Paths to multimorbidity: a longitudinal perspective on disease accumulation in Catalonia, Spain

Abstract

With increasing life expectancy, multimorbidity represents a growing global challenge, affecting quality of life.

We analyzed electronic health records of individuals aged 45–74 from Catalonia, Spain (2007–2021), who were healthy at the end of 2007. We use sequence and cluster analysis to describe and categorize monthly disease accumulation patterns, from healthy to one or multiple conditions or death, involving four disease groups with the highest global morbidity and mortality burdens. We further investigate the association between identified clusters and sociodemographic factors and the association of cluster membership with healthcare utilization.

Approximately 36% of individuals remained healthy throughout the study, while the remainder transitioned to single or multiple morbidity and/or died. A higher number of conditions in a given month increased transition probability, with metabolic and hypertensive conditions being the most common entry points. We identified nine disease accumulation trajectories linked to sociodemographic characteristics: Women were more likely to be in clusters involving neurodegenerative conditions and men in those involving cardiovascular and cancer conditions. Higher-income individuals were more likely to be in lower morbidity clusters, except the cancer multimorbidity cluster. Healthcare utilization was elevated in all clusters relative to the healthy group, with notably higher emergency use in cardiovascular clusters and more hospital admissions in cancer-related clusters. The study shows that disease accumulation follows identifiable patterns linked to sociodemographic factors and underscores the propagative nature of multimorbidity. Further, healthcare utilization is shaped more by condition type than by the number, highlighting the need for targeted health services.

Introduction

The co-existence of chronic diseases, or multimorbidity, represents a growing challenge worldwide, affecting more than half of the global population above 50 years of age [1]. The burden is only expected to rise, given the global advancement in life expectancy, in many instances outpacing improvements in healthy life expectancy [2–4], and the rapidly growing number of people with or at risk of long-term conditions, such as diabetes and cancer [5, 6].

Public health approaches to disease prevention, management, and treatment were historically designed to address single isolated conditions, making them ill-suited to address the growing challenge of rising multimorbidity [7]. The co-existence of disease implies complex health needs with implications for disease management, an individual’s quality of life, healthcare utilization, and system-level costs. This in turn mandates urgent prioritization of multimorbidity research in order to guide the adaptation of health systems and improve the prevention and management of chronic conditions globally.

Studies show that sociodemographic factors play a critical role in disease accumulation. Socioeconomic status and rural versus urban residence have all been associated with differing risks of developing multiple chronic diseases [8, 9]. These findings suggest that the social context can play a significant role in how diseases accumulate, and thus, a classification approach that accounts for these factors could yield actionable insights. Multimorbidity is increasingly understood as a dynamic and socially stratified process in which diseases interact biologically while unfolding within unequal social contexts that shape exposure, vulnerability, and access to care [10, 11]. This perspective highlights the need for analytical approaches that capture how chronic conditions evolve not only over time but also across social and demographic strata [12].

A key limitation of current multimorbidity research is its reliance on cross-sectional data, which fails to capture the dynamic evolution of disease patterns over time [12]. Some longitudinal studies [12] have attempted to address this gap by analyzing the sequence and timing of disease accumulation; however, there remains a need to explore whether similar patterns hold across different populations. A broader perspective is necessary to determine the extent to which multimorbidity trajectories are influenced by biological versus social determinants of health. Comparative studies across different healthcare systems and demographic settings are essential to generalize findings and inform globally relevant policies and clinical strategies.

Healthcare utilization is known to increase with a higher number of conditions, and this relationship is observed across primary, specialist, and emergency care [13]. However, simply counting conditions may mask differences in how specific diseases co-occur, the sequence in which they emerge, and the social contexts in which they develop. These “qualitative” aspects of multimorbidity can significantly shape care pathways and resource needs, ultimately influencing both clinical outcomes and healthcare costs. In turn, a more nuanced understanding of how disease accumulation patterns drive healthcare utilization is useful for designing targeted interventions and allocating resources more effectively.

The current study focuses on multimorbidity in Catalonia, Spain; however, its findings have broader implications for aging populations worldwide. Catalonia serves as an ideal setting for longitudinal multimorbidity research due to its comprehensive public healthcare system, extensive electronic health records (EHRs), and high life expectancy [14], which is associated with a higher risk of multimorbidity. Previous studies have demonstrated significant variations in multimorbidity prevalence and healthcare utilization within the region [15], suggesting that findings from Catalonia may generalize to other aging societies with similar healthcare infrastructures. By identifying commonalities and differences in multimorbidity progression across populations, this study contributes to the global discourse on multimorbidity management and healthcare system adaptation. The insights generated can inform policy discussions. Additionally, as EHRs become more comprehensive and available for public health research, our study demonstrates ways by which this data can be used to derive insights relevant both clinically and on a policy level.

Overall, this study aims to contribute to the literature on multimorbidity by providing a longitudinal perspective on disease accumulation in older adults (aged 45–74) and applying sequence analysis methods to EHRs. Specifically, this study [1] describes individual trajectories of disease accumulation over time [2], classifies these trajectories into distinct groups based on the timing and sequence of diagnoses [3], identifies sociodemographic characteristics associated with each trajectory group, and [4] examines health system utilization across these groups. In general, this research seeks to deepen the understanding of multimorbidity progression and its associated social and healthcare implications, ultimately supporting more precise and equitable health interventions.

Methods

Data and health measures

We use the HEALIN cohort [16], a population-based dataset that follows a random sample of 1 551 126 individuals in Catalonia (Spain) (accounting for 22% of the total Catalan population) from 2005 up to 2021. This sample is representative in terms of age, sex, and region. All data within the HEALIN cohort are anonymized and comply with Spanish regulations regarding population-observational studies. This dataset contains information about all diseases that are registered in health records in Catalonia, resulting in thousands of unique ICD codes in the dataset.

We use sequence analysis methods (explained below), where the number of possible states is determined by the number of unique conditions we consider. To balance the inclusion of multiple conditions while keeping the number of unique states manageable, we group conditions based on the affected organ system. We categorize these conditions into four distinct groups that are highly relevant from a prevention perspective.

The first group (Metabolic/HTN) is comprised of hypertensive and metabolic disorders (diabetes & lipid disorders), which are highly prevalent, often affect multiple organ systems, have a lower mortality risk compared to other conditions, and are considered risk factors for the development of other diseases [17, 18]. The second group (CVD) is comprised of cardiovascular diseases, which account for most deaths from non-communicable diseases (NCDs) [19]. The third (Neuro) is comprised of neurodegenerative diseases, the leading causes of disability globally [20], and especially in aging populations. The fourth group (Cancer) was comprised of malignant (cancer) tumors and leukemia, the second leading cause of NCD deaths globally [19]. The conditions were identified using ICD-10 and ICD-9 codes listed in Supplementary Annex A. All possible combinations of these four groups, in addition to the healthy state and the death state, are coded as unique states, resulting in 17 unique states. Multimorbidity in our study refers to having more than one group of conditions. The “healthy” state is defined as the absence of any of the four condition groups presented, and not the absence of any disease. This definition was chosen to isolate the transition dynamics relevant to the NCD groups mentioned above.

Sociodemographic variables included are age, sex, and income. For income, the HEALIN cohort data provides a rough measure that uses the pharmacy copayment category a person falls in: “exempt” (low), “<18 000 euros per year” (low-middle), “18 000–100 000” (high-middle), and “>100 000 euros per year” (high).

For healthcare service use, we analyze the total number of primary healthcare (PHC) visits, emergency health visits, and hospital admissions during the observation period.

Sequence and cluster analysis

In this analysis, we include individuals aged 45–74 in 2005 who are healthy by the end of 2007. This way we allow for a 2-year washout period, meaning that individuals are considered healthy if they have not been diagnosed with any of the conditions in our study between 2005 and 2007. The washout period is used to reduce left-truncation issues. A longitudinal disease‐transition sequence was constructed for each individual from 2007 to 2021, recording monthly states from 2007 to 2021. Each person could remain in a healthy state throughout the entire observation period, transition into one, two, three, or four comorbid conditions, or move into death (an absorbing state). The transition unit was set to 1 month to capture some granularity while maintaining a manageable sequence length.

To group individuals with comparable disease‐accumulation and progression patterns, we applied sequence clustering based on the Longest Common Subsequence (LCS) measure. A common subsequence is any sequence of states that can be matched in order within both sequences, ignoring any gaps. The longest such subsequence is LCS. LCS is chosen because it is a robust, computationally efficient measure suited for long sequences and large datasets [21]. Given the large size of our dataset, we used the Clustering Large Applications (CLARA) algorithm, which was specifically designed for large datasets [22]. CLARA draws multiple subsamples, performs cluster analyses on each, and then combines the results to assign all individuals to the most suitable cluster. We explored a range of 2–15 clusters, and we ran 100 independent iterations, each time sampling 10 000 sequences for clustering and then classifying all sequences based on the resulting medoids. We employed the CRISP method, ensuring stable cluster assignment across iterations and reducing the risk of overfitting in large datasets (Supplementary Annex B).

Results are presented for nine clusters, which were determined based on standard sequence-clustering criteria (Supplementary Annex B).

Regression analysis

We use multivariate logistic regression to analyze the association between sociodemographic characteristics, namely, sex, age, and income level, and the odds of being a member of each of the clusters we identify in the cluster analysis. This allows us to investigate whether specific patterns of disease accumulation and progression are associated with specific characteristics.

Additionally, we use a negative binomial model to investigate the effect of cluster membership on the number of PHC visits, emergency visits, and admissions, adjusting for age, income, sex, and number of transitions, while considering time at risk.

Results

Description of the individual trajectories

From the HEALIN cohort of 1.5 million individuals, 255 121 participants aged 45–74 in 2005 were healthy by the end of 2007. Approximately 49.0% were female, with 67.0% aged 45–59 and 33.0% aged 60–74. Income data was available for 72.9% of participants; 42.1% were low-middle income, 27.5% were in the high-middle, 2.1% were low income, and 1.2% had high income. Individuals missing income data tended to be older and predominantly male. On average, participants had 52.9 PHC visits, 2.2 emergency visits, and 1.7 hospital admissions between 2007 and 2021.

All individuals started in a healthy state, but over time, single, multimorbidity, and death states increased. About 36.1% remained healthy throughout, 15.3% died, 22.5% experienced one transition ending in a single morbidity state, and 26.1% progressed to a multimorbid state (Supplementary Fig. SC1).

In all states, the most likely state in the following month is the same state (>90%). The most common initial transition out of a healthy state was to a metabolic/hypertension condition (0.3%). Notably, higher morbidity states show higher transition rates to further morbidity; e.g. the rate from healthy to a cardiovascular state was 0.1%, compared to 0.3% from a metabolic/hypertension state to a metabolic/hypertension + cardiovascular state (Supplementary Fig. SC2).

Figure 1 shows the mean time spent in each state by sex, age, and income. Younger cohorts spent more time in single morbidity states, most notably in the Metabolic/HTN only state (32.2 vs 18.4 months), while death and multimorbidity states dominated in the older cohort.

Figure 1.

Time spent in each state by demographic characteristics. This graph shows the average time spent in each state by individuals in each category, aggregated over the observation period (2007–2021). Source: HEALIN cohort.

Women on average spent more time in states involving neurodegenerative conditions, most notably in the Neuro-only state (3.1 vs 1.9 months), while men spent longer time in the death state than women (15.5 vs 9.1 months) and in most states involving cancer conditions, most notably in the state of cardiovascular and cancer multimorbidity (CVD+Cancer; 0.6 vs 0.4 months, respectively) (Fig. 1).

A clear income gradient was observed: higher-income groups spent more time healthy, while lower-income groups accumulated more time in morbidity states (e.g. Metabolic/HTN and Neuro state). Interestingly, cancer multimorbidity states (Metabolic/HTN and Cancer, CVD and Cancer) showed a unique pattern, with middle-income individuals recording the highest durations, whereas the cancer-only state exhibited an inverse gradient with higher-income groups. Those missing the income variable had the greatest time in both the healthy and death states but negligible time in other morbidity states, suggesting potential under‐ascertainment or misclassification of diagnoses in this subset.

Identifying sequence clusters

Using our clustering criteria, we identified nine optimal clusters (see Supplementary Annex B). The state distribution by month for each cluster is shown in Fig. 2. Cluster 1 (Healthy Group) includes those with no or a single late transition, representing 43.9% of participants. Cluster 2 (Early Metabolic/HTN) consists of individuals transitioning early to a Metabolic/HTN state, some of whom experienced additional transitions to multimorbidity or death. Cluster 3 (Late Metabolic/HTN) is similar to cluster 2 but with a later transition from health to a Metabolic/HTN condition. Clusters 4 (CVD-start) and 5 (Neuro-start) include those beginning with a cardiovascular or neurodegenerative condition, respectively, with about half progressing to multimorbidity or death. Clusters 6 (Metabolic/HTN+CVD), 7 (Metabolic/HTN+Neuro), and 8 (Metabolic/HTN+Cancer) include participants who, after a Metabolic/HTN condition, quickly develop a second condition and spend a significant period in a multimorbid state, with roughly a quarter advancing to further multimorbidity or death.

Figure 2.

State distribution by month and cluster from 2007–2021 for adults aged 45–74. Stacked area charts showing the proportion of individuals in each health or morbidity state in each month from 2007 to 2021 (inclusive). Source: HEALIN cohort.

Figure 3 presents multivariate regression results linking cluster membership with sociodemographic factors. Younger individuals had higher odds of being in healthy or low-morbidity clusters compared to older cohorts. Men were slightly less likely to belong to the healthy, Neuro-start, or Metabolic/HTN+Neuro clusters, but more likely to be in the High Mortality, Metabolic/HTN+Cancer, and Metabolic/HTN+CVD clusters. A negative income gradient (higher odds in lower income groups) was evident for all but two clusters, the Healthy group and the Metabolic/HTN+Cancer group. Those in the high-middle income group had 53% higher odds of being in the Healthy cluster compared to the low-income group (95% CI 1.44–1.63) and 37% higher odds of being in the Metabolic/HTN+Cancer group (95% CI 1.37–1.39).

Figure 3.

Odds ratios and confidence intervals for the association of cluster membership with age, sex, and income from multivariate regression models.

In terms of healthcare utilization, we find that, as expected, clusters with higher transitions have higher primary and emergency healthcare visits and admissions.

[Table 1 Association between cluster membership and healthcare utilization]

Table 1.

Mean number of transitions, primary healthcare (PHC) visits, emergency visits and admissions by cluster, and association between healthcare use and cluster membership

Cluster	Transitions	PHC visits		Emergency visits		admissions
	Mean	Mean	Rate ratio (95% CI)	Mean	Rate ratio (95% CI)	Mean	Rate ratio (95% CI)
Late Metabolic/HTN	1.5	73.9	1.59 (1.57–1.62)	2.8	1.04 (1.02–1.06)	1.6	1.16 (1.14–1.18)
Early Metabolic/HTN	1.5	82.0	1.72 (1.69–1.74)	2.9	1.05 (1.03–1.07)	1.7	1.59 (1.55–1.64)
CVD start	1.8	95.8	1.76 (1.72–1.8)	4.1	1.34 (1.31–1.38)	2.7	1.39 (1.35–1.44)
Neuro start	1.8	89.6	1.59 (1.55–1.64)	3.8	1.23 (1.19–1.27)	2.3	1.43 (1.4–1.47)
Metabolic/HTN+CVD	2.3	121.0	1.7 (1.67–1.73)	4.5	1.14 (1.11–1.17)	3.0	1.18 (1.15–1.22)
Metabolic/HTN+Neuro	2.4	110.0	1.38 (1.35–1.42)	4.2	1.01 (0.98–1.04)	2.5	1.81 (1.76–1.86)
Metabolic/HTN+Cancer	2.2	88.5	1.3 (1.27–1.33)	3.9	1.06 (1.03–1.09)	3.4	4.47 (4.38–4.56)
High mortality	1.61	18.8	2.84 (2.78–2.89)	1.1	1.01 (0.98–1.04)	3.3	1.16 (1.14–1.18)
Total	1.13	52.8	–	2.2	–	1.7	–

In fully adjusted models (controlling for age, sex, income, number of diagnostic transitions, and time at risk), all clusters exhibited significantly higher PHC visit Rate Ratios (RR) compared to the Healthy reference cluster (Table 1). The High Mortality cluster had the greatest increase in PHC visits (RR 2.84, 95% CI 2.78–2.89), followed by the CVD-Start (RR 1.76, 95% CI 1.72–1.80) and Metabolic/HTN+CVD (RR 1.70, 95% CI 1.67–1.73) clusters, when compared to the Healthy cluster. Emergency visit rates were also higher across all clusters when compared to the Healthy cluster, ranging from an RR of 1.01 (95% CI 0.98–1.04; Metabolic/HTN+Neuro) to 1.34 (95% CI 1.31–1.38; CVD-Start), with the CVD-Start and Neuro-Start (RR 1.23, 95% CI 1.19–1.27) clusters showing the highest RR. By contrast, the highest RR for hospital admissions, when compared to the Healthy cluster, was observed among individuals in the Metabolic/HTN+Cancer cluster (RR 4.47, 95% CI 4.38–4.56) and Metabolic/HTN+Neuro cluster (RR 1.81, 95% CI 1.76–1.86).

Discussion

This study provides a longitudinal perspective on multimorbidity, analyzing patterns of disease accumulation, their sociodemographic correlates, and associated healthcare utilization in relation to four major disease groups affecting health and disability globally. We identify substantial heterogeneity in the dynamic process of disease accumulation and multimorbidity, with marked differences by age, sex, and income levels.

Our study highlights the propagative nature of chronic diseases, where having any condition seems to increase the probability of developing additional conditions. This was true across most disease combinations but is most notable for Metabolic/HTN conditions—the most prevalent initial conditions—which often preceded multimorbidity patterns involving the remaining three condition groups. Prior studies [23] similarly document clustering of chronic conditions, particularly cardiovascular and metabolic disorders, due to shared pathophysiological mechanisms. This finding emphasizes the role of metabolic and hypertensive conditions as risk factors for neurodegenerative, cardiovascular, and cancer conditions development and progression [24], offering a clear opportunity for reducing multimorbidity by targeting these initial conditions, which are often easier and less expensive to manage.

The nine clusters identified highlight the diversity of health and disease progression trajectories in the population. Some individuals retain their health for long periods, while others go on to develop one or multiple conditions. This diversity is not random, as aligned with previous studies, and seems to be influenced by observed sociodemographic characteristics of individuals. Younger cohorts were more likely to remain healthy or in single morbidity states. Sex predicted specific trajectories reflecting the known sex distribution of the studied conditions—for instance, higher female membership in clusters involving neurodegenerative conditions, while men were more likely to be in the high mortality cluster. Potential reasons for these sex differences include differences in survival and healthcare-seeking behaviors, as well as biological and social determinants [25, 26]. Unlike other studies, however, women spent a similar time in a healthy or disease-free state as men [27, 28], though this may be an artifact of the diseases included in this study compared to others.

Income also shows a clear gradient, with individuals in higher-income brackets spending more years disease-free and fewer years in states of morbidity or death. This is consistent with evidence that socioeconomic advantage confers better health outcomes [28]. The exception to this is states involving cancer, where the high- or middle-income groups spent more time compared to the lowest groups, a pattern we also see for the Metabolic/HTN+Cancer cluster. This, however, does not necessarily contradict the overarching thesis and may instead reflect higher cancer survival rates among higher earners rather than a true disadvantage [29]. The observed variations by sex and income levels highlight the dynamic interplay between social and biological determinants of multimorbidity in Catalonia, which reproduce the broader socio-economic inequalities seen across Europe [4, 11].

Multimorbidity patterns are difficult to compare between studies due to substantial variation in measurement and methods, yet we can still note some commonalities identified with other studies. Most notably, metabolic and cardiovascular combination patterns are among the common patterns [28]. More specifically, three clusters identified in a sequence analysis of a Scottish sample (cancer starts with a slow transition to multimorbidity, and diabetes starts with a slow transition to multimorbidity and a fast transition to both the diabetes and cardiovascular clusters) map almost perfectly to three of the clusters we identified in our study [30]. These similarities across populations indicate the potential for biological and physiological patterns to determine some of the disease accumulation patterns in older populations.

The heterogeneous progression patterns observed in this study have important implications for healthcare service utilization. In addition to the resource-saving potential of prolonged health, we also find that the type of condition is strongly associated with patterns of service use. Both disease type and disease count contribute to multimorbidity complexity. However, when controlling for the number of transitions, some condition types showed distinct associations with healthcare use, indicating that disease type may explain additional variation beyond disease count alone. PHC visits—the most fundamental form of service utilization—were particularly elevated in the high mortality cluster, suggesting frequent engagement with healthcare systems in the context of deteriorating health. In contrast, emergency visits were highest among individuals in the CVD Start and Metabolic/HTN +CVD clusters, perhaps owing to the acute nature of cardiovascular events. Hospital admissions, the most resource-intensive form of care examined, were disproportionately concentrated among members of the Metabolic/HTN+Cancer cluster, reflecting the intensive treatment pathways often associated with cancer. These findings underscore the need for a more nuanced approach to health service planning that considers the qualitative aspects of multimorbidity.

This study contributes to the growing body of research on multimorbidity by offering a longitudinal lens on disease development and progression, considering the most prevalent condition groups contributing to global morbidity and mortality. The large cohort size, combined with a 17-year follow-up period, addresses a critical gap in the existing literature, which often relies on shorter time frames or cross-sectional designs [30]. Grouping conditions into four distinct but large categories provides a broad yet sufficiently granular perspective on disease accumulation, offering clinically and policy-relevant findings. Moreover, the clustering algorithm applied enabled the classification of a large dataset, thereby overcoming methodological difficulties common in cluster analysis. Finally, by linking these trajectories to healthcare utilization patterns, the study offers actionable results for both clinicians and policymakers.

Nonetheless, this study comes with some limitations. Firstly, our definition of health potentially biases our estimates of health upwards, since we do not include all chronic diseases. Second, the reliance on administrative data limits the inclusion of potentially relevant variables, such as education or lifestyle factors, which may provide valuable insights for health policy. Finally, the clustering approach, while robust, inherently depends on algorithmic choices (e.g. clustering method), which could influence results.

In conclusion, this study demonstrates that disease progression follows identifiable patterns closely linked to sociodemographic characteristics. It underscores the propagative nature of multimorbidity, with metabolic and hypertensive disorders frequently serving as entry points, highlighting critical opportunities for the prevention and delay of multimorbidity through early intervention. Importantly, we show that healthcare utilization patterns are shaped more by the type of condition than by the number of conditions, underscoring the need for a more nuanced approach to health service planning. While the analysis is based on data from Catalonia, Spain, the insights are broadly applicable to other aging populations and support the design of more precise, equitable health strategies aimed at improving outcomes for individuals living with chronic conditions.

Funding

I.P. and A.S.A. acknowledge financial support for this research from the Spanish Ministry of Science and Innovation under the program “Generación de Conocimiento” Project LONGHEALTH (PID2021-128892OB-I00; PI: I.P. [CED-CERCA]. Support from the ERC CoG HEALIN project (864616; PI: I.P.) is gratefully acknowledged. This paper has also received support from the 2021 SGR 00957 project (PI: A.S.A.) and the 2021 SGR 01337 project (PI: I.P.) funded by the Agency for Management of University and Research Grants (AGAUR).

Data availability

Data for this study is available through request from the Agència de Qualitat i Avaluació Sanitàries de Catalunya (AQuAS).

Key points.

• Multimorbidity follows identifiable progression patterns influenced by sociodemographic factors such as age, sex, and income.

• Metabolic and hypertensive conditions are common initial diseases that often precede other chronic conditions, highlighting them as critical targets for early intervention.

• In addition to the number of conditions, disease type also shapes healthcare utilization, indicating a need for differentiated service planning.

• Higher-income individuals experience longer disease-free periods, emphasizing persistent health inequalities across socioeconomic strata.

• Public health policy should prioritize early detection and management of metabolic and hypertensive conditions to delay or prevent multimorbidity and reduce healthcare burden.