{"paper_id":"0658428c-9b3d-461e-9411-9356e00f6b05","body_text":"The lifetime accumulation of multimorbidity and its influence on dementia risk: a UK\nBiobank Study\nPatel, R.ab, Mackay, C.E.ab, Griffanti, L. ab, Gillis, G. ab, Ebmeier, K.P .b, Suri, S. ab\na Oxford Centre for Human Brain Activity, Wellcome Centre for Integrative Neuroimaging, University of Oxford,\nOxford, UK\nb Department of Psychiatry, University of Oxford, Oxford, UK\nCorresponding Author:\nDr Raihaan Patel\nOxford Centre for Human Brain Activity\nWellcome Centre for Integrative Neuroimaging\nDepartment of Psychiatry\nUniversity of Oxford\nOxford, UK\nEmail: mohammed.patel@psych.ox.ac.uk\n1\n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted January 23, 2024. ; https://doi.org/10.1101/2024.01.21.24301584doi: medRxiv preprint \nNOTE: This preprint reports new research that has not been certified by peer review and should not be used to guide clinical practice.\n\nAbstract\nThe number of people living with dementia worldwide is projected to reach 150 million by 2050,\nmaking prevention a crucial priority for health services 1. The co-occurrence of two or more\nchronic health conditions, termed multimorbidity, occurs in up to 80% of dementia patients 2,\nraising the potential of multimorbidity as an important risk factor for dementia. However, precise\nunderstanding of which specific conditions, as well as their age of onset, drive the link between\nmultimorbidity and dementia is unclear. We defined the patterns of accumulation of 46 chronic\nconditions over their lifetime in 282,712 individuals from the UK Biobank. By grouping\nindividuals based on their life-history of chronic illness, we show here that risk of incident\ndementia can be stratified by both the type and timing of their accumulated chronic conditions.\nWe identified several distinct clusters of multimorbidity, and their associated risks varied in an\nage-specific manner. Compared to low multimorbidity, cardiometabolic and neurovascular\nconditions acquired before 55 years were most strongly associated with dementia. Acquisition of\nmental health and neurovascular conditions between the ages of 55 and 70 was associated with\nan over two-fold increase in dementia risk compared to low multimorbidity. The age-dependent\nrole of multimorbidity in predicting dementia risk could be used for early stratification of\nindividuals into high and low risk groups and inform targeted prevention strategies based on a\nperson’s prior history of chronic disease.\n2\n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted January 23, 2024. ; https://doi.org/10.1101/2024.01.21.24301584doi: medRxiv preprint \n\nMain\n50 million people worldwide are diagnosed with dementia, and this number is projected to grow\nrapidly. The global increase in life expectancy makes dementia prevention an urgent public\nhealth priority 3. Up to 80% of dementia patients have two or more other chronic health\nconditions by the time they receive their diagnosis 2. This makes multimorbidity a potentially\ncrucial target for optimising dementia treatment and possibly prevention on a population level.\nMultimorbidity has significant impacts on patients, their families, and health systems.\nIndividuals with multimorbidity are more likely to use health services 4,5, have impaired\nfunctional status 6, and are at higher risk for mortality 7. Like dementia, multimorbidity is a global\nhealth concern that is increasingly common in older age 8. The rising prevalence of\nmultimorbidity has driven prominent and pressing calls to action for the medical sciences to\nmove away from single disease focused models of treatment and research 9–11.\nHowever, despite the fact that dementia has consistently been associated with multimorbidity\n12,13, there are large gaps in our understanding of this relationship. For instance, multimorbidity\nhas most commonly been quantified as a binary classification (yes/no), or as a count of the\nnumber of pre-existing conditions 11,14. Neither framework considers interactions between\nconditions, or recognizes that patterns of co-occurring conditions can cluster in a systematic,\npredictable fashion 9. Advancing beyond a simplistic measurement of multimorbidity has\npreviously been limited by lack of large datasets with linked electronic health records. Moreover,\nmany studies do not examine a comprehensive range of conditions; a review of 556\nmultimorbidity studies reported the median number of included conditions to be only 17, with\n3\n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted January 23, 2024. ; https://doi.org/10.1101/2024.01.21.24301584doi: medRxiv preprint \n\nonly 8 conditions consistently included (diabetes, stroke, cancer, chronic obstructive pulmonary\ndisease, hypertension, coronary heart disease, chronic kidney disease, and heart failure) 14.\nFinally, the assessment of longitudinal trajectories of multimorbidity is lacking, even in those\nstudies which include a relatively comprehensive list of conditions 10,11. Although multimorbidity\nis increasingly common in older age, it is not solely confined to late life 8,15. For example, in a\nScottish population, nearly two-thirds of 65-84 year olds were multimorbid but almost a third of\n45 - 64 year olds were also identified as multimorbid 8. Therefore, characterising the sequential\naccumulation of multimorbidity patterns throughout the lifespan is a necessary step in order to\nbring our understanding of multimorbidity as a risk factor in line with the current life-course\nperspective of other dementia risk factors 16.\nHere, we address these challenges by characterising the sequential patterns of multimorbidity\nthroughout the life course. We use the UK Biobank, a large population cohort of over 502,000\nindividuals17. The baseline assessment was conducted in the late 2000s, when participants were\n40 to 69 years old. Importantly, the UK Biobank has continuously updated linked electronic\nhealth records since the baseline assessment occurred, allowing us to study chronic conditions\ndiagnosed prior to the baseline assessment as well as up to an average of 12 years after. To fully\ntake advantage of this unique resource, we identify patterns of co–occurring chronic conditions\n(termed multimorbidity clusters) in each of three distinct age ranges: 0-55 years, 55 - 65 years,\nand 65-70 years. Within each age range, we apply a clustering procedure which classifies\nparticipants into distinct multimorbidity clusters based on their diagnoses within the given\nportion of the lifespan. This enables us to identify the patterns of chronic conditions which tend\nto co-occur across the study population in each age range, cluster participants in distinct fashion\n4\n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted January 23, 2024. ; https://doi.org/10.1101/2024.01.21.24301584doi: medRxiv preprint \n\nbased on their unique accumulated conditions within each age range, and study how participants\ntransition between different multimorbidity clusters across the lifecourse. This framework\nallows us to answer three key questions: 1) how do patterns of multimorbidity change during the\nlifespan?; 2) do different multimorbidity patterns confer different levels of risk for future\ndementia?; and 3) how does multimorbidity accumulate over the life-course in dementia\npatients? (Figure 1)\n5\n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted January 23, 2024. ; https://doi.org/10.1101/2024.01.21.24301584doi: medRxiv preprint \n\nFigure 1: Tracking lifetime accumulation of multimorbidity: Study overview showing the\nderivation of lifetime multimorbidity patterns for two sample subjects. A) For each participant,\nwe ascertain the diagnosis of a variety of chronic conditions using linked electronic health\nrecords. Based on the age of onset, diagnoses are considered in only one of three distinct age\nranges: 0-55 years, 55-65 years, and 65-70 years, representing early, mid-, and late-life. B)\nWithin each age range, a clustering analysis is performed. At age 55, participants are clustered\ninto groups based on their conditions diagnosed from birth to age 55. At age 65, participants are\nclustered based on their conditions diagnosed between ages 55 and 65. At age 70, participants\nare clustered based on their conditions diagnosed between ages 65 and 70. C) We track the\nmovement of participants between clusters across the different age ranges. Participants may\nmove between clusters describing different multimorbidity patterns (eg gastrointestinal ->\nrespiratory), or, may move between clusters defined by similar conditions (eg. cardiovascular ->\ncardiovascular), depending on their unique pattern of accumulated conditions.\n6\n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted January 23, 2024. ; https://doi.org/10.1101/2024.01.21.24301584doi: medRxiv preprint \n\nResults\nWe tracked the movement of participants across distinct life-course patterns of multimorbidity. In\neach of the three age ranges assessed, we found multimorbidity clusters defined by\ncardiometabolic, neurovascular, and mental health conditions. Additionally, in the 0-55 age range\nonly, we also identified a cluster of eye conditions, while a peripheral vascular cluster was found\nin both of the later age ranges. Crucially, we not only show that dementia risk is most associated\nwith cardiovascular, mental health, and neurological related multimorbidity, but also that the\nodds of developing dementia varies based on when in the lifespan these conditions are\ndiagnosed. Prior to midlife (55 years of age), development of cardiometabolic conditions was\nassociated with the highest risk of incident dementia. However this changed from mid- to\nlate-life; in the 55-65 and 65-70 age ranges, the accumulation of mental health and neurovascular\nconditions was associated with the highest risk, a two-fold increased risk of dementia. The\ndementia-specific trajectories of multimorbidity identified in this study could be used for early\nstratification of individuals into high and low risk groups and inform targeted prevention\nstrategies based on a person’s prior history of chronic disease.\nParticipants\nWe used data from 282,712 UK Biobank 17 participants with available medical records after the\nage of 65, who were dementia free at age 65, and who were diagnosed with at least one chronic\ncondition other than dementia. The mean age at baseline was 61.6 (SD=4.65), and 150,687\nparticipants were female (53.3%). After the age of 65, 6,922 (2.45%) participants had developed\ndementia. Table 1 displays demographic statistics stratified by incident dementia status. The\nmajority of dementia patients (70.75%) had at least four other conditions diagnosed by the time\n7\n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted January 23, 2024. ; https://doi.org/10.1101/2024.01.21.24301584doi: medRxiv preprint \n\nthey received a diagnosis of dementia, compared to 51.78% of controls having four or more\nchronic conditions diagnosed throughout the study period.\nTable 1: Demographic characteristics of the analysis sample. Mean and standard deviation are given for\nage at the baseline UK biobank assessment, and years of education. For sex and number of conditions, the\nquantity and percentage of the sample is listed.\nControls (N = 275790) Dementia (N = 6922)\nAge at baseline, years\n(SD) 61.51 (4.65) 65.29 (3.24)\nFemale (%) 147,409 (53.45%) 3,278 (47.36%)\nEducation, years (SD) 13.13 (3.07) 12.40 (3.05)\nNumber of chronic conditions (%)\n1 46,792 (16.97%) 496 (7.17%)\n2 45,390 (16.46%) 749 (10.82%)\n3 40,806 (14.80%) 780 (11.27%)\n>=4 142,802 (51.78%) 4,897 (70.75%)\nAge-specific patterns of multimorbidity\nTo estimate life-course patterns of multimorbidity, we studied 46 chronic conditions diagnosed\nwithin three distinct age ranges: 0-55 years, 55-65 years, and 65-70 years. Within each age range,\nwe clustered participants based on new chronic conditions accumulated only within that age\nrange (Figure 1). We employed a two-stage multivariate clustering procedure (Methods) which\ncreated clusters based on chronic conditions which co-occurred together in a non-random,\nsystematic fashion. The resulting multimorbidity clusters delineate participants into distinct\ngroups, each with a unique pattern of chronic conditions diagnosed within a given age range.\n8\n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted January 23, 2024. ; https://doi.org/10.1101/2024.01.21.24301584doi: medRxiv preprint \n\nOutput clusters are named according to prevalence of, and comorbidity between, chronic\nconditions within the cluster relative to patterns in the full sample (Methods). Five clusters were\nidentified as the best fit within each age range.\nMultimorbidity from 0-55 years\nThe five multimorbidity clusters are described as: 1) a relatively healthy group of participants\nwith low multimorbidity burden (LOW, N=243855); 2) a group with mental health conditions\n(MH, N=23116); 3) a group with cardiometabolic conditions (CVMTB, N=10971); 4) a group\nwith eye conditions (EYE, N=3302); and 5) a group with neurovascular conditions (NV ASC,\nN=1468). The multimorbidity burden was lowest in the LOW group (mean number of chronic\nconditions = 0.27 +- 0.57), and highest in the NV ASC group (2.92 +- 2.05). The MH group had a\nhigher percentage of females (67.74%) than males, while the CVMTB (35.92%) and NV ASC\n(41.49%) groups had a lower percentage of females (Table 2).\n9\n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted January 23, 2024. ; https://doi.org/10.1101/2024.01.21.24301584doi: medRxiv preprint \n\nTable 2: Demographic characteristics and disease burden of multimorbidity clusters. Age at\nbaseline, number and proportion of females, years of education, and mean (sd) number of chronic\nconditions are given for each multimorbidity cluster. Note that the number of chronic conditions refers\nonly to the average number of newly acquired diagnoses received in that age range.\nMultimorbidity patterns at 0-55 years\nCVMTB\n(N = 10,971,\n3.88%)\nEYE\n(N = 3,302,\n1.17%)\nLOW\n(N = 243,855,\n86.26%)\nMH\n(N = 23,116,\n8.18%)\nNV ASC\n(N = 1,468,\n0.52%)\nAge at baseline,\nyears (SD) 58.66 (4.16) 59.10 (4.30) 61.99 (4.57) 59.37 (4.56) 58.68 (4.20)\nFemale (%) 3,941 (35.92%) 1,723 (52.18%) 128,756 (52.80%) 15,658 (67.74%) 609 (41.49%)\nEducation, years\n(SD) 12.83 (3.07) 13.38 (3.09) 13.12 (3.08) 13.19 (3.05) 12.76 (2.98)\nNumber of\nchronic\nconditions in this\nage range (SD) 2.83 (1.71) 2.12 (1.52) 0.27 (0.57) 2.31 (1.35) 2.92 (2.05)\nMultimorbidity patterns at 55-66 years\nCVMTB\n(N = 43,501,\n15.39%)\nLOW\n(N = 214,888,\n76.01%)\nMH\n(N = 16,136,\n5.71%)\nNV ASC\n(N = 4,689,\n1.66%)\nPV ASC\n(N = 3,498,\n1.24%)\nAge at baseline,\nyears (SD) 60.71 (4.50) 61.93 (4.63) 59.86 (4.61) 61.06 (4.65) 61.28 (4.65)\nFemale (%) 19,379 (44.55%) 117,344 (54.61%) 10,646 (65.98%) 1,910 (40.73%) 1,408 (40.25%)\nEducation, years\n(SD) 12.74 (3.06) 13.22 (3.08) 12.90 (3.01) 12.82 (3.12) 12.37 (3.02)\nNumber of new\nchronic\nconditions in this\nage range (SD) 3.76 (1.56) 0.72 (0.85) 3.32 (1.97) 3.79 (2.27) 4.28 (2.57)\nMultimorbidity patterns at 65-70 years\nCVMTB\n(N = 39,960,\n18.47%)\nLOW\n(N = 162,318,\n75.02%)\nMH\n(N = 7,795,\n3.6%)\nNV ASC\n(N = 3,994,\n1.85%)\nPV ASC\n(N = 2,305,\n1.07%)\nAge at baseline,\nyears (SD) 63.41 (3.33) 63.46 (3.45) 62.55 (3.16) 63.50 (3.34) 63.47 (3.40)\nFemale (%) 18,896 (47.29%) 87,709 (54.04%) 5,201 (66.72%) 1,729 (43.29%) 927 (40.22%)\nEducation, years\n(SD) 12.68 (3.05) 13.06 (3.07) 12.67 (3.03) 12.79 (3.12) 12.34 (2.96)\nNumber of new\nchronic 3.14 (1.26) 0.47 (0.65) 3.20 (1.85) 3.16 (1.83) 3.65 (2.11)\n10\n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted January 23, 2024. ; https://doi.org/10.1101/2024.01.21.24301584doi: medRxiv preprint \n\nconditions in this\nage range (SD)\nThe observed/expected (OE) ratio (OE) is used to describe the notable conditions within each\ncluster. For example, an OE ratio of 2 for diabetes in a given cluster indicates that there are twice\nas many participants in that cluster with diabetes than expected, based on the prevalence of\ndiabetes in the entire sample (Methods).\nComorbidity clusters are represented by chord diagrams, circular figures with nodes along the\ncircumference and lines (edges) connecting each pair of nodes. The size of nodes is scaled to the\nOE of a given condition, while the edges are colour coded to show high and low comorbidity\n(Figures 2-4), such that orange-yellow links between a pair of conditions represents a larger than\nexpected comorbidity (Methods).\nFigure 2 illustrates each cluster within the 0-55 age range. As expected, no condition is\noverrepresented in the LOW group, suggesting low/no multimorbidity in this group. The\nCVMTB group is so named because of an over-representation of cardiovascular and metabolic\nconditions in this cluster. This group also includes a higher than expected number of individuals\nwith chronic kidney disease, liver disease, and respiratory conditions, although these were less\nprevalent than the cardio-metabolic conditions. The EYE group is dominated by the\nco-occurence of eye conditions such as macular degeneration, glaucoma, and cataract, but also\ncontains a higher than expected number of individuals with chronic kidney disease and diabetes.\n11\n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted January 23, 2024. ; https://doi.org/10.1101/2024.01.21.24301584doi: medRxiv preprint \n\nThe NV ASC group is dominated by stroke and TIA, but also includes a larger than expected\nproportion of individuals with other neurological conditions, cardiovascular conditions, stress\ndisorders, and metabolic disorders. The MH group is defined by the over-representation of\nmental health conditions (psychoses, anxiety disorders, depression, and stress disorders). There\nwas also a larger than expected proportion of chronic fatigue syndrome, psoriasis,\ngastrointestinal conditions (irritable bowel syndrome, inflammatory bowel disease), neurological\nconditions (epilepsy, multiple sclerosis, Parkinson’s disease), and respiratory conditions in the\nMH group.\n12\n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted January 23, 2024. ; https://doi.org/10.1101/2024.01.21.24301584doi: medRxiv preprint \n\nFigure 2: Multimorbidity patterns from birth to age 55 are defined by mental health,\ncardiometabolic, neurovascular, eye, and low disease burden patterns. For each multimorbidity\ncluster identified from age 0-55 years, we plot a chord diagram describing the defining conditions and\ntheir comorbidity patterns. Each condition is represented by a dot along the circumference of the cluster\ndiagram. The dots are colour coded according to ICD-10 classification (eg. starting with anaemia and\nmoving clockwise, cardiovascular conditions are in red, neurological conditions are in light blue,\nmetabolic conditions are in yellow, genitourinary conditions are in teal, thyroid conditions are in light\ngreen, gastrointestinal conditions are in green, sensory (hearing) conditions are in blue, sensory (eye)\nconditions are in black, mental health conditions are in purple, chronic fatigue syndrome, psoriasis, and\ncancer are in pink (grouped together as ‘other’ conditions), musculoskeletal conditions are in grey, and\nrespiratory conditions are in brown). The size of the dot is scaled to match the observed/expected (OE)\nratio, such that a larger OE represents a larger than expected proportion of participants in the cluster\nhaving a diagnosis of a given chronic condition. Lines connecting pairs of conditions are colour coded to\nrepresent the strength of comorbidity between two given conditions, such that orange-yellow lines\nrepresent strong comorbidity, while black lines indicate low comorbidity. The clusters are named\naccording to the OE and comorbidity patterns. Moving clockwise from top left, the multimorbidity\nclusters from birth to age 55 are a mental health cluster (MH), a low disease burden/health cluster (LOW),\na neurovascular cluster (NV ASC), a cardiometabolic cluster (CVMTB), and an eye cluster (EYE).\n13\n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted January 23, 2024. ; https://doi.org/10.1101/2024.01.21.24301584doi: medRxiv preprint \n\nMultimorbidity from 55-65 years\nThe five multimorbidity clusters are described as: 1) a relatively healthy group of participants\n(LOW, N=214888); 2) a group with mental health conditions (MH, N=16136); 3) a group with\ncardiometabolic conditions (CVMTB, N=43501); 4) a group with peripheral vascular conditions\n(PV ASC, N=3498); and 5) a group with neurovascular conditions (NV ASC, N=4689). The\nmultimorbidity burden was lowest in the LOW group (mean number of chronic conditions = 0.72\n+- 0.85), and highest in the PV ASC group (4.28 +- 2.57).\nThe MH group had a higher percentage of females (65.98%), while the CVMTB (44.55%),\nNV ASC (40.73%), and PV ASC (40.25%) groups had a lower percentage of females than males.\n(Table 2). Compared to the multimorbidity patterns in the 0-55 age range, the 55-65 range had\nfewer participants in the LOW and MH groups, and more participants in the CVMTB group.\nWhile disease burden in the LOW group remained similar as in the younger age range, in all\nother groups participants developed, on average, over three new chronic conditions between the\nages of 55 and 65 (Table 2).\nFigure 3 plots the OE patterns of each cluster. As expected, no condition is over-represented in\nthe LOW group. The CVMTB group has an over-representation of cardiovascular and metabolic\nconditions and also a higher than expected number of individuals with chronic kidney disease,\nliver disease, and chronic obstructive pulmonary disorder. The NV ASC group is again dominated\nby stroke and TIA, but also includes a larger than expected proportion of individuals with other\nneurological conditions (epilepsy, migraine), cardiovascular conditions (heart failure, atrial\nfibrillation), chronic fatigue syndrome, lipid disorders, chronic kidney disease, and stress\n14\n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted January 23, 2024. ; https://doi.org/10.1101/2024.01.21.24301584doi: medRxiv preprint \n\ndisorders. The MH group has an over-representation of mental health conditions (psychoses,\nanxiety disorders, depression, and stress disorders), but there was also a larger than expected\nproportion of chronic fatigue syndrome, psoriasis, and neurological conditions (epilepsy,\nmigraine, Parkinson’s disease). In the PV ASC group, which did not appear in the 0-55 age range,\nthe most prominent conditions are atherosclerosis and other peripheral vascular disorders. Other\nconditions over represented include cardiovascular conditions (heart failure, CHD, atrial\nfibrillation), metabolic conditions, chronic kidney diseases, depression, and chronic obstructive\npulmonary disorder. The EYE group from 0-55 years did not appear in the 55-65 year age range.\n15\n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted January 23, 2024. ; https://doi.org/10.1101/2024.01.21.24301584doi: medRxiv preprint \n\nFigure 3: Multimorbidity patterns from age 55-65 are defined by mental health, cardiometabolic,\nneurovascular, peripheral vascular, and low disease burden patterns. Figure 2 plots, for each\nmultimorbidity cluster identified from age 55-65 years, a chord diagram describing the defining\nconditions and their comorbidity patterns. Moving clockwise from top left, the multimorbidity clusters\nfrom age 55-65 are a mental health cluster (MH), a low disease burden/health cluster (LOW), a\nneurovascular cluster (NV ASC), a cardiometabolic cluster (CVMTB), and a peripheral vascular (PV ASC).\nMultimorbidity from 65 - 70 years\nWe identified five multimorbidity clusters, similar to the 55-65 age range : 1) LOW N=162318;\n2) MH, N=7795; 3) CVMTB, N=39960), defined by an over-representation of cardiovascular\n(except atherosclerosis) and metabolic conditions and a higher than expected number of\nindividuals with chronic kidney disease, endometriosis, gastrointestinal conditions, respiratory\nconditions, and sleep disorders. ; 4) PV ASC (N=2305),; and 5) NV ASC (N=3994), dominated by\nstroke and TIA, but also including a larger than expected proportion of individuals with other\n16\n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted January 23, 2024. ; https://doi.org/10.1101/2024.01.21.24301584doi: medRxiv preprint \n\nneurological conditions (epilepsy, migraine), cardiovascular conditions, chronic fatigue\nsyndrome, lipid disorders, chronic kidney disease, and chronic obstructive pulmonary disorder..\nAs in the 55-65 range, the multimorbidity burden was lowest in the LOW group had the lowest\nmultimorbidity burden (mean number of chronic conditions = 0.47 +- 0.65), and highest in the\nPV ASC group (3.65 +- 2.11)\nThe MH group had more females (66.72%), while the CVMTB (47.29%), NV ASC (43.29%),\nand PV ASC (40.22%) groups had fewer females than males. (Table 2). There were fewer\nparticipants in the LOW and MH groups at 65-70 than at 55-65, despite the shorter age range.\nNumbers in the CVMTB, PV ASC, and NV ASC groups remained relatively stable across the\nages. While disease burden in the LOW group remained similar, in all other groups participants\ndeveloped, on average, over three new chronic conditions between the ages of 65 and 70 (Table\n2). 7,465 individuals who had a death record between ages 65-70, 1102 who received a diagnosis\nof dementia between ages 65-70, and 57,773 individuals who did not have health records past\nage 70 were not included in the multimorbidity clustering procedure in the 65-70 age range.\n216,372 were included in this analysis (Methods).\nFigure 4 plots the OE patterns of each cluster, which largely shows similar characteristics as\nclusters in the 55-65 age range. Thus, the same naming conventions are used for clusters of\nconditions developed between age 65-70.\n17\n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted January 23, 2024. ; https://doi.org/10.1101/2024.01.21.24301584doi: medRxiv preprint \n\nFigure 4: Multimorbidity patterns from age 65-70 are defined by mental health, cardiometabolic,\nneurovascular, peripheral vascular, and low disease burden patterns. Figure 2 plots, for each\nmultimorbidity cluster identified from age 65-70 years, a chord diagram describing the defining\nconditions and their comorbidity patterns. Moving clockwise from top left, the multimorbidity clusters\nfrom age 65-70 are a mental health cluster (MH), a low disease burden/health cluster (LOW), a\nneurovascular cluster (NV ASC), a cardiometabolic cluster (CVMTB), and a peripheral vascular (PV ASC).\nMultimorbidity Clusters associated with dementia\nTo assess the association of multimorbidity clusters with incident dementia, we used a logistic\nregression with incident dementia as outcome, multimorbidity cluster as predictor, and age, sex,\nand years of education as covariates. The LOW group was used as the reference group. To first\nassess the cross-sectional association of each multimorbidity cluster, irrespective of prior or\nfuture multimorbidity, we analysed each age range independently and found all multimorbidity\nclusters were significantly associated with increased risk of incident dementia ( < 0.05) with𝑃𝐹𝐷𝑅\n18\n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted January 23, 2024. ; https://doi.org/10.1101/2024.01.21.24301584doi: medRxiv preprint \n\nthe exception of the EYE cluster in the 0-55 age range (Methods). To assess how the\naccumulation of multimorbidity over the lifespan is associated with incident dementia, we\ninstead performed a logistic regression with incident dementia as outcome, multimorbidity\ncluster from 0-55, multimorbidity cluster from 55-65, and multimorbidity cluster from 65-70 as\npredictors, and age, sex, and years of education as covariates. This identifies the risk associated\nwith multimorbidity patterns at each time point while covarying for a participant’s past and\nfuture multimorbidity. This analysis identified both disease- and time-specific patterns of\nmultimorbidity associated with increased risk.\nIn the 0-55 age range, the NV ASC (OR = 1.58 95% CI = [1.11, 2.19]) and CVMTB (OR = 1.51,\nCI = [1.31, 1.72]) clusters were associated with increased dementia risk in comparison to the\nLOW cluster. At this age, belonging to the MH or EYE cluster was not associated with increased\nrisk of dementia when compared to having low/no multimorbidity. However, this pattern was\ndifferent in the 55-65 age range, where relative to the LOW cluster, the NV ASC (OR = 1.84,\n[1.58, 2.14]), MH (OR = 1.72 [1.55, 1.9]), CVMTB (OR = 1.5 [1.4, 1.6]), and PV ASC (OR=1.4,\n[1.16, 1.69]) clusters were associated with increased dementia risk. In the 65-70 age range, the\nMH (OR = 2.48 [2.21, 2.78]) and NV ASC (OR = 2.46, [2.13, 2.83]) clusters stood out as having\nover a two-fold increase in the odds of developing dementia in comparison to the LOW cluster.\nThe CVMTB (OR = 1.22 [1.14, 1.31]) cluster also showed increased risk in this time frame.\nTable 3 displays the OR of each multimorbidity cluster across the age ranges investigated.\n19\n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted January 23, 2024. ; https://doi.org/10.1101/2024.01.21.24301584doi: medRxiv preprint \n\nTable 3: Multimorbidity impact on risk of incident dementia varies across the life course. Table 3\ndisplays odds ratios (OR) of each multimorbidity group at each age range, representing the risk associated\nwith each group while co-varying for a participant’s age, sex, years of education, and life course patterns\nof multimorbidity. 95% confidence intervals and statistical significance of the computed ORs is also\nshown.\nMultimorbidity patterns at 0-55 years\nPredictor beta Odds Ratio [95% CI] 𝑝 𝑝𝐵𝐻\nIntercept -17.09 0 0\nCVMTB 0.61 1.84 [1.61, 2.1] 1.43E-19 2.86E-19\nEYE 0.06 1.06 [0.78, 1.4] 0.69 0.69\nMH 0.12 1.13 [1.02, 1.26] 0.02 0.03\nNV ASC 0.67 1.96 [1.37, 2.71] 9.33E-05 1.24E-04\nAge 0.22 1.24 [1.24, 1.25] 0 0\nMale Sex 0.16 1.18 [1.12, 1.24] 7.49E-11 1.20E-10\nEducation (years) -0.05 0.95 [0.95, 0.96] 1.88E-27 5.02E-27\nMultimorbidity patterns at 55-65 years\nPredictor beta Odds Ratio [95% CI] 𝑝 𝑝𝐵𝐻\nIntercept -17.51 0 0\nCVMTB 0.49 1.64 [1.53, 1.75] 2.06E-49 5.50E-49\nEYE 0.63 1.88 [1.7, 2.08] 2.83E-34 5.66E-34\nNV ASC 0.77 2.16 [1.85, 2.5] 3.10E-23 4.95E-23\nPV ASC 0.55 1.73 [1.43, 2.08] 8.97E-09 8.97E-09\nAge 0.22 1.25 [1.24, 1.26] 0 0\nMale Sex 0.15 1.16 [1.1, 1.22] 4.33E-09 4.95E-09\nEducation (years) -0.04 0.96 [0.95, 0.97] 5.75E-22 7.67E-22\nMultimorbidity patterns at 65-70 years\nPredictor beta Odds Ratio [95% CI] 𝑝 𝑝𝐵𝐻\nIntercept -19.84 0 0\n20\n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted January 23, 2024. ; https://doi.org/10.1101/2024.01.21.24301584doi: medRxiv preprint \n\nCVMTB 0.21 1.24 [1.16, 1.33] 5.21E-10 6.94E-10\nEYE 0.93 2.54 [2.26, 2.84] 2.66E-56 7.10E-56\nNV ASC 0.96 2.61 [2.27, 2.99] 8.87E-42 1.77E-41\nPV ASC 0.33 1.39 [1.09, 1.73] 0.01 0.01\nAge 0.26 1.29 [1.28, 1.3] 0 0\nMale Sex 0.14 1.15 [1.09, 1.22] 2.75E-07 3.14E-07\nEducation (years) -0.04 0.96 [0.95, 0.97] 1.19E-20 1.91E-20\nMultimorbidity Trajectories associated with dementia\nHaving identified the differential association of multimorbidity phenotypes with dementia at\neach distinct age-range, we sought to characterise the trajectory of accumulation of\nmultimorbidity over the life-course in dementia patients. Here we use a sankey diagram to\nvisualise the movement of dementia patients between multimorbidity categories prior to\nreceiving a dementia diagnosis (Figure 5).\n21\n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted January 23, 2024. ; https://doi.org/10.1101/2024.01.21.24301584doi: medRxiv preprint \n\nFigure 5: Multimorbidity trajectories of dementia patients. Our clustering analysis separates\nparticipants into groups in each of three age ranges assessed. Here, a sankey diagram is used to visualise\nthe movement of dementia patients between multimorbidity clusters from 55 to 65, and from 65 to 70.\nParameters of the figure (block height, link width, link colour) are coordinated to highlight the\nmultimorbidity trajectories with larger proportions of participants who received a future diagnosis of\ndementia. The height of each block is scaled to the log of the total number of future dementia cases in a\ngiven cluster (taller blocks represent larger numbers of future patients). Similarly, the height of each link\nis scaled to the log of the total number of future dementia cases who follow a given trajectory. To\nhighlight the trajectories which occur more frequently in future dementia patients, the colour of the links\nis scaled to the proportion of participants in the full sample following a trajectory who receive a future\ndiagnosis of dementia. For example, the wide link between the LOW clusters at 55 and 65 indicate a\nrelatively larger number of participants in the full sample following this trajectory who receive a future\ndiagnosis. However, the dark purple colour of this link conveys that a relatively lower percentage (2.38%)\nof individuals in the full sample following this trajectory receive a future diagnosis of dementia.\nConversely, the bright link between CVMTB at 65 and NV ASC at 70 indicates a higher percentage\n(8.13%) of participants in the full sample following this trajectory receive a future diagnosis of dementia,\nthough the overall number of individuals following this path is not as large as the LOW at 55 to LOW at\n65 path.\nVisualisation of the multimorbidity trajectories of dementia patients allows us to assess\ndifferential risk based on a participant’s unique pattern of accumulation of chronic conditions\n22\n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted January 23, 2024. ; https://doi.org/10.1101/2024.01.21.24301584doi: medRxiv preprint \n\nover their lifespan. Several notable observations can be made. Trajectories on the left side of\nFigure 5 generally represent lower dementia prevalence than those on the right side, indicating\nthat delaying or reducing multimorbidity burden until the age of 65 could have key implications\nfor risk reduction. There was a higher prevalence of incident dementia in all pathways leading to\nNV ASC membership in the 65-70 age range, highlighting the significant risk associated with\ndiagnosis of conditions such as stroke, TIA, and epilepsy between the ages of 65 and 70. In\nspecific cases, the risk associated with membership in a multimorbidity cluster at a given time\nvaried based on which diseases a person accumulated over the next five or ten years. For\nexample, while the CVMTB group from 0-55 years is associated with increased dementia risk,\nthis risk is most prominent for the individuals in the CVMTB group who progress to the NV ASC\ngroup over the next ten years of their lives. Conversely, those in the CVMTB group at 55 who\nhad minimal disease burden (LOW) in the next ten years had a lower risk of dementia.\nInterestingly, the CVMTB-NV ASC pathway contained a higher proportion of incident dementia\ncases than the CVMTB-CVMTB pathway. Progression to the NV ASC cluster at 70, as opposed\nto progression to any other cluster, was also associated with higher risk for participants who were\nin the NV ASC cluster at 55 and 65, the MH cluster at 65, and the CVMTB cluster at 65. Other\ntrajectories notably containing high prevalence of incident dementia include PV ASC at 65 to MH\nat 70 and NV ASC at 65 to either CVMTB or PV ASC at 70. Throughout the lifespan, progression\nto membership in the LOW cluster was associated with a significantly lower risk in comparison\nto any other trajectory.\n23\n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted January 23, 2024. ; https://doi.org/10.1101/2024.01.21.24301584doi: medRxiv preprint \n\nDiscussion\nIdentification of at-risk individuals is essential for any successful modification of the disease\ncourse of dementia. We clustered participants based on accumulation of chronic conditions\nwithin three distinct age ranges, and present three novel and salient suggestions which help bring\nunderstanding of multimorbidity as a risk factor in line with the current life-course perspective of\nother dementia risk factors 16. First, we note that the link between multimorbidity and increased\nrisk of incident dementia can be stratified by both the type and timing of their accumulated\nchronic conditions.\nSecond, we find that participants who repeatedly accumulated chronic conditions throughout\ntheir lifespan had significantly higher odds of dementia, whereas in those who remained in the\n“low” cluster until age 65, and who moved from a multimorbidity cluster to the “low” cluster at\nthe next time point, odds of developing dementia was considerably lower. This suggests that\ndelaying the development of a new chronic condition before the age of 65 could have significant\nimplications for dementia risk reduction for all individuals, regardless of the diseases they\naccumulated prior to age 55.\nThird, we found that a diagnosis of either cardiometabolic or neurovascular conditions by the age\nof 55 was associated with increased risk of future dementia relative to low/no multimorbidity.\nDeveloping a new cardiometabolic, neurovascular, mental health, or peripheral vascular\ncondition between the ages of 55 and 65 was also associated with higher odds of dementia. In\nthis critical age range, compared to the multimorbidity-dementia link at age 55, the risk\nassociated with cardiometabolic conditions remained relatively stable while the risk associated\n24\n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted January 23, 2024. ; https://doi.org/10.1101/2024.01.21.24301584doi: medRxiv preprint \n\nwith new mental health and neurovascular conditions increased. This trend continued in the\n65-70 age range, in which the accumulation of either mental health or neurovascular conditions\nwas associated with an over two-fold increase in the odds of future dementia. Development of\ncardiometabolic conditions remained associated with increased risk, although to a lesser degree\nthan previous age ranges.\nFurther, we present the unique trajectories of multimorbidity that were most strongly associated\nwith dementia. From age 55 until 65, the NV ASC-NV ASC, CVMTB-NV ASC, and\nLOW-NV ASC trajectories were most strongly associated with future dementia, with 6.3%, 5.9%,\nand 4.1% of all participants following these trajectories going on to develop dementia,\nrespectively. This highlights the importance of delaying onset of neurovascular and\ncardiometabolic conditions until age 65 for reducing dementia risk. Between age 65 and 70, the\nPV ASC-MH, MH-NV ASC, and CVMTB-NV ASC trajectories were most strongly associated\nwith future dementia, with 9.2%, 9.1%, and 8.1% of all participants following these trajectories\ngoing on to develop dementia, respectively. This highlights the emergence of mental health\nconditions between ages 65 and 70 as strong risk factors for dementia, as well as the continued\nrisk associated with neurovascular conditions.\nTo date, few studies have assessed either trajectories 18 or clusters 19,20 of multimorbidity in the\ncontext of dementia risk, and no study has assessed how the longitudinal accumulation of\nchronic conditions influences dementia risk. A large study using Danish registry data computed\ntemporal trajectories of individual chronic conditions prior to a diagnosis of Alzheimer’s disease\nand compared them to trajectories preceding vascular dementia 18. Diabetes, hypertension, heart\n25\n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted January 23, 2024. ; https://doi.org/10.1101/2024.01.21.24301584doi: medRxiv preprint \n\nfailure, depression, intracerebral haemorrhage, and atherosclerosis were among the conditions\nidentified as commonly occurring prior to an AD or V aD diagnosis. While this study looked at\naccumulation of multimorbidity, it did not consider age of onset or multimorbidity clusters. Our\nfindings add vital new information on how risk varies based on when and in what combination\nthe above conditions are diagnosed.\nTwo recent studies have studied multimorbidity clusters in the context of dementia risk. In one, a\nmental health cluster and a cardiometabolic cluster were most strongly associated with dementia\nrisk. An inflammatory/autoimmune cluster centred on arthritis and psoriasis was also moderately\nlinked to higher risk 20. In the second study, a sex stratified analysis was performed. In women, a\nhypertension, diabetes, and CHD cluster, along with an osteoporosis and dyspepsia cluster were\nassociated with highest risk. In men, a diabetes and hypertension cluster, along with a CHD,\nhypertension, and stroke cluster were associated with highest risk 19. Both these studies analysed\nthe UK Biobank population, though neither considered the longitudinal progression of\nmultimorbidity. These findings and ours add to overwhelmingly strong evidence linking\ncardiometabolic (eg. CHD, hypertension, diabetes, lipid disorders), neurovascular (e.g. stroke),\nand mental health (e.g. schizophrenia, depression, anxiety) conditions to increased risk of\ndementia.\nThese studies both describe clusters with an inflammatory/immune component (e.g. conditions\nsuch as arthritis, psoriasis, and osteoporosis), as being linked to higher risk of dementia. While\nnone of the clusters in our study were dominated by the prevalence or comorbidity of these\nconditions, arthritis, psoriasis, and osteoporosis did occur in greater than expected proportions\n26\n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted January 23, 2024. ; https://doi.org/10.1101/2024.01.21.24301584doi: medRxiv preprint \n\n(i.e. OE > 2) in certain clusters, particularly in the 0-55 age range. Thus, we also find some\nevidence of inflammatory/immune conditions contributing to dementia. However, given that\nneither of the above studies considered the sequential progression of multimorbidity, it is\npossible that the higher risk attributable to these conditions would be masked, or reduced, by the\ndevelopment of cardiometabolic, neurovascular, or mental health conditions in the future and not\nconsidered in their cross-sectional analyses.\nStrengths\nThe primary strength of our study is the integration of a clustering and longitudinal approach to\nadvance understanding of multimorbidity as a risk factor in line with the current life-course\nperspective of other dementia risk factors. To derive multimorbidity clusters we used MCA, a\ndimensionality reduction approach which strives to identify pairs of conditions which co-occur in\ngreater proportions than is expected based on their prevalence in the sample. This gives our study\nthe ability to identify strong comorbidity patterns between conditions even if they have a lower\noverall prevalence. This is crucial for identifying mental health clusters which have proven to be\na common finding in studies of multimorbidity patterns 2,6,21,22, but may be hard to detect if the\nprevalence of mental health conditions is lower in the sample 19. We also employed a\nlongitudinal approach by delineating electronic health record derived diagnoses into specific age\nranges, allowing us to identify critical periods in which multimorbidity conveys higher or lower\nrisk, along with the differential risk associated with certain clusters. The use of age ranges, as\nopposed to utilising exact ages, both accounts for uncertainty associated with diagnosis dates in\n27\n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted January 23, 2024. ; https://doi.org/10.1101/2024.01.21.24301584doi: medRxiv preprint \n\nelectronic health records, and eliminates methodological issues arising when multiple diagnoses\nare linked to the same date.\nLimitations\nFour key limitations require consideration in this study. First, the UK Biobank lacks a gold\nstandard, clinically adjudicated diagnosis of dementia. However our approach, combining\nhospital records, death records, and primary care data, has demonstrated a positive predictive\nvalue of 82.5% compared to clinical adjudication 23. Combined with the size of the dataset, this\nmakes UK Biobank a unique resource to study dementia outcomes. The positive predictive value\nof dementia subtypes was reported to be lower, and thus we only consider all-cause dementia as\nopposed to Alzheimer’s Disease, vascular dementia, or other subtypes. Second, the UK Biobank\ncohort is predominantly Caucasian, and has a lower incidence of chronic conditions than the\ngeneral population 24. As dementia risk and onset varies by race and ethnicity 25, further evaluation\nin more diverse cohorts is required before large generalisation of our findings. Third, the\nobservational nature of this cohort and long prodrome of dementia precludes us from making any\ninferences regarding causation. We also note that the presence of multimorbidity may increase\nthe likelihood of an individual seeking medical attention and further compounding the number of\ndiagnoses received. Fourth, we did not consider treatments being administered to participants\n(eg. medication). While we focus on age of diagnosis, we did not have access to other\ncontinually updated measures of disease severity. Thus, our findings remain associative in\nnature.\n28\n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted January 23, 2024. ; https://doi.org/10.1101/2024.01.21.24301584doi: medRxiv preprint \n\nConclusion\nWe studied the accumulation of multimorbidity throughout the life course in a large prospective\ncohort and found that the strong association between multimorbidity and a future diagnosis of\ndementia varies based on the specific type of conditions an individual has, as well as when they\nreceived a diagnosis. Until midlife (age 55), the accumulation of cardiometabolic conditions,\nsuch as coronary heart disease, atrial fibrillation, heart failure, diabetes, and lipid disorders, was\nstrongly associated with dementia risk. However from ages 55 to 70, the accumulation of mental\nhealth conditions, such as anxiety, depression, psychoses, and stress, as well as neurovascular\nconditions, such as stroke, transient ischaemic attack, and epilepsy, was most strongly linked to\ndementia. Crucially, individuals who continuously and sequentially accumulate cardiometabolic,\nmental health, and neurovascular conditions were at greatest risk. These findings promote the\nimportance of considering multimorbidity patterns across the life course when assessing an\nindividual’s risk for dementia.\n29\n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted January 23, 2024. ; https://doi.org/10.1101/2024.01.21.24301584doi: medRxiv preprint \n\nMethods\nParticipants\nWe used data from the UK Biobank, a large population study of 502,386 individuals 17.\nAssessments were conducted at one of 22 centres throughout Scotland, England, and Wales. At\nthe baseline assessment conducted between 2006-2010, participants were aged between 40-69\nyears old. Health outcomes are available via linked medical records, enabling the ascertainment\nof dementia and other chronic conditions throughout the lifespan, prior to and after the baseline\nassessment. To reduce the likelihood of including monogenetic cases of early-onset dementia 13,\nwe excluded any individuals for whom we could not confirm dementia free status at the age of\n65. Thus, we excluded any individuals who did not have available medical records after the age\n65 (N=158,793) and any individuals who received a diagnosis of dementia (N=859) or had a\ndeath record prior to age 65 (N=4460). We also excluded any individuals who had zero\ndiagnoses of chronic conditions (55,562) in order to focus our analysis on patterns of\nmultimorbidity. This produced an analysis sample of 282,712 individuals. The UK Biobank\nstudy received ethical approval from the Northwest Multi-Centre Research Ethics Committee.\nAll participants provided their informed written consent.\nAscertainment of chronic conditions\nAscertainment of chronic condition diagnoses was completed using three complementary\nsources: hospital inpatient data, primary care data, and death records. Hospital inpatient data was\navailable for the full cohort, and describes the date of admission and diagnoses using\nInternational Classification of diseases, Tenth V ersion (ICD-10) coding. Primary care data was\n30\n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted January 23, 2024. ; https://doi.org/10.1101/2024.01.21.24301584doi: medRxiv preprint \n\navailable for 230,000 participants, and includes coded diagnoses and dates from GP systems via\nREAD2 and CTV-3 codes. Data on the date and cause of death was available via linkage to\nnational death registries, coded using ICD-10 codes. Hospital and death records were available\nup to October 1st, 2021. Primary care data was available up to 2017.\nWe developed a list of 46 chronic conditions to include based on previous recommendations 21\nand a survey of the multimorbidity literature. For each, we parsed the above sources to ascertain\nif and when an individual received a diagnosis. If an individual received multiple diagnoses of\nthe same condition (via repetition in the same data source or across multiple sources), the date of\ndiagnosis was taken as the earliest available record. Importantly, diagnoses of chronic conditions\nwere only considered if they occurred prior to a dementia diagnosis. The age at which a\ncondition was diagnosed was then computed by comparing diagnosis date to the participant’s\ndate of birth. We had access to year and month of birth, and so the midpoint of the birth month\nwas taken as the date of birth for all participants. The ICD-10, READ2, and CTV-3 codes used to\nidentify each condition were developed via merging of related works 26,27 as well UK Biobank\nguidelines (https://biobank.ndph.ox.ac.uk/showcase/refer.cgi?id=592 ). Codings are listed at\nhttps://git.fmrib.ox.ac.uk/psp365/mm_risk/-/blob/main/ltc_codes.csv?ref_type=heads.\nDementia Ascertainment\nAll-cause dementia status was ascertained based on a combination of hospital inpatient data,\nprimary care data, and death report data as done in several papers based on this cohort 28,29. To\nincrease diagnostic accuracy, we excluded 117 individuals who self-reported having dementia\n31\n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted January 23, 2024. ; https://doi.org/10.1101/2024.01.21.24301584doi: medRxiv preprint \n\nbut had no corresponding medical record. The list of ICD-10 codes are available at\nhttps://git.fmrib.ox.ac.uk/psp365/mm_risk/-/blob/main/ltc_codes.csv?ref_type=heads.\nMultimorbidity Clusters\nMultimorbidity clusters were derived using a two step procedure. First, a matrix with dimensions\n282712 x 46 (number of participants x number of chronic conditions) was derived in which each\nmatrix element was a binary (yes/no) categorization of disease diagnosis. This matrix was\nsubmitted to multiple correspondence analysis (MCA), a multivariate technique used for\nidentifying patterns in categorical data 30,31 and previously used to study multimorbidity patterns\n32. Commonly described as a version of principal component analysis for categorical data, MCA\nmaps input data to a multidimensional space by applying correspondence analysis to an indicator\nmatrix. MCA provides components and factor scores for both participants and conditions which\ndescribe their loading within each component. Crucially, associations between categorical\nvariables is weighted by the statistic 31. Thus, output MCA components are driven by pairs ofχ\n2\nchronic conditions for which the observed proportion of individuals with both conditions is\nsignificantly different from the expected number based on overall sample prevalence. After\nMCA analysis, participant factor scores were inputted to k-means clustering to identify clusters\nof individuals loading similarly onto MCA components. The resulting groups of individuals thus\nshare similar patterns of comorbid chronic conditions (i.e. multimorbidity patterns). The number\nof clusters was selected via the silhouette and bayesian information criterion as goodness of fit\nstatistics.\n32\n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted January 23, 2024. ; https://doi.org/10.1101/2024.01.21.24301584doi: medRxiv preprint \n\nIn defining multimorbidity clusters, we note key considerations. First, the above workflow was\nseparately applied in three distinct age ranges: 0-55 years, 55-65 years, and 65-70 years old. For\neach age-specific analysis, we consider diagnoses only occurring within the given time frame.\nFor example, if an individual was diagnosed with diabetes at age 60, the matrix element\ncorresponding to their diabetes diagnosis would be 0 (no) in both the 0-55 and 65-70 years, but 1\n(yes) in the 55-65 analysis. Second, all participants are included in each age-range analysis .\nThat is, an individual with linked health record data up to the age of 72 would be included in\neach of the three age-clustering workflows, though their input data would vary at each age,\ndepending on when they received a diagnosis of a particular condition. Third, while clustering in\nthe 65-70 age range, we excluded individuals who had a death record between ages 65-70,\nreceived a diagnosis of dementia between ages 65-70, or who did not have linked health records\npast the age of 70. Fourth, for each analysis described below, a group representing low\nmultimorbidity load (i.e. a relatively healthier group) was used as the reference group for\nstatistical comparisons.\nDescribing Clusters\nEach cluster was characterised and named according to the prevalence and patterns of chronic\nconditions within the cluster. Specifically, for each condition and within each cluster we\ncomputed the observed/expected ratio (OE) as the prevalence of a condition within a cluster\ndivided by the prevalence of a condition in the entire sample. For example, an OE of 2 for\ndiabetes in a particular cluster indicates that the prevalence of diabetes in that cluster is double\nthe prevalence of diabetes in the sample within the age range analysed. As MCA is weighted\ntowards identifying strong comorbidity, we quantified and considered comorbidity of each pair\n33\n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted January 23, 2024. ; https://doi.org/10.1101/2024.01.21.24301584doi: medRxiv preprint \n\nof conditions on a per cluster basis. Specifically, a contingency table is formed for a pair of\nconditions, from which the statistic is computed. This is repeated for each pair of conditionsχ\n2\nwithin each cluster. The magnitude of the resulting statistic indicates the strength ofχ\n2\ncomorbidity between a pair of conditions, based on the observed number of individuals\ndiagnosed with each condition. Comorbidity clusters are represented by chord diagrams, circular\nfigures with nodes along the circumference and lines (edges) connecting each pair of nodes. The\nsize of nodes is scaled to the OE of a given condition, while the edges are colour coded to show\nhigh and low comorbidity (Figures 2-4). statistics are capped at a maximum of 8 based on theχ\n2\nskewed (right tail) distribution of comorbidity metrics across all condition-condition pairs. Thus,\nthe colour scheme does not distinguish between pairs of conditions with comorbidity >8 and any\nsuch pairing will be viewed as having equally strong comorbidity.\nStatistical Analysis\nWe performed a logistic regression to test the association between multimorbidity clusters across\nthe lifespan and risk of incident dementia. The clustering workflow described above provides\nthree categorical variables (MM_55, MM_65, MM_70), each of which denotes the\nmultimorbidity cluster a participant is assigned to in each age range (e.g. a participant could\nbelong to a cardiometabolic cluster in the 0-55 age range, and then to a mental health cluster in\nthe 55-65 age range depending on the new conditions they acquired from age 55 to 65). We\nincluded incident dementia as the outcome variable and multimorbidity cluster as the predictor of\ninterest, and covaried for age at baseline, sex, and years of education. We first performed this\nanalysis for each age range separately, irrespective of participants' multimorbidity patterns at\nother age ranges (i.e. three separate models, each containing only one of MM_55, MM_65,\n34\n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted January 23, 2024. ; https://doi.org/10.1101/2024.01.21.24301584doi: medRxiv preprint \n\nMM_70). Then, to assess the impact of multimorbidity over the lifespan, we performed a logistic\nregression with incident dementia as outcome, all three of MM_55, MM_65, and MM_70 as\npredictors, and age, sex, and education as covariates. For this analysis, individuals excluded in\nthe 65-70 years clustering due to unavailability of health records were all assigned to a separate\ngroup (EXCLUDE) at age 70. Throughout each analysis, a group representing low\nmultimorbidity load (i.e. a relatively healthy group) was used as the reference group for\nstatistical comparisons. P values were corrected for multiple comparisons using false discovery\nrate correction ( < 0.05 considered significant).𝑃𝐹𝐷𝑅\nData and Code Availability\nThe data used in this study are available from the UK Biobank\n(https://www.ukbiobank.ac.uk/enable-your-research/apply-for-access). As restrictions apply to\nthe availability of these data, which were used under licence for the current study, the authors\ncannot publicly share this data. The use of data from the UK Biobank was approved by the UK\nBiobank Access Committee (Project No. 95758). Code for this project is available at\nhttps://git.fmrib.ox.ac.uk/psp365/mm_risk .\nFunding\nThe UK Biobank resource is supported by funding from the UK Medical Research Council and\nthe Wellcome Trust. The authors report the following funding: a Canadian Institutes of Health\nResearch Fellowship (R.P , Grant 458882 ); a UK Alzheimer’s Society Research Fellowship\n(S.S.; Grant 441), Academy of Medical Sciences/the Wellcome Trust/the Government\n35\n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted January 23, 2024. ; https://doi.org/10.1101/2024.01.21.24301584doi: medRxiv preprint \n\nDepartment of Business, Energy and Industrial Strategy/the British Heart Foundation/Diabetes\nUK Springboard Award (S.S; Grant SBF006\\1078), HDH Wills 1965 Charitable Trust (Nr:\n1117747), UK Medical Research Council (K.P .E.; G1001354, MR/K013351/1), the European\nCommission (K.P .E.; Horizon 2020, Grant agreement number: 732592) and Wellcome Trust.\nThis work was supported by the NIHR Oxford Health Biomedical Research Centre and the\nWellcome Centre for Integrative Neuroimaging (WIN). The WIN is supported by core funding\nfrom the Wellcome Trust (203139/Z/16/Z). The funders of this study had no role in the study\ndesign or collection, interpretation, analysis or reporting of the data.\nReferences\n1. Nicholson, K. et al. Examining the prevalence and patterns of multimorbidity in Canadian primary\nhealthcare: a methodologic protocol using a national electronic medical record database. J Comorb 5,\n150–161 (2015).\n2. Poblador-Plou, B. et al. Comorbidity of dementia: a cross-sectional study of primary care older\npatients. BMC Psychiatry 14, 84 (2014).\n3. Winblad, B. et al. Defeating Alzheimer’s disease and other dementias: a priority for European\nscience and society. Lancet Neurol. 15, 455–532 (2016).\n4. van Oostrom, S. H. et al. Multimorbidity of chronic diseases and health care utilization in general\npractice. BMC Fam. Pract. 15, 61 (2014).\n5. Palladino, R., Tayu Lee, J., Ashworth, M., Triassi, M. & Millett, C. Associations between\nmultimorbidity, healthcare utilisation and health status: evidence from 16 European countries. Age\nAgeing 45, 431–435 (2016).\n36\n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted January 23, 2024. ; https://doi.org/10.1101/2024.01.21.24301584doi: medRxiv preprint \n\n6. Calderón-Larrañaga, A. et al. Multimorbidity and functional impairment-bidirectional interplay,\nsynergistic effects and common pathways. J. Intern. Med. 285, 255–271 (2019).\n7. Gijsen, R. et al. Causes and consequences of comorbidity: A review. J. Clin. Epidemiol. 54, 661–674\n(2001).\n8. Barnett, K. et al. Epidemiology of multimorbidity and implications for health care, research, and\nmedical education: a cross-sectional study. Lancet 380, 37–43 (2012).\n9. Whitty, C. J. M. et al. Rising to the challenge of multimorbidity. BMJ 368, l6964 (2020).\n10. Pearson-Stuttard, J., Ezzati, M. & Gregg, E. W. Multimorbidity-a defining challenge for health\nsystems. Lancet Public Health 4, e599–e600 (2019).\n11. Skou, S. T. et al. Multimorbidity. Nat Rev Dis Primers 8, 48 (2022).\n12. Ben Hassen, C. et al. Association between age at onset of multimorbidity and incidence of dementia:\n30 year follow-up in Whitehall II prospective cohort study. BMJ 376, e068005 (2022).\n13. Tai, X. Y . et al. Cardiometabolic multimorbidity, genetic risk, and dementia: a prospective cohort\nstudy. Lancet Healthy Longev 3, e428–e436 (2022).\n14. Ho, I. S.-S. et al. Examining variation in the measurement of multimorbidity in research: a\nsystematic review of 566 studies. Lancet Public Health 6, e587–e597 (2021).\n15. Jindai, K., Nielson, C. M., V orderstrasse, B. A. & Quiñones, A. R. Multimorbidity and Functional\nLimitations Among Adults 65 or Older, NHANES 2005-2012. Prev. Chronic Dis. 13, E151 (2016).\n16. Livingston, G. et al. Dementia prevention, intervention, and care: 2020 report of the Lancet\nCommission. Lancet 396, 413–446 (2020).\n17. Sudlow, C. et al. UK biobank: an open access resource for identifying the causes of a wide range of\ncomplex diseases of middle and old age. PLoS Med. 12, e1001779 (2015).\n18. Jørgensen, I. F., Aguayo-Orozco, A., Lademann, M. & Brunak, S. Age-stratified longitudinal study\nof Alzheimer’s and vascular dementia patients. Alzheimers. Dement. 16, 908–917 (2020).\n19. Calvin, C. M., Conroy, M. C., Moore, S. F., Kuzma, E. & Littlejohns, T. J. Association of\nMultimorbidity, Disease Clusters, and Modification by Genetic Factors With Risk of Dementia.\n37\n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted January 23, 2024. ; https://doi.org/10.1101/2024.01.21.24301584doi: medRxiv preprint \n\nJAMA Netw Open 5, e2232124 (2022).\n20. Khondoker, M. et al. Multimorbidity pattern and risk of dementia in later life: an 11-year follow-up\nstudy using a large community cohort and linked electronic health records. J. Epidemiol. Community\nHealth 77, 285–292 (2023).\n21. Busija, L., Lim, K., Szoeke, C., Sanders, K. M. & McCabe, M. P . Do replicable profiles of\nmultimorbidity exist? Systematic review and synthesis. Eur . J. Epidemiol.34, 1025–1053 (2019).\n22. Schäfer, I. et al. Multimorbidity patterns in the elderly: a new approach of disease clustering\nidentifies complex interrelations between chronic conditions. PLoS One 5, e15941 (2010).\n23. Wilkinson, T. et al. Identifying dementia outcomes in UK Biobank: a validation study of primary\ncare, hospital admissions and mortality data. Eur . J. Epidemiol.34, 557–565 (2019).\n24. Fry, A. et al. Comparison of sociodemographic and health-related characteristics of UK Biobank\nparticipants with those of the general population. Am. J. Epidemiol. 186, 1026–1034 (2017).\n25. Stephan, B. C. M. et al. Prediction of dementia risk in low-income and middle-income countries (the\n10/66 Study): an independent external validation of existing models. The Lancet Global Health 8,\ne524–e535 (2020).\n26. Ronaldson, A. et al. Associations between physical multimorbidity patterns and common mental\nhealth disorders in middle-aged adults: A prospective analysis using data from the UK Biobank.\nLancet Reg Health Eur 8, 100149 (2021).\n27. Shang, X. et al. Association of a wide range of chronic diseases and apolipoprotein E4 genotype with\nsubsequent risk of dementia in community-dwelling adults: A retrospective cohort study.\nEClinicalMedicine 45, 101335 (2022).\n28. Gong, J., Harris, K., Peters, S. A. E. & Woodward, M. Sex differences in the association between\nmajor cardiovascular risk factors in midlife and dementia: a cohort study using data from the UK\nBiobank. BMC Med. 19, 110 (2021).\n29. Lourida, I. et al. Association of Lifestyle and Genetic Risk With Incidence of Dementia. JAMA 322,\n430–437 (2019).\n38\n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted January 23, 2024. ; https://doi.org/10.1101/2024.01.21.24301584doi: medRxiv preprint \n\n30. Sourial, N. et al. Correspondence analysis is a useful tool to uncover the relationships among\ncategorical variables. J. Clin. Epidemiol. 63, 638–646 (2010).\n31. Salkind, N. J. Encyclopedia of Measurement and Statistics. (Sage Publications, Inc., 2011).\ndoi:10.4135/9781412952644.\n32. V etrano, D. L. et al. Twelve-year clinical trajectories of multimorbidity in a population of older\nadults. Nat. Commun. 11, 3223 (2020).\n39\n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted January 23, 2024. ; https://doi.org/10.1101/2024.01.21.24301584doi: medRxiv preprint","source_license":"CC-BY-4.0","license_restricted":false}