012. Prev Chronic Dis. 13, E151 (2016). 1. Calderon-Larranaga, A. et al. Multimorbidity and functional impairment- 32. Calderon-Larranaga, A. et al. Assessing and measuring chronic bidirectional interplay, synergistic effects and common pathways. J. Intern. multimorbidity in the older population: a proposal for its operationalization. J. Med. 285, 255–271 (2019). Gerontol. A Biol. Sci. Med. Sci. 72, 1417–1423 (2017). 8 NATURE COMMUNICATIONS | (2020)11:3223 | https://doi.org/10.1038/s41467-020-16780-x | www.nature.com/naturecommunications NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-020-16780-x ARTICLE 33. Jackson, C. A. et al. Multimorbidity patterns are differentially associated with Research Council for Health, Working life and Welfare (Forte; 2016-07175; 2017-01764); functional ability and decline in a longitudinal cohort of older women. Age Gamla Tjanarinnor (2019-00897), and the Ermenegildo Zegna Foundation. The funders Ageing 44, 810–816 (2015). had no role in study design, data collection and analysis, decision to publish, or pre- 34. Hemingway, H. et al. Prognosis research strategy (PROGRESS) 1: a paration of the paper. Open access funding provided by Karolinska Institute. framework for researching clinical outcomes. BMJ 346, e5595 (2013). 35. National Institute for Health and Care Excellence (NICE). Multimorbidity: clinical assessment and management. (2016). https://www.nice.org.uk/ Author contributions Conception or design of the work: D.L.V., A.R.L., A.C.L., S.F., C.V., A.M. Data analysis: guidance/ng56/resources/multimorbidity-clinical-assessment-and- A.R.L., S.F., D.L.V., A.C.L. Interpretation of the results: D.L.V., A.R.L., A.C.L., S.F., C.V., management-pdf-1837516654789 Accessed on 1st January 2020. A.M., M.G.C., G.O., L.F. Drafting the article: D.L.V., A.R.L., A.C.L., A.M. Critical revision 36. Ruppar, T. M. et al. Medication adherence interventions improve heart failure of the paper: D.L.V., A.R.L., A.C.L., S.F., C.V., A.M., M.G.C., G.O., L.F. Final approval of mortality and readmission rates: systematic review and meta-analysis of the paper: D.L.V., A.R.L., A.C.L., S.F., C.V., A.M., M.G.C., G.O., L.F. All the authors fulfill controlled trials. J Am. Heart Assoc. 5. https://doi.org/10.1161/ the ICMJE criteria for authorship. jaha.115.002606. (2016). 37. Salisbury, C. et al. Management of multimorbidity using a patient-centred care model: a pragmatic cluster-randomised trial of the 3D approach. Lancet 392, Competing interests 41–50 (2018). The authors declare no competing interests. 38. Marengoni, A., Vetrano, D. L., & Onder, G. Target population for clinical trials on multimorbidity: is disease count enough? J. Am. Med. Dir. Assoc. https://doi.org/10.1016/j.jamda.2018.10.012. (2018) Additional information 39. Lagergren, M. et al. A longitudinal study integrating population, care and Supplementary information is available for this paper at https://doi.org/10.1038/s41467- social services data. The Swedish National study on Aging and Care (SNAC). 020-16780-x. Aging Clin. Exp. Res. 16, 158–168 (2004). 40. Spector, W. D. & Fleishman, J. A. Combining activities of daily living with Correspondence and requests for materials should be addressed to D.L.V. instrumental activities of daily living to measure functional disability. J. Gerontol. B Psychol. Sci. Soc. Sci. 53, S46–S57 (1998). Peer review information Nature Communications thanks Anders Jensen, Marjan Van 41. Studenski, S. et al. Gait speed and survival in older adults. Jama 305, 50–58 (2011). Den Akker, and the other, anonymous, reviewers for their contribution to the peer review 42. Folstein, M. F., Folstein, S. E. & McHugh, P. R. “Mini-mental state”. A of this work. Peer reviewer reports are available. practical method for grading the cognitive state of patients for the clinician. J. Psychiatr. Res. 12, 189–198 (1975). Reprints and permission information is available at http://www.nature.com/reprints 43. Sourial, N. et al. Correspondence analysis is a useful tool to uncover the relationships among categorical variables. J. Clin. Epidemiol. 63, 638–646 Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in (2010). published maps and institutional affiliations. 44. Zhao Q. Cluster validity in clustering methods. Publications of the University of Eastern Finland. (2012) http://cs.joensuu.fi/sipu/pub/qinpei-thesis.pdf. Accessed on 1st January 2020. Open Access This article is licensed under a Creative Commons 45. Benjamini, Y. & Hochberg, Y. Controlling the false discovery rate: a practical Attribution 4.0 International License, which permits use, sharing, and powerful approach to multiple testing. J. R. Stat. Soc. Ser. B. 57, 289–300 adaptation, distribution and reproduction in any medium or format, as long as you give (1995). appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless Acknowledgements indicated otherwise in a credit line to the material. If material is not included in the We thank the SNAC-K participants and the SNAC-K Group for their collaboration in article’s Creative Commons license and your intended use is not permitted by statutory data collection and management, and scientific editors Kimberly Kane and Karen regulation or exceeds the permitted use, you will need to obtain permission directly from Hagersten for useful comments on the text. This work was supported by the funders of the copyright holder. To view a copy of this license, visit http://creativecommons.org/ the Swedish National study on Aging and Care (SNAC): the Ministry of Health and licenses/by/4.0/. Social Affairs, Sweden; the participating County Councils and Municipalities; and the Swedish Research Council. Specific grants were received from The Swedish Research Council for Medicine (VR; 521-2013-8676; 2017-06088; 2016-00981); the Swedish © The Author(s) 2020 NATURE COMMUNICATIONS | (2020)11:3223 | https://doi.org/10.1038/s41467-020-16780-x | www.nature.com/naturecommunications 9