www.aging-us.com AGING 2022, Vol. 14, No. 24 Research Paper 12-year evolution of multimorbidity patterns among older adults based on Hidden Markov Models Albert Roso-Llorach1,2,3,*, Davide L. Vetrano4,5,*, Caterina Trevisan4,6, Sergio Fernández1,2, Marina Guisado-Clavero7, Lucía A. Carrasco-Ribelles1,8, Laura Fratiglioni4,5, Concepción Violán2,9,#, Amaia Calderón-Larrañaga4,5,# 1 Fundació Institut Universitari per a la recerca a l’Atenció Primària de Salut Jordi Gol i Gurina (IDIAPJGol), Barcelona, Spain 2 Universitat Autònoma de Barcelona, Bellaterra (Cerdanyola de Vallès), Spain 3 Programa de Doctorat en Metodologia de la Recerca Biomèdica i Salut Pública, Universitat Autònoma de Barcelona, Bellaterra (Cerdanyola del Vallès), Spain 4 Aging Research Center, Department of Neurobiology, Care Sciences and Society, Karolinska Institutet and Stockholm University, Stockholm, Sweden 5 Stockholm Gerontology Research Center, Stockholm, Sweden 6 Department of Medical Sciences, University of Ferrara, Ferrara, Italy 7 Unidad Docente Multiprofesional de Atención Familiar y Comunitaria Norte, Gerencia Asistencial Atención Primaria, Madrid Health Service, Madrid, Spain 8 Signal Theory and Communications Department, Universitat Politecnica de Catalunya, Barcelona, Spain 9 Unitat de Suport a la Recerca Metropolitana Nord, Fundació Institut Universitaria per a la recerca a l’Atenció Primària de Salut Jordi Gol i Gurina (IDIAPJGol), Mataró, Barcelona, Spain *Shared first authorship # Shared last authorship Correspondence to: Albert Roso-Llorach, Amaia Calderón-Larrañaga; email: aroso@idiapjgol.org, amaia.calderon.larranaga@ki.se Keywords: multimorbidity, older adults, longitudinal population-based study, aging, Hidden Markov Models Received: July 13, 2022 Accepted: November 14, 2022 Published: November 23, 2022 Copyright: © 2022 Roso-Llorach et al. This is an open access article distributed under the terms of the Creative Commons Attribution License (CC BY 3.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. ABSTRACT Background: The evolution of multimorbidity patterns during aging is still an under-researched area. We lack evidence concerning the time spent by older adults within one same multimorbidity pattern, and their transitional probability across different patterns when further chronic diseases arise. The aim of this study is to fill this gap by exploring multimorbidity patterns across decades of age in older adults, and longitudinal dynamics among these patterns. Methods: Longitudinal study based on the Swedish National study on Aging and Care in Kungsholmen (SNAC-K) on adults ≥60 years (N=3,363). Hidden Markov Models were applied to model the temporal evolution of both multimorbidity patterns and individuals' transitions over a 12-year follow-up. Findings: Within the study population (mean age 76.1 years, 66.6% female), 87.2% had ≥2 chronic conditions at baseline. Four longitudinal multimorbidity patterns were identified for each decade. Individuals in all decades showed the shortest permanence time in an Unspecific pattern lacking any overrepresented diseases (range: 4.6-10.9 years), but the pattern with the longest permanence time varied by age. Sexagenarians remained longest in the Psychiatric-endocrine and sensorial pattern (15.4 years); septuagenarians in the Neuro-vascular www.aging-us.com 9805 AGING and skin-sensorial pattern (11.0 years); and octogenarians and beyond in the Neuro-sensorial pattern (8.9 years). Transition probabilities varied across decades, sexagenarians showing the highest levels of stability. Interpretation: Our findings highlight the dynamism and heterogeneity underlying multimorbidity by quantifying the varying permanence times and transition probabilities across patterns in different decades. With increasing age, older adults experience decreasing stability and progressively shorter permanence time within one same multimorbidity pattern. INTRODUCTION new preventive and, eventually, therapeutic approaches. Hidden Markov Models (HMM) over- Extended human longevity is a goal achieved in the last come several of the limitations of previously employed century, and a reality in middle- and high-income methods, which were unable to account for the countries [1]. Improvements in health resources and variability in chronic disease interactions throughout medical sciences, and decreases in preventable mortality time [9]. HMM consider diseases in each person to be have been key to living longer [2]. However, increasing random variables conditioned by a hidden state or life expectancy comes along with a higher burden of cluster. Despite the technique’s potential, only one chronic diseases [3]. The coexistence of multiple chronic previous register-based study has used HMM for the diseases in a single person is known as multimorbidity. longitudinal study of multimorbidity [9], but the Multimorbidity is associated with a higher risk of folllow-up time was insufficient to draw any relevant polypharmacy and decreased quality of life, and conclusions. Cohort studies with homogeneously challenges the decision-making of clinicians that lack collected data over long periods of time represent a effective guidelines for the management and treatment unique resource for the longitudinal analysis of of patients with cohexisting complex diseases [4]. multimorbidity patterns, and their use for such a purpose is warranted. In an attempt to understand how chronic diseases are inter-related, several studies have explored so-called The aims of this study were: 1) to explore longitudinal multimorbidity patterns [5–7]. In a previous systematic multimorbidity patterns across decades of age after review, three patterns of multimorbidity involving 60 using HMM, and 2) to detect the dynamics cardiometabolic diseases, mental health problems, and underlying such patterns in terms of the time subjects musculoskeletal disorders have been consistently remained within the same pattern, and the probability suggested to be the most prevalent in the older population of transitioning across different patterns. [5]. Diseases cluster in specific patterns due to common pathophysiological pathways and risk factors, or because RESULTS they may be the cause or consequence of other coexisting diseases. Along with the above mentioned patterns, a Multimorbidity patterns high number of less reproducible and sparse disease combinations have been described, often inconsistently The study population included 3,363 individuals aged across studies. Several factors may explain such disparate 60+ of whom 87.2% had multimorbidity at baseline. observations: first, the use of cross-sectional designs, Participants’ mean age at baseline was 76.1 years, and which do not account for the dynamic nature of 66.6% were female. Over the 12-year follow-up, 1346 multimorbidity in old age; second, the use of different (40%) deaths occurred (25% within the first 6 years and disease lists, spanning from less than ten to more than 15% within the next 6 years). Moreover, 719 (21.4%) two hundred conditions; and third, the employement of individuals dropped out (13.7% within the first 6 years statistical methods that cannot properly manage the and 7.7% within the next 6 years). Descriptive statistics complexity of the phenomenon. Recently, several of each age cohort at each follow-up wave can be found advanced machine-learning techniques such as non- in Table 1. hierarchical and hierarchical clustering tehcniques have been used to explore multimorbidity patterns. In the three age groups, a total of 44, 49 and 47 chronic disease categories, respectively, showed a median Exploring how multimorbidity patterns evolve prevalence ≥2% during the study period, and were thus throughout people’s lives and the time subjects remain included in the HMM estimations (Supplementary Table within specific patterns is still an under-researched 1). Overall, four multimorbidity patterns were identified area [7, 8]. The understanding of how diseases cluster for each age group, and two additional patterns were longitudinally in specific age groups would pave the artificially added to account for death and dropout way to the design of new prognostic tools, as well as during the follow-up period (Supplementary Table 2). www.aging-us.com 9806 AGING Table 1. Sociodemographic, clinical, and functional characteristics of the study population by baseline age group (N=3,363). Sexagenarians Septuagenarians Octogenarians and beyond 6 years 12 years 6 years 12 years 6 years 12 years Baseline Baseline Baseline follow-up follow-up follow-up follow-up follow-up follow-up N=1304 N=939 N=1120 N=1045 N=846 N=639 N=358 N=374 N=94 Age, mean (SD) 63.0 (2.91) 68.9 (2.89) 74.9 (2.88) 75.3 (3.00) 81.1 (2.98) 86.6 (2.89) 87.9 (5.10) 91.5 (4.11) 95.5 (2.84) Female, n (%) 735 (56.4%) 603 (57.7%) 503 (59.5%) 598 (63.7%) 419 (65.6%) 245 (68.4%) 849 (75.8%) 276 (73.8%) 71 (75.5%) Education, n (%) Elementary 93 (7.14%) 61 (5.84%) 45 (5.32%) 150 (16.1%) 95 (14.9%) 48 (13.4%) 347 (31.7%) 95 (25.7%) 22 (23.4%) High school 561 (43.1%) 445 (42.6%) 346 (40.9%) 514 (55.1%) 343 (53.7%) 189 (52.8%) 576 (52.6%) 197 (53.4%) 53 (56.4%) University 648 (49.8%) 539 (51.6%) 455 (53.8%) 269 (28.8%) 201 (31.5%) 121 (33.8%) 173 (15.8%) 77 (20.9%) 19 (20.2%) # chronic diseases, mean (SD) 2.72 (1.78) 4.87 (2.78) 7.70 (3.57) 4.24 (2.28) 7.71 (3.46) 12.0 (4.56) 5.47 (2.51) 9.70 (3.58) 14.2 (4.41) # drugs, mean (SD) 2.66 (2.77) 4.13 (3.37) 5.18 (3.92) 4.39 (3.42) 6.10 (3.92) 7.44 (4.50) 5.37 (3.48) 7.25 (3.97) 8.47 (4.46) Walking speed, mean (SD) 1.26 (0.31) 1.20 (0.35) 1.08 (0.35) 1.00 (0.38) 0.79 (0.41) 0.66 (0.41) 0.54 (0.41) 0.43 (0.36) 0.37 (0.35) MMSE, mean (SD) 29.3 (1.45) 28.7 (1.59) 28.5 (2.27) 28.4 (3.32) 26.9 (4.25) 25.5 (5.73) 24.8 (7.39) 24.1 (6.99) 21.9 (8.64) Abbreviations: MMSE, Mini Mental State Examination; SD, standard deviation. Among sexagenarians, subjects in the Unspecific towards smaller ones over time, and the smallest ones pattern were the youngest across all follow-ups, while (i.e., presumably containing the sickest subjects) tended those in the Cardiovascular and anemia pattern were to become larger over time. For example, among the oldest (Supplementary Table 3). Subjects in the sexagenarians, subjects in the Unspecific pattern Cardio-metabolic pattern were more frequently male represented 80% of the study population at baseline, but while those in the Psychiatric-endocrine and sensorial the figure went down to 52.4% after 6 years and to pattern were more likely to be female. Subjects in the 22.6% after 12 years. The prevalence of the death and latter pattern showed the highest level of education. dropout patterns increased in older age groups; an important part of the transitions among octogenarians Among septuagenarians, subjects in the Unspecific and beyond were in fact towards death. pattern were the youngest, while those in the Neuro- vascular and skin-sensorial pattern were the oldest. The estimated mean permanence times were computed Subjects in the Cardiovascular and diabetes pattern for each age group. As an example, for sexagenarians were more frequently male while those in the Neuro- belonging to the Cardiovascular and anemia pattern at vascular and skin-sensorial and Neuro-psychiatric and baseline, it was estimated that they would remain in the sensorial patterns were more likely to be female. same pattern for a mean time of 14.9 years before Subjects in the Cardiovascular and diabetes pattern had transitioning to other patterns. In all age groups, the the lowest proportion of university education. Unspecific patterns showed the shortest sojourn times, and the Psychiatric-endocrine and sensorial, Neuro- In the group of octogenarians and beyond, those in the vascular and skin-sensorial and Neuro-sensorial were Respiratory-circulatory and skin pattern were the the patterns with the longest sojourn time for youngest, while those in the Cardio-respiratory and sexagenarians, septuagenarians and octogenarians and neurological were the oldest. All patterns had a higher beyond, respectively. proportion of females. Subjects in the Neuro-sensorial pattern showed the highest level of education. The transition probability matrices by age group are shown in Figure 2. Regarding the interpretation of these Evolution and transitions across multimorbidity probabilities, the models show that, for example, patterns sexagenarians belonging to the Unspecific pattern at baseline had a probability of 0.9% of transitioning to the The evolution and transitions of and among Cardiovascular and anemia pattern and of 20.0% of multimorbidity patterns are graphically represented staying in the same pattern in the next 12 years. In through river plots in Figure 1. For all age groups, general, sexagenarians showed the highest levels of pattern prevalence varied over time, showing that people stability, as the probabilities of staying in the same commonly transition from one pattern to another. A pattern were higher than in the other age groups. More general trend was that the most represented patterns at specifically, among sexagenarians, the most likely baseline (i.e., containing the healthiest subjects) evolved transition between patterns was from the Unspecific to www.aging-us.com 9807 AGING Figure 1. Evolution and transitions of multimorbidity patterns over time by age group (N=3,363). Sexagenarians: Unspecific (Unsp); Cardiovascular and anemia (CV and Anemia); Cardio-metabolic (Cardio-Meta) and Psychiatric-endocrine and sensorial (Psy-Endoc and Sens). Septuagenarians: Unspecific (Unsp); Cardiovascular and diabetes (CV and Diab); Neuro-vascular and skin-sensorial (NeuroVasc and Skin); and Neuro-psychiatric and sensorial (NeuroPsy and Sens). Octogenarians and beyond: Unspecific (Unsp); Respiratory-circulatory and skin (Resp-Circula and Skin); Cardio-respiratory and neurological (CardioResp and Neuro); and Neuro-sensorial (Neuro-Sens). www.aging-us.com 9808 AGING Figure 2. Transition probability matrices by age group from baseline to the 12-year follow-up (N=3,363). Sexagenarians: Unspecific (Unsp); Cardiovascular and anemia (CV and Anemia); Cardio-metabolic (Cardio-Meta) and Psychiatric-endocrine and www.aging-us.com 9809 AGING sensorial (Psy-Endoc and Sens). Septuagenarians: Unspecific (Unsp); Cardiovascular and diabetes (CV and Diab); Neuro -vascular and skin-sensorial (NeuroVasc and Skin); and Neuro-psychiatric and sensorial (NeuroPsy and Sens). Octogenarians and beyond: Unspecific (Unsp); Respiratory-circulatory and skin (Resp-Circula and Skin); Cardio-respiratory and neurological (CardioResp and Neuro); and Neuro-sensorial (Neuro-Sens). the Psychiatric-endocrine and sensorial pattern (30.0%) MMSE in all age groups. While subjects in the after 12 years. Among septuagenarians, the most likely Unspecific patterns showed the slowest changes over transition was from the Unspecific to the Neuro- time, except for octogenarians, those in the patterns psychiatric and sensorial pattern (24.0%) after 12 years. characterized by cardiovascular and/or neurological Finally, in octogenarians and beyond, the transition diseases showed the worse baseline values and fastest from the Unspecific to the Cardio-respiratory and declines for all studied variables. neurological pattern (5.0%) after 12 years was the likeliest. The Cardiovascular and anemia, Neuro- DISCUSSION vascular and skin-sensorial, and Respiratory-circulatory and skin patterns showed the highest probabilities of In this study we identified and characterized longitudinal transitioning to death after 12 years in the three age multimorbidity patterns among older adults from a groups, respectively. Swedish urban population, and estimated the time they spent in each pattern as well as the probability of Characterization of multimorbidity patterns transitioning across different patterns throughout a 12-year follow-up period. Estimations of the longitudinal trends (predicted values from linear mixed models) for different clinical and Our findings highlight the dynamism and heterogeneity functional variables by patterns and for each age group underlying multimorbidity. The dynamism among are shown in Figure 3. An increasing trend was observed multimorbidity patterns was reflected by the for the number of chronic conditions and drugs across varying sojourn times across patterns, which differed age groups, with subjects in the Unspecific patterns by age group, and the specific patterns people consistently showing the lowest values. Conversely, a presented with. In sexagenarians, the average time decreasing trend was observed for walking speed and was 13.3 years, while in octogenarians and beyond, it Figure 3. Longitudinal trends (predicted values from linear mixed models) in clinical and functional characteristics associated with the multimorbidity patterns by age group (N=3,363). Sexagenarians: Unspecific (Unsp); Cardiovascular and anemia (CV and Anemia); Cardio-metabolic (Cardio-Meta) and Psychiatric-endocrine and sensorial (Psy-Endoc and Sens). Septuagenarians: Unspecific (Unsp); Cardiovascular and diabetes (CV and Diab); Neuro-vascular and skin-sensorial (NeuroVasc and Skin); and Neuro-psychiatric and sensorial (NeuroPsy and Sens). Octogenarians and beyond: Unspecific (Unsp); Respiratory-circulatory and skin (Resp-Circula and Skin); Cardio- respiratory and neurological (CardioResp and Neuro); and Neuro-sensorial (Neuro-Sens). MMSE: Mini Mental State Examination. www.aging-us.com 9810 AGING was 6.5 years. This observation implies that, as expected, progression and multimorbidity pattern trajectories the time of permanence in each pattern is greater in the using primary care electronic health records in the younger age groups, especially when less burdensome United Kingdom [15] and the Netherlands [16]. The patterns are at play. For example, the Unspecific pattern studies by Strauss et al. [15] and Lappenschaar et al. was characterized in all age groups by a lack of [16] were carried out on adult populations, older than overrepresentation of any of the low-severity chronic 35 and 50, and with a follow-up period of 3 and 5 conditions the pattern was composed of (e.g., cardio- years, respectively; and both included a lower number vascular risk factors, osteoarthritis, hearing impairment, of chronic diseases than that used in this study. In etc.). Consequently, people belonging to this pattern terms of the analytical approach, the latent class could be regarded as being the healthiest, and thus the growth models employed by Strauss et al. are designed target for primary and secondary preventive strategies. to identify longitudinal trajectories, but one cannot Indeed, almost one third of sexagenarians in the infer transitions among classes. Also, Lappenschaar et Unspecific pattern at baseline transitioned to the al. used multilevel temporal Bayesian networks, which Psychiatric and sensorial pattern, and almost one in ten are aimed at analyzing the relationships between to the Cardio-metabolic pattern during the follow-up. diseases (i.e., networks) but not the transitions across The heterogeneity of multimorbidity was evidenced by clusters. Other studies [17, 18] have focused on the the different patterns obtained within, but especially, incidence of new chronic diseases across time, but across age groups. Despite being similar, patterns at failed to examine patterns of multimorbidity. In brief, different ages represent different states of the disease none of the previously applied statistical methods severity continuum. These different stages may be makes it possible to study the evolution and transitions associated with differential probabilities of developing between patterns of multimorbidity. In contrast, when complications and functional decline, and may trigger applying HMM, one can explore the variability of different pharmacological and non-pharmacological chronic disease evolution over time by considering treatments. In relation to mortality, trajectories each subject’s diseases as random variables characterized by cardiovascular and circulatory diseases conditioned by a hidden or conglomerate state, which were found to concentrate the highest death probabilities. further enables depicting people’s transitions among All these aspects may contribute to increase the different patterns of multimorbidity. Other studies heterogeneity of the multimorbidity landscape. looking at multimorbidity patterns within large databases have considered disease trajectories rather Moreover, our study serves as an example of how than individual trajectories as the main axis of interest longitudinal data may be used to explore the trajectories [19]. This approach, which is somewhat disease- rather of multimorbidity – that is, the evolution of and than person-oriented, is limited by the inability to transitions among patterns of diseases. To date, studies identify homogenous groups of patients. Another on patterns of multimorbidity have predominantly example is the work by Giannoula et al., which focused on analyzing the association between diseases, focused on the identification of complex time- paying less attention to individuals’ “journeys” in and dependent disease associations using dynamic time out of these patterns [5, 10]. This is mainly because most warping, a machine learning technique [20]. Similar studies, even those using longitudinal data [11], were problems are present in the study by Xu et al., which based on cross-sectional designs. Indeed, studies moreover only considered three pathologies [21]. incorporating the entire longitudinal structure of the data are scarce [12–14]. Studying patterns of multimorbidity This study has several strengths. First, thanks to longitudinally is a challenging endeavor given that the the exhaustive clinical evaluation that SNAC-K heterogeneity in disease clustering originates both from participants undergo in each follow-up wave, the the cross-sectional and longitudinal axes. Therefore, to reliability of the diagnostic data, which moreover understand the interdependence among diseases when integrates data from electronic health records, lab tests looking at longitudinal multimorbidity patterns, dynamic and drug use, is optimal. Second, the statistical machine learning methodologies such as the HMM are methods applied allowed us to cluster people by their required. These models integrate a dynamic Bayesian co-occurring diseases taking both the cross-sectional network that accounts for the temporal sequence of the and longitudinal axes into account: HMM and the person-level data observed. This allows considering the fuzzy c-means cluster algorithm. The latter is the longitudinal structure of the data (i.e., time series) and choice method for pattern recognition when clusters the correlations among observations. tend to overlap, which is often the case as older adults show a high prevalence of co-occurring conditions. Comparing our results with those from previous Furthermore, in this study we were able to explore studies is difficult for the reasons mentioned above. longitudinal multimorbidity patterns by age group and Nevertheless, two previous studies analyzed disease the time that people remained in each pattern. As far as www.aging-us.com 9811 AGING we know, these aspects have not been previously expanded, for instance, by aggregating information on studied and are key to personalized clinical decision- complementary health indicators such as frailty and making. Moreover, by stratifying our study sample by biological and physiological variables, which could decade age groups, we were able to account for the further optimize patient stratification and management selection bias inherent to aging cohorts, whereby the efforts. oldest age groups tend to represent healthier individuals characterized by better biological and Our study provides evidence that multimorbidity is environmental living conditions. dynamic and heterogeneous in old age. With increasing age, older adults experience decreasing clinical stability Some limitations must also be considered. First, the and progressively shorter permanence time within one relatively small size of the SNAC-K cohort and the same multimorbidity pattern. Moreover, a significant further stratification of the study sample into three proportion ranging between 5.9%-22.6% belongs to an different age groups led to some of the patterns including Unspecific pattern with a low burden of diseases few people (i.e., <14 people). However, the methods and a promising preventive potential. Adding new applied have been shown to be responsive enough for the variables related to drug use, environmental and genetic identification of subgroups of people even in small factors, and/or frailty to the longitudinal analysis of samples. Additionally, the iterative estimation process multimorbidity patterns may allow optimizing the and the number of realizations allowed us to maximize epidemiological understanding and applicability of these the likelihood of the models applied given the data. models for patient-tailored prevention and management Second, participant dropout (14% within the first 6 years strategies. and 8% within the next 6 years) may have affected the cluster definition process. Still, to the best of our MATERIALS AND METHODS knowledge, this is an exceptionally low figure compared with studies of this type. Third, the discontinuous follow- Study population up carried out in SNAC-K (i.e., every 3 or 6 years depending on the age of participants) may have affected Longitudinal data from the population-based Swedish the rate of disease detection and, consequently, the National study on Aging and Care in Kungsholmen longitudinal cluster analysis, especially among people (SNAC-K) was used [22]. The study population who died or dropped out during the observation consisted of adults ≥60 years of age living in the period. To adapt to the assumptions of our study community or in institutions in the Kungsholmen district design, participant data were analyzed in accordance of Stockholm, Sweden. A random sample of 11 age with the available follow-up waves, avoiding any data cohorts (ages 60, 66, 72, 78, 81, 84, 87, 90, 93, 96 and extrapolation. Last, differences in the baseline ≥99) born between 1898 and 1943 (the youngest and composition and evolution of patterns across age groups oldest age cohorts were oversampled) was invited to could be due to variations in exposure history, and not participate in the study. People who agreed to participate only to age, given that there is up to 40 years of a gap were evaluated for the first time between 2001 and 2004. between the youngest and oldest subjects at study Participants who were <78 years of age were then baseline. followed up every six years and participants ≥78 every three years. The present study is based on data collected The analysis of longitudinal multimorbidity patterns is at baseline, the six-year follow-up, and the 12-year fundamental for the provision of personalized medical follow-up. At baseline, 3363 people were examined care that is not based merely on the application of (participation rate: 73%). For our study, the sample was guidelines targeting each chronic condition individually. stratified into three age groups: sexagenarians (age While some of our findings can be explained through cohorts of 60 and 66 years), septuagenarians (age cohorts known pathophysiological mechanisms, others may serve of 72 and 78 years) and octogenarians and beyond (age to generate new hypotheses worth exploring in future cohorts of 81 years and over). studies. Our statistical approach enabled us to model the evolution and transitions of multimorbidity over time, Chronic diseases and the results of this could be applied in the interests of healthier aging. Moreover, the age-stratified analyses At each follow-up wave, SNAC-K participants undergo allowed us to identify which disease combinations and an approximately five-hour-long comprehensive clinical transitions were more prevalent in each decade. This and functional assessment carried out by trained information is key to defining specific care plans to physicians, nurses, and neuropsychologists. Physicians prevent or delay the negative consequences of the most collect information on diagnoses via physical frequent diseases identified. The characterization of examination, medical history, examination of medical multimorbidity patterns using HMM could moreover be charts, self-reported information, and/or proxy www.aging-us.com 9812 AGING interviews. Clinical parameters, lab tests, drug dimensionality of the longitudinal dataset. To prevent information, and inpatient and outpatient care data are statistical noise and spurious findings from the models, also used to identify specific conditions. All diagnoses only diseases that achieved a median prevalence of 2% are coded in accordance with the International across all follow-up waves were included (Supplementary Classification of Diseases, 10th revision (ICD-10). In Table 1). Afterwards, a fuzzy segmentation procedure the current study we classified all the ICD-10 codes into (Fuzzy C-means algorithm, FCM) [11] was applied on the 60 chronic disease categories in accordance with a new dataset to identify an initial set of clusters, which was clinically driven methodology [23]. In SNAC-K, drugs used to initialize some of the HMM parameters in the next are coded in accordance with the Anatomical stage. Finally, two more clusters were added in order to Therapeutic Chemical (ATC) classification. account for dropout and/or death. Covariates The set of HMM parameters, composed of the initial cluster probabilities, the inter-cluster transition Information on demographics (age, sex, education) was probabilities and the emission distributions provided by collected during nurse interviews. We divided education the FCM, were fitted into the observation dataset by into elementary, secondary, university or higher. applying the Baum-Welch (BW) algorithm. This made Information about vital status was derived from death it possible to infer the longitudinal trajectories followed certificates provided by Statistics Sweden, the Swedish by each individual. The best cluster trajectory was governmental statistics agency. Survival status was identified by maximizing the probability of the assessed throughout the follow-up period. Participants observed sequence conditioned to the computed model were considered lost to follow-up if they or a proxy parameters (Viterbi Algorithm). To validate the model, declined to participate, could not be contacted, had a comparison between BW and Viterbi transition moved out of the study area, or cancelled an probability matrices was conducted, showing a assessment. Walking speed (m/s) was assessed by good agreement between theoretical and observed asking participants to walk 6 m at their usual speed or values [26]. 2.44 m if the participant reported walking quite slowly [24]. Cognitive status was assessed by physicians using The time unit considered for each transition across the Mini-Mental State Examination (MMSE), with a clusters/states was the time between follow-up waves, 6 score range of 30 at best to 0 at worst [25]. years for sexagenarians and septuagenarians and 3 years for octogenarians and beyond. The time spent in a Statistical analysis specific cluster/state before moving to other clusters/ states was assumed to follow a geometric distribution. The sample characteristics at baseline, the 6-year Subsequently, the expected average time spent or mean follow-up and the 12-year follow-up for all age groups sojourn (permanence) time was computed. were described as appropriate. Additionally, 3-year and 9-year follow-up data was considered for the group of To optimize the performance of the selected octogenarians and beyond. mathematical model, the iterative process involved in the application of the BW algorithm was initialized To model the temporal evolution of multimorbidity using a range of 100 different values of the parameters patterns and individuals’ transitions across these to be learned. The best model was selected using a patterns, a dynamic random process represented by a procedure that is equivalent to applying the Bayes HMM was assumed [9]. Disease information from all Information Criterion to choose the best set of HMM individuals and across all follow-up waves is used by the parameters [9]. HMM to identify so-called hidden states (i.e., longitudinal multimorbidity pattern). HMM estimates Multimorbidity patterns the transition probabilities between patterns, i.e., the probability that any individual moves from one pattern For each age group, a final number of longitudinal to another in a given time-frame. Furthermore, by using patterns was selected. To evaluate the consistency and HMM, one can examine individuals’ probability of utility of the final clusters, we contrasted the clinical following different longitudinal multimorbidity patterns, relevance of our findings in the context of previous and subsequently identify the one that is most likely to literature, and we dicussed the findings within the happen. research team (2 GPs, 2 geriatricians, 3 epidemiologists and 2 statisticians). The dataset was pre-processed by applying a Multiple Correspondence Analysis (MCA) to the categorical To characterize the multimorbidity patterns, we features (i.e., diseases), in order to reduce the calculated the frequency of chronic diseases in each www.aging-us.com 9813 AGING cluster. Observed/expected ratios (O/E-ratios) were FUNDING calculated by dividing the prevalence of a given disease within a cluster by its prevalence in the overall This work was supported by the funders of the population. The exclusivity of different diseases, Swedish National study on Aging and Care (SNAC): defined as the fraction of participants with the disease in the Ministry of Health and Social Affairs, Sweden; the cluster over the total number of participants with the the participating county councils and municipalities; disease, was also calculated. We considered a disease to and the Swedish Research Council. Specific grants be associated with a given cluster of individuals when were obtained from the Swedish Research Council the O/E ratio was ≥2 or the exclusivity was ≥ 20% [12]. for Medicine (VR; 521-2013-8676; 2016-00981; Such criteria were used to name multimorbidity patterns 2017-06088; 2021-03324) and the Swedish Research after the diseases that predominantly characterized Council for Health, Working Life and Welfare them. (Forte; 2016-07175; 2017-01764). This work was supported by Spain’s Ministry of Science and The longitudinal trends of clinical and functional Innovation through the Carlos III Health Institute and characteristics (no. of chronic diseases, no. of drugs, by European Union ERDF (European Regional walking speed and MMSE) associated with the Development Fund) funds through the Research multimorbidity patterns were estimated through linear Network in Preventive Activities and Health mixed models, assuming a random intercept and Promotion in Primary Care (redIAPP, RD16/0007/ including an interaction between the patterns and 0001). This study has also been funded by the Carlos follow-up time, both as linear and quadratic. The III Health Institute through the project “PI19/00535” models were additionally adjusted by age, sex and (co-funded by the ERDF; “A way to make Europe”). education. The IDIAP Jordi Gol also supported this work by offering the first author office space and access to the The analyses were carried out using Stata version 17 institution’s resources throughout the research and R version 4.1.2. The significance level was set at process. α=0.05. REFERENCES AUTHOR CONTRIBUTIONS 1. Kontis V, Bennett JE, Mathers CD, Li G, Foreman K, ARL, DLV, CV and ACL developed the study Ezzati M. Future life expectancy in 35 industrialised concept and design. ARL performed the data analysis, countries: projections with a Bayesian model and DLV, CV and ACL contributed to the ensemble. Lancet. 2017; 389:1323–35. interpretation of the results. ARL drafted the https://doi.org/10.1016/S0140-6736(16)32381-9 manuscript. All authors provided critical revisions and PMID:28236464 approved the final version of the manuscript for 2. Oeppen J, Vaupel JW. Demography. Broken limits to submission. life expectancy. Science. 2002; 296:1029–31. https://doi.org/10.1126/science.1069675 ACKNOWLEDGMENTS PMID:12004104 We thank the SNAC-K participants and the SNAC-K 3. Chudasama YV, Khunti K, Gillies CL, Dhalwani NN, Group for their collaboration in data collection and Davies MJ, Yates T, Zaccardi F. Healthy lifestyle and life management. expectancy in people with multimorbidity in the UK Biobank: A longitudinal cohort study. PLoS Med. 2020; CONFLICTS OF INTEREST 17:e1003332. https://doi.org/10.1371/journal.pmed.1003332 The authors have no conflicts. PMID:32960883 4. González-González AI, Meid AD, Dinh TS, Blom JW, van ETHICAL STATEMENT AND CONSENT den Akker M, Elders PJ, Thiem U, De Gaudry DK, Swart KM, Rudolf H, Bosch-Lenders D, Trampisch HJ, The study was approved by the Regional Ethics Meerpohl JJ, et al. A prognostic model predicted Review Board in Stockholm. Participants in the study deterioration in health-related quality of life in older completed a written informed consent form as patients with multimorbidity and polypharmacy. J Clin stipulated by the ethics board. For participants with Epidemiol. 2021; 130:1–12. prevalent or incident cognitive impairment, consent https://doi.org/10.1016/j.jclinepi.2020.10.006 was obtained from next of kin. PMID:33065164 www.aging-us.com 9814 AGING 5. Prados-Torres A, Calderón-Larrañaga A, Hancco- https://doi.org/10.1186/s12877-018-0705-7 Saavedra J, Poblador-Plou B, van den Akker M. PMID:29338690 Multimorbidity patterns: a systematic review. J Clin 13. Violán C, Foguet-Boreu Q, Fernández-Bertolín S, Epidemiol. 2014; 67:254–66. Guisado-Clavero M, Cabrera-Bean M, Formiga F, https://doi.org/10.1016/j.jclinepi.2013.09.021 Valderas JM, Roso-Llorach A. Soft clustering using real- PMID:24472295 world data for the identification of multimorbidity 6. Violan C, Foguet-Boreu Q, Flores-Mateo G, Salisbury C, patterns in an elderly population: cross-sectional study Blom J, Freitag M, Glynn L, Muth C, Valderas JM. in a Mediterranean population. BMJ Open. 2019; Prevalence, determinants and patterns of 9:e029594. multimorbidity in primary care: a systematic review of https://doi.org/10.1136/bmjopen-2019-029594 observational studies. PLoS One. 2014; 9:e102149. PMID:31471439 https://doi.org/10.1371/journal.pone.0102149 14. France EF, Wyke S, Gunn JM, Mair FS, McLean G, PMID:25048354 Mercer SW. Multimorbidity in primary care: a 7. Ho IS, Azcoaga-Lorenzo A, Akbari A, Black C, Davies J, systematic review of prospective cohort studies. Br J Hodgins P, Khunti K, Kadam U, Lyons RA, McCowan C, Gen Pract. 2012; 62:e297–307. Mercer S, Nirantharakumar K, Guthrie B. Examining https://doi.org/10.3399/bjgp12X636146 variation in the measurement of multimorbidity in PMID:22520918 research: a systematic review of 566 studies. Lancet 15. Strauss VY, Jones PW, Kadam UT, Jordan KP. Distinct Public Health. 2021; 6:e587–97. trajectories of multimorbidity in primary care were https://doi.org/10.1016/S2468-2667(21)00107-9 identified using latent class growth analysis. J Clin PMID:34166630 Epidemiol. 2014; 67:1163–71. 8. Cezard G, McHale CT, Sullivan F, Bowles JK, Keenan K. https://doi.org/10.1016/j.jclinepi.2014.06.003 Studying trajectories of multimorbidity: a systematic PMID:25063556 scoping review of longitudinal approaches and 16. Lappenschaar M, Hommersom A, Lucas PJ, Lagro J, evidence. BMJ Open. 2021; 11:e048485. Visscher S, Korevaar JC, Schellevis FG. Multilevel https://doi.org/10.1136/bmjopen-2020-048485 temporal Bayesian networks can model longitudinal PMID:34810182 change in multimorbidity. J Clin Epidemiol. 2013; 9. Violán C, Fernández-Bertolín S, Guisado-Clavero M, 66:1405–16. Foguet-Boreu Q, Valderas JM, Vidal Manzano J, Roso- https://doi.org/10.1016/j.jclinepi.2013.06.018 Llorach A, Cabrera-Bean M. Five-year trajectories of PMID:24035172 multimorbidity patterns in an elderly Mediterranean population using Hidden Markov Models. Sci Rep. 17. Vos R, van den Akker M, Boesten J, Robertson C, 2020; 10:16879. Metsemakers J. Trajectories of multimorbidity: https://doi.org/10.1038/s41598-020-73231-9 exploring patterns of multimorbidity in patients with PMID:33037233 more than ten chronic health problems in life course. BMC Fam Pract. 2015; 16:2. 10. Ng SK, Tawiah R, Sawyer M, Scuffham P. Patterns of https://doi.org/10.1186/s12875-014-0213-6 multimorbid health conditions: a systematic review of PMID:25608728 analytical methods and comparison analysis. Int J Epidemiol. 2018; 47:1687–704. 18. Ruel G, Lévesque JF, Stocks N, Sirois C, Kroger E, Adams https://doi.org/10.1093/ije/dyy134 RJ, Doucet M, Taylor AW. Understanding the evolution PMID:30016472 of multimorbidity: evidences from the North West Adelaide Health Longitudinal Study (NWAHS). PLoS 11. Vetrano DL, Roso-Llorach A, Fernández S, Guisado- One. 2014; 9:e96291. Clavero M, Violán C, Onder G, Fratiglioni L, Calderón- https://doi.org/10.1371/journal.pone.0096291 Larrañaga A, Marengoni A. Twelve-year clinical PMID:24798485 trajectories of multimorbidity in a population of older adults. Nat Commun. 2020; 11:3223. 19. Jensen AB, Moseley PL, Oprea TI, Ellesøe SG, Eriksson https://doi.org/10.1038/s41467-020-16780-x R, Schmock H, Jensen PB, Jensen LJ, Brunak S. PMID:32591506 Temporal disease trajectories condensed from population-wide registry data covering 6.2 million 12. Guisado-Clavero M, Roso-Llorach A, López-Jimenez T, patients. Nat Commun. 2014; 5:4022. Pons-Vigués M, Foguet-Boreu Q, Muñoz MA, Violán C. https://doi.org/10.1038/ncomms5022 PMID:24959948 Multimorbidity patterns in the elderly: a prospective cohort study with cluster analysis. BMC Geriatr. 2018; 20. Giannoula A, Gutierrez-Sacristán A, Bravo Á, 18:16. Sanz F, Furlong LI. Identifying temporal patterns in www.aging-us.com 9815 AGING patient disease trajectories using dynamic time Rizzuto D, Welmer AK, et al. Assessing and Measuring warping: A population-based study. Sci Rep. 2018; Chronic Multimorbidity in the Older Population: A 8:4216. Proposal for Its Operationalization. J Gerontol A Biol Sci https://doi.org/10.1038/s41598-018-22578-1 Med Sci. 2017; 72:1417–23. PMID:29523868 https://doi.org/10.1093/gerona/glw233 PMID:28003375 21. Xu X, Mishra GD, Dobson AJ, Jones M. Progression of diabetes, heart disease, and stroke multimorbidity in 24. Studenski S, Perera S, Patel K, Rosano C, Faulkner K, middle-aged women: A 20-year cohort study. PLoS Inzitari M, Brach J, Chandler J, Cawthon P, Connor EB, Med. 2018; 15:e1002516. Nevitt M, Visser M, Kritchevsky S, et al. Gait speed and https://doi.org/10.1371/journal.pmed.1002516 survival in older adults. JAMA. 2011; 305:50–8. PMID:29534066 https://doi.org/10.1001/jama.2010.1923 PMID:21205966 22. Lagergren M, Fratiglioni L, Hallberg IR, Berglund J, Elmståhl S, Hagberg B, Holst G, Rennemark M, Sjölund 25. Folstein MF, Folstein SE, McHugh PR. “Mini-mental BM, Thorslund M, Wiberg I, Winblad B, Wimo A. A state”. A practical method for grading the cognitive longitudinal study integrating population, care and state of patients for the clinician. J Psychiatr Res. 1975; social services data. The Swedish National study on 12:189–98. Aging and Care (SNAC). Aging Clin Exp Res. 2004; https://doi.org/10.1016/0022-3956(75)90026-6 16:158–68. PMID:1202204 https://doi.org/10.1007/BF03324546 26. Rabiner LR. A Tutorial on Hidden Markov Models and PMID:15195992 Selected Applications in Speech Recognition. Proc IEEE. 23. Calderón-Larrañaga A, Vetrano DL, Onder G, Gimeno- 1989; 77:257–86. Feliu LA, Coscollar-Santaliestra C, Carfí A, Pisciotta MS, https://doi.org/10.1109/5.18626 Angleman S, Melis RJ, Santoni G, Mangialasche F, www.aging-us.com 9816 AGING SUPPLEMENTARY MATERIALS Supplementary Tables Please browse Full Text version to see the data of Supplementary Tables 1–3. Supplementary Table 1. Disease prevalence by age group and follow-up wave. Supplementary Table 2. Description of multimorbidity patterns in terms of the top 10 diseases characterizing them by age group and follow-up wave. Supplementary Table 3. Description of multimorbidity patterns in terms of sociodemographic, clinical and functional characteristics by age group and follow-up wave. www.aging-us.com 9817 AGING