Multimorbidity patterns, sociodemographic characteristics, and mortality: Data science insights from low-resource settings

Abstract

Multimorbidity data typically are analyzed by tallying disease counts, an approach that overlooks nuanced relationships among conditions. We identified clusters of multimorbidity and subpopulations with varying risks and examined their association with all-cause mortality using a data-driven approach. We analyzed 8-year follow-up data of people aged 35 years or older who were part of the CRONICAS Cohort Study, a multisite cohort from Peru. First, we used Partitioning Around Medoids and multidimensional scaling to identify multimorbidity clusters. We then estimated the association between multimorbidity clusters and all-cause mortality. Second, we identified subpopulations using finite mixture modeling. Our analysis revealed three clusters of chronic conditions: respiratory (cluster 1: bronchitis, chronic obstructive pulmonary disease, and asthma); lifestyle, hypertension, depression, and diabetes (cluster 2); and circulatory (cluster 3: heart disease, stroke, and peripheral artery disease). Although only the cluster comprising circulatory diseases showed a significant association with all-cause mortality in the overall population, we identified two latent subpopulations (named I and II) exhibiting differential mortality risks associated with specific multimorbidity clusters. These findings underscore the importance of considering multimorbidity clusters and sociodemographic characteristics in understanding mortality risks. They also highlight the need for tailored interventions to address the unique needs of different subpopulations living with multimorbidity to reduce mortality risks effectively.

Introduction

Multimorbidity poses a significant global challenge, particularly in low- and middle-income countries (LMICs).^1-4 The recent COVID-19 pandemic has shown that people with multimorbidity have a higher risk of severe disease and death due to SARS-CoV-2 infection.^5-7 Thus far, the bulk of literature on multimorbidity has originated from high-income settings, with limited cross-sectional evidence originating from LMICs^2,8 and even fewer longitudinal studies.⁹

Multimorbidity in relation to increased mortality has become a significant public health concern. A meta-analysis by Nunes et al.¹⁰ showed that individuals with three or more chronic conditions have a 2- to 3-fold higher risk of death compared with those without multimorbidity. The burden is especially critical in LMICs, where health systems are often under-resourced. Evidence from a systematic review highlighted that in LMICs, multimorbidity is associated with a 1.5- to 2-fold increase in mortality.¹¹ Yet, health care systems are often designed to treat single diseases rather than manage the complexity of multiple conditions.

Despite growing awareness of multimorbidity and its link to increased mortality, significant gaps remain in understanding how these conditions interact, particularly in LMICs. Recently, Whitty and Watt¹² emphasized that recognizing patterns of diseases that occur together holds significant clinical relevance, particularly in individuals across all ethnicities and socioeconomic levels. However, there is limited research on the specific patterns of multimorbidity in LMIC populations and how socioeconomic, cultural, and environmental factors influence these patterns and their impact on mortality.^13,14

Sociodemographic characteristics are crucial for studying multimorbidity patterns; however, studying them properly is challenging.^15-17 Although multimorbidity is often associated with elderly age groups, there is a growing appreciation that younger adults also experience multimorbidity.^18,19 In LMIC settings, diverse populations reside in rural and urban areas, further challenging the study of multimorbidity.²⁰ Moreover, multimorbidity tends to affect socioeconomically disadvantaged populations disproportionately.^1,21 All these sociodemographic factors do not operate in isolation. Thus, they need to be studied from a multifactorial angle that facilitates the description and detection of those subpopulations at more risk.

Benefiting from an 8-year ascertainment of mortality in an adult cohort spanning various geographical settings of an LMIC, we aimed to study the association between multimorbidity and mortality and how this can vary across subpopulations defined by different sociodemographic features. To accomplish this, we set three objectives: 1) to delineate clusters of physical and mental chronic conditions within the target population, 2) to assess the relationship between these clusters (at baseline) and all-cause 8-year mortality across the entire population, and 3) pinpoint and characterize subpopulations of participants based on their sociodemographic profiles and explore whether the studied association varies across these groups.

Methods

Study design and settings

This follow-up analysis builds upon the CRONICAS Cohort Study, conducted across four distinct Peruvian settings characterized by variations in urbanization, air pollution levels, and altitude. Comprehensive details of this study have been previously published.^20,22 The study sites encompassed highly urbanized Lima, Peru's capital, alongside semi-urban communities in Tumbes, located in northern Peru at sea level. Additionally, high-altitude sites included urban and rural communities in Puno, positioned at an elevation of 3825 m above sea level in the southern Peruvian Andes.

Participants

The study enrolled full-time residents of the area aged 35 years or older. We used a stratified random sampling method, using sex and age categories (35-44, 45-54, 55-64, and ≥65 years). Only one participant per household was included in the study.

Study procedures at baseline

Baseline assessments were conducted from September 2010 to January 2012, with detailed procedures outlined elsewhere.²² Fieldwork personnel and site coordinators underwent comprehensive training covering participant selection, ethical considerations of human participants in research, informed consent procedures, and clinical assessments.

Sociodemographic information, risk factors, and cardiopulmonary symptom history were collected through a questionnaire administered at enrollment.²² In rural areas, fieldworkers proficient in local languages assisted participants with poor literacy in completing surveys. Physical measurements included weight, height, blood pressure, and spirometry before and after bronchodilator administration. Measurement techniques are described in previous publications.²² Blood samples were analyzed at a single laboratory facility.

Follow-up procedures

Mortality status was ascertained in 2018 using Peru's National Registry of Identification and Civil Status, the national authority responsible for vital records. Vital status and date of death or censoring (September 2018 in Lima, Tumbes, and Puno) were the only data used in this study. For deceased participants, the date of death was used for analysis; for those still alive, the date of the database search was considered the censoring date. Due to typical delays in death recording, there may be a 1-3 month lag between the actual date of death and its recording in the registry.

Study variables

The primary outcome of interest was all-cause mortality, defined as the incidence of any fatal event from the baseline enrollment until any point during the follow-up period. The duration between the baseline assessment and the date of death or censorship was measured in years.

At baseline, multimorbidity was defined as the presence of two or more chronic conditions from a predefined list of 12 conditions: alcohol disorder, asthma, chronic bronchitis, chronic obstructive pulmonary disease, depression, heart disease, hypertension, peripheral artery disease, stroke, and type 2 diabetes mellitus, consistent with our previous study.²⁰ Additionally, for this analysis, gastroesophageal reflux (n = 3) and lung cancer (n = 2) were excluded due to the small number of cases. Instead, we incorporated obesity, defined as a body mass index (BMI) ≥30 kg m^–2, and total cholesterol, with high total cholesterol defined as ≥240 mg dL^–1.²³ Although multimorbidity was defined as having two or more chronic conditions, we further quantified the number of diseases within each cluster to create a Multimorbidity Index, as described in “Statistical Methods.”

The analyses considered various sociodemographic characteristics, including sex, age (in years), education level (categorized as low, low-middle, middle, middle-high, and high), wealth index (also classified into five levels: low, low-middle, middle, middle-high, and high), and geographic site (Lima, Urban Puno, Rural Puno, Tumbes). These covariates were all assessed at baseline. More information on geographic sites and wealth index definitions can be found elsewhere.^22,24

Statistical methods

Our analysis strategy used a two-stage approach, summarized in Figure 1. In the first stage, we used two algorithms, Partitioning Around Medoids (PAM) and multidimensional scaling, to identify clusters of multimorbidity. We then developed a multimorbidity index for each cluster to facilitate comparisons in subsequent survival analyses. We then estimated the association between multimorbidity clusters and all-cause mortality in the full population by fitting Weibull survival models, adjusting for specific comorbidities and covariates. In the second stage, we identified subpopulations within the full population using finite mixture modeling^25,26 based on socio-demographic characteristics while considering multimorbidity clusters and mortality risk. For survival models, we assessed the proportional hazard assumption graphically. Missing data on covariates were quantified and addressed using an imputation approach. We conducted formal analyses using Stata 18.0 and R, version 4.3.1. To ensure reproducibility, we include essential details of all models, along with analysis codes, in the Supplementary Information.

Figure 1.

Two-stage approach for clustering diseases, detecting subpopulations based on their sociodemographic characteristics, and analyzing mortality risk related to multimorbidity. FMM, finite mixture modelling; PAM, Partition Around Medoids.

Stage 1: Clustering of diseases, multimorbidity index, and associations with mortality

First, clusters of multimorbidity were identified using two algorithms: PAM^27-29 and multidimensional scaling.³⁰ In this process, all health conditions were dichotomized into yes/no categories, and a distance matrix was computed based on the Jaccard distance, measuring the proximity of each condition. PAM operates by searching for k representative objects or medoids in this matrix. Once a set of k medoids is identified, the algorithm constructs k clusters by assigning each observation to the nearest medoid. The optimal k value is the one that minimizes the sum of dissimilarities between each health condition and its closest representative object. The multidimensional scaling algorithm takes the distance matrix and returns a set of points such that the dissimilarities are approximately equal to the distances between points. The algorithm adjusts the positions of points in a low-dimensional space iteratively to minimize the selected stress function, which quantifies the difference between the original dissimilarities and the distances between points. Combining both approaches, PAM and multidimensional scaling, we drew a plot to visualize the clusters of multimorbidities. This way, we integrated information from both techniques to determine the final clusters. To verify the consistency of the clusters identified, we applied an alternative hierarchical clustering approach to the same data and contrasted its output against that of the PAM and multidimensional scaling. Finally, two medical doctors verified the clinical plausibility of the clusters. Ours contrasts with the traditional approach of merely tallying disease counts, which overlooks nuanced relationships among conditions. Thus, we were able to identify linked comorbidity groups to be analyzed in terms of mortality risk, providing more actionable insights for clinical practice.

Second, a multimorbidity index was calculated for each multimorbidity cluster by first counting the number of conditions present per patient. Next, we subtracted the mean and then divided it by the SD, resulting in a z score with a mean of zero and an SD of 1. This index was created to enable meaningful comparisons in subsequent survival analyses. Because clusters contain varying numbers of diseases, hazard ratios (HRs) for each cluster cannot be directly compared when using simple counts. The standardization provided by the index addresses this issue. It is important to note that multimorbidity index values are directly proportional to the number of diseases within clusters; if a person presents more diseases in the same cluster, the multimorbidity index of that cluster increases accordingly. Additionally, with k clusters, each person has k values of the multimorbidity index, representing one value per cluster.

Third, to explore the relationship between clusters of multimorbidity and death, we fitted Weibull survival models to estimate HRs as a measure of relative risk. Specifically, HR indicates the increase or decrease in mortality risk per unit change in the multimorbidity index. For example, an HR of 1.3 means a 30% extra risk per each 1-unit increase in the multimorbidity index. We fitted one model per cluster, adjusting for specific comorbid conditions within each model. This approach helped us avoid the inclusion of comorbid conditions as confounders when they are in the causal pathway. Consequently, we reported unadjusted HRs for each cluster and adjusted HRs after accounting for covariates.

Stage 2: Subpopulations based on sociodemographic characteristics and mortality risk associated with multimorbidity within each subpopulation

We delineated classes or subgroups of individuals within the target population, here referred to as subpopulations, based on their socio-demographic characteristics (e.g., sex, age, wealth, and site), while considering their multimorbidity cluster profile (i.e., multimorbidity index values for all clusters) and mortality risk. Performing finite mixture modeling (FMM),³¹ we identified these latent subpopulations. Finite mixture modeling integrates two models: a multinomial regression to estimate the probability of belonging to each subpopulation considering socio-demographics and a survival regression to predict time-to-death based on multimorbidity clusters. After FMM analysis, we calculated posterior probabilities to assign each patient to the most likely subpopulation. For each identified subpopulation, we report their socio-demographic attributes, morbidities, and HRs reflecting mortality risk (both per subpopulation and per cluster within each subpopulation). To provide a comprehensive overview of our findings, we generated a plot illustrating individuals categorized by age, subpopulation, and comorbidity cluster.

Research ethics

The CRONICAS Cohort Study received approval from the ethics committees of Universidad Peruana Cayetano Heredia (UPCH) and Asociación Benéfica PRISMA in Peru, as well as from the Johns Hopkins Bloomberg School of Public Health in the United States. Additionally, the ethics committee at UPCH reviewed and approved the mortality follow-up assessments.

Data and code availability

Data sets used for the present analysis could be available from the corresponding author on reasonable request. Analysis code is presented in Supplementary Information.

Results

Data from 3326 participants (51% women; mean age, 55.6 [SD 12.6] years) were included in the analysis. Nearly half of the population fell into the low and low-middle education categories. Urban areas (Lima and urban Puno) were home to 55% of participants, the remaining 45% resided in rural areas (Tumbes and rural Puno). More than half (63%) lived at sea level (Lima and Tumbes), and 37% were in high-altitude settings (Puno). Only 25% of the sample had a normal BMI. The most prevalent conditions at baseline were overweight/obesity (65%), hypertension (25%), and depressive symptoms (18%), followed by alcohol disorders (14%) and diabetes (8%). The prevalence of other conditions studied was not greater than 5%. Missing data were consistently below 18% in each study variable (Table 1).

Table 1.

Sociodemographics and health conditions at baseline of the study cohort (n = 3326).

Sociodemographics and health conditions	No.	%
Sex
Female	1699	51.1
Male	1624	48.8
Missing data	3	0.1
Age, mean (SD), years	55.6	(12.6)
Wealth index
Low	578	17.4
Low-middle	646	19.4
Middle	697	21.0
Middle-high	692	20.8
High	713	21.4
Education level
Low	773	23.2
Low-middle	690	20.8
Middle	1083	32.6
Middle-high	123	3.7
High	654	19.7
Missing data	3	0.1
Site
Lima	1079	32.4
Urban Puno	756	22.7
Rural Puno	473	14.2
Tumbes	1018	30.6
Alcohol disorder
Normal	2848	85.6
Hazardous	477	14.4
Missing	1	0.0
Asthma
No	3250	97.7
Yes	73	2.2
Missing	3	0.1
Chronic bronchitis
No	2762	83.0
Yes	171	5.2
Missing data	393	11.8
Chronic obstructive pulmonary disease
No	2605	78.3
Yes	154	4.6
Missing data	567	17.1
Heart disease diagnosed
No	3200	96.2
Yes	124	3.7
Missing data	2	0.1
Hypertension
No	2204	66.2
Yes	831	25.0
Missing data	291	8.8
Peripheral artery disease
No	2917	87.7
Yes	72	2.2
Missing data	337	10.1
Stroke diagnosed
No	3309	99.5
Yes	15	0.5
Missing data	2	0.1
Diabetes
No	2657	79.9
Yes	262	7.9
Missing data	407	12.2
Body mass index
Normal	832	25.0
Overweight/obesity	2163	65.0
Missing data	331	10.0
Total cholesterol, mg dL–1
<240	2784	83.7
≥240	121	3.6
Missing data	421	12.7
Depressive symptoms
No	2739	82.3
Yes	584	17.6
Missing data	3	0.1

Clusters of multimorbidity and all-cause mortality in the overall population

PAM identified two clusters, named “cluster 1, respiratory” and “cluster 2, lifestyle, hypertension, depression and diabetes.” The remaining three health conditions—peripheral artery disease, stroke, and heart disease—were nonclustered by PAM (Figure 2). However, our multidimensional scaling plot revealed that these conditions form another cluster designated “cluster 3, circulatory.” Medical professionals on the study team validated these clusters, concluding they were clinically plausible. Although the three clusters were partially consistent with those obtained from the alternative hierarchical clustering analysis, some variations were noted (see Supplementary Information). Given that cluster 2 encompasses six health conditions and the other two clusters contain three each, using a multimorbidity index seemed justified to standardize measures and facilitate comparisons of HRs among clusters 1, 2, and 3.

Figure 2.

Multidimensional scaling plot with Partition Around Medoids (PAM) clustering. Points are coded by cluster assignment (cluster 1, cluster 2, and cluster 3; see legend). The x- and y-axes are the first two dimensions obtained from multidimensional scaling. Points that are closer together are more likely to belong to the same cluster. BMI, obesity based on body mass index; CBR, chronic bronchitis; COPD, chronic obstructive pulmonary disease; PAD, peripheral artery disease.

In the overall population (Table 2), unadjusted estimates revealed a longitudinal association between specific clusters and all-cause mortality. Specifically, there was a 37% higher mortality for cluster 1, and a 32% higher all-cause mortality for cluster 3, per each unit increase in the multimorbidity index. However, these estimates were attenuated after further adjustment for covariates, with only cluster 3, showing evidence of an association with all-cause mortality in the overall population (HR = 1.16).

Table 2.

All-cause mortality and multimorbidity clusters (standardized index) by subpopulations,^a which are described in terms of sociodemographics and health conditions (n = 3326).

	Overall		Subpopulation I		Subpopulation II
MMI, b  sociodemographics, health conditions	N = 3326 (100%)		n = 2679 (80.5%)		n = 647 (19.5%)
	No.	%	No.	%	No.	%
All-cause mortality
No	3127	94.0	2552	95.3	575	88.9
Yes	199	6.0	127	4.7	72	11.1
Subpopulations, hazard ratio (95% CI)c			Reference		2.6 (2.0-3.5)
MMI (standarized): hazard ratio (95% CI)d
Cluster 1 (respiratory)	1.37	(1.24-1.53)	1.33	(1.16-1.51)	1.49	(1.25-1.77)
Cluster 2 (lifestyle, depression, and diabetes)	1.02	(0.87-1.19)	1.18	(0.98-1.41)	0.86	(0.65-1.11)
Cluster 3 (circulatory)	1.32	(1.21-1.44)	1.40	(1.27-1.54)	1.26	(1.02-1.55)
MMI (standarized): adjusted hazard ratio (95% CI)e
Cluster 1 (respiratory)	1.11	(0.98-1.25)	1.06	(0.90-1.23)	1.31	(1.07-1.60)
Cluster 2 (lifestyle, depression, and diabetes)	1.05	(0.89-1.23)	1.18	(0.97-1.43)	0.91	(0.68-1.23)
Cluster 3 (circulatory)	1.16	(1.04-1.29)	1.19	(1.07-1.33)	1.18	(0.90-1.55)
	%	(95% CI)	%	(95% CI)	%	(95% CI)
Sex
Female	51.1	(49.4-52.8)	52.9	(51.0-54.8)	44.0	(40.2-47.8)
Male	48.9	(47.2-50.6)	47.1	(45.2-49.0)	56.0	(52.2-59.8)
Age, mean (95% CI), years	55.6	(55.2-56.0)	55.2	(54.8-55.7)	57.1	(56.2-58.1)
Wealth index
Low	17.4	(16.1-18.7)	12.4	(11.1-13.6)	38.0	(34.3-41.8)
Low-middle	19.4	(18.1-20.8)	18.7	(17.2-20.2)	22.4	(19.2-25.6)
Middle	21.0	(19.6-22.3)	24.0	(22.4-25.6)	8.3	(6.2-10.5)
Middle-high	20.8	(19.4-22.2)	24.8	(23.2-26.5)	4.2	(2.6-5.7)
High	21.4	(20.0-22.8)	20.1	(18.6-21.6)	27.0	(23.6-30.5)
Education level
Low	23.2	(21.8-24.7)	22.4	(20.8-23.9)	26.7	(23.3-30.2)
Low-middle	20.8	(19.4-22.1)	20.3	(18.8-21.8)	22.7	(19.5-26.0)
Middle	32.6	(31.0-34.2)	34.3	(32.5-36.1)	25.5	(22.1-28.9)
Middle-high	3.7	(3.1-4.3)	4.1	(3.4-4.9)	2.0	(0.9-3.1)
High	19.7	(18.3-21.0)	18.9	(17.4-20.4)	23.0	(19.8-26.3)
Site
Lima	32.4	(30.8-34.0)	39.7	(37.8-41.5)	2.5	(1.3-3.7)
Urban Puno	22.7	(21.3-24.2)	21.9	(20.3-23.5)	26.1	(22.7-29.5)
Rural Puno	14.2	(13.0-15.4)	0.6	(0.3-0.8)	70.6	(67.1-74.1)
Tumbes	30.6	(29.0-32.2)	37.8	(35.9-39.6)	0.8	(0.1-1.4)
Alcohol disorder
Normal	85.7	(84.5-86.8)	86.9	(85.6-88.2)	80.5	(77.5-83.6)
Hazardous	14.3	(13.2-15.5)	13.1	(11.8-14.4)	19.5	(16.4-22.5)
Asthma
No	97.8	(97.3-98.3)	97.3	(96.7-97.9)	99.7	(0.99-100.0)
Yes	2.2	(1.7-2.7)	2.7	(2.0-3.3)	0.3	(0.1-0.7)
Chronic bronchitis
No	93.9	(93.0-94.8)	94.2	(93.3-95.2)	92.5	(90.1-94.9)
Yes	6.1	(5.2-7.0)	5.8	(4.8-6.7)	7.5	(5.1-9.9)
Chronic obstructive pulmonary disorder
No	93.7	(92.7-94.7)	94.4	(93.4-95.4)	90.9	(87.9-93.8)
Yes	6.3	(5.3-7.3)	5.6	(4.6-6.6)	9.1	(6.2-12.1)
Heart disease diagnosed
No	96.3	(95.6-96.9)	96.1	(95.3-96.8)	97.1	(95.8-98.4)
Yes	3.7	(3.1-4.4)	3.9	(3.2-4.7)	2.9	(1.6-4.2)
Hypertension
No	73.0	(71.5-74.6)	71.5	(69.8-73.3)	79.1	(75.8-82.4)
Yes	27.0	(25.4-28.5)	28.5	(26.7-30.2)	20.9	(17.6-24.2)
Peripheral artery disease
No	97.6	(97.0-98.2)	97.3	(96.7-97.9)	98.7	(97.7-99.7)
Yes	2.4	(1.8-3.0)	2.7	(2.0-3.3)	1.3	(0.3-2.3)
Stroke diagnosed
No	99.5	(99.3-99.8)	99.5	(99.2-99.8)	99.8	(99.5-100.0)
Yes	0.5	(0.2-0.7)	0.5	(0.2-0.8)	0.2	(0-0.5)
Diabetes
No	91.0	(89.9-92.0)	90.1	(88.9-91.3)	94.6	(92.5-96.6)
Yes	9.0	(8.0-10.1)	9.9	(8.7-11.1)	5.4	(3.4-7.5)
Body mass index
Normal	28.2	(26.5-29.9)	24.1	(22.3-25.9)	45.5	(41.4-49.6)
Overweight/obese	71.8	(70.1-73.4)	75.9	(74.1-77.7)	54.5	(50.4-58.6)
Total cholesterol, mg dL–1
<240	95.7	(94.8-96.4)	95.4	(94.5-96.3)	96.7	(95.1-98.3)
≥240	4.3	(3.6-5.1)	4.6	(3.7-5.5)	3.3	(1.7-4.9)
Depressive symptoms
No	82.4	(81.1-83.7)	84.3	(82.9-85.7)	74.8	(71.4-78.1)
Yes	17.6	(16.3-18.9)	15.7	(14.3-17.1)	25.2	(21.9-28.6)

Abbreviation: MMI, multimorbidity index.

^aSubpopulations were detected by finite mixture modeling. Modeling details can be seen in “Methods.”

^bFor each multimorbidity cluster, we counted the number of conditions present and then calculated a z score (mean = 0, SD = 1). We called this score the Multimorbidity Index (MMI).

^cHazard ratio estimate from Weibull survival models (reference group is Subpopulation I; the outcome is time to death). The hazard ratios for subpopulations are unadjusted.

^dHazard ratio estimates from Weibull survival models (1 model per cluster; outcome is time to death).

^eHazard ratio adjusted for sex, age, education, wealth index, site, and all conditions from other clusters except body mass index and cholesterol.

Subpopulations and all-cause mortality

Using FMM, we uncovered two latent groups within the overall population: subpopulation I (n = 2679; 80.5%) and subpopulation II (n = 647; 19.5%) (Table 2). Individuals in subpopulation II had a significantly higher risk of death, being 2.6 times more likely to die than those in subpopulation I.

When investigating the relationship between multimorbidity clusters and all-cause mortality in both subpopulations, cluster 2, lifestyle, hypertension, depression and diabetes, showed no association with all-cause mortality. Unadjusted HRs, per each 1-unit increase in the multimorbidity index, were consistent in magnitude across cluster 1 (respiratory) and cluster 3 (circulatory). However, after adjusting for covariates, cluster 1 remained associated with mortality in subpopulation II (HR = 1.31); cluster 3 remained associated with mortality in subpopulation I (HR = 1.19).

In simpler terms, our analysis revealed that respiratory multimorbidity poses a higher risk of all-cause mortality among individuals from subpopulation II. In contrast, cardiovascular multimorbidity presents a greater risk among those from subpopulation I. Interestingly, this distinction was not apparent when analyzing the entire population.

Characteristics of subpopulations

The profile of individuals in subpopulation II predominantly consisted of men, with a higher proportion falling into the lowest categories of wealth and residing primarily in highland areas, mainly rural Puno, followed by urban Puno. In contrast, individuals in subpopulation I were predominantly women, slightly younger (by approximately 2 years), and predominantly resided in sea-level areas such as Lima and Tumbes. The morbidity profile of both subpopulations is detailed in Table 2.

Figure 3 illustrates that individuals from subpopulation II have fewer morbidities within cluster 3 (circulatory) than those in subpopulation I. Moreover, individuals from subpopulation II tend to have more simultaneous morbidities within cluster 2 (lifestyle, hypertension, depression, diabetes) before the age of 70 years, whereas those in subpopulation I may experience a similar number of comorbid conditions before and after that age. Additionally, any subpopulation may present with more than one morbidity from cluster 1 (respiratory) between the ages of 40 and 80 years.

Figure 3.

Multimorbidity clusters (1, 2, and 3) in people of two subpopulations (subpopulation I and subpopulation II, n = 3326) . Note: The larger the symbol, the higher the number of comorbid conditions within that specific cluster. Circles indicate cluster 1; V-shaped symbols indicate cluster 2; triangles indicate cluster 3. Subpopulations are labelled I and II in the legend. Plots show 20% of the sample for cluster 2 only (randomly selected to facilitate visualization).

Discussion

Main findings

We used mortality data from a cohort study spanning various socioeconomic and environmental settings to identify clusters of multimorbidity and subpopulations with varying risks and examined their association with all-cause mortality. Our analysis revealed three distinct clusters of chronic conditions within the target population: respiratory (cluster 1), lifestyle, hypertension, depression, and diabetes (cluster 2), and circulatory (cluster 3). Interestingly, only the cluster comprising circulatory diseases had a significant association with all-cause mortality in the overall population.

Additionally, our investigation identified two latent subpopulations within the target population (based on socioeconomic factors), each exhibiting different mortality risks and, more interestingly, differential risks associated with specific multimorbidity clusters. Individuals in subpopulation II, who mostly were men and of older age, lower wealth, and highland residence (mostly rural), faced 1) more than twice the mortality risk of subpopulation I and 2) a higher mortality risk linked to respiratory multimorbidity (cluster 1). Conversely, individuals in subpopulation I, primarily women, younger, wealthier, and residing at sea level, faced a higher mortality risk associated with circulatory multimorbidity (cluster 3).

Clustering in the study of multimorbidity

Chronic conditions manifest in clusters rather than randomly arising, often sharing common risk factors.^31,32 However, there is currently no standardized method for identifying and analyzing these cluster patterns, with factor analysis and hierarchical clustering being two commonly employed techniques.^32,33 Replicating multimorbidity clusters using different approaches may prove challenging, highlighting the importance of appropriate measures and analysis to understand multimorbidity patterns and inform adequate responses, including prevention strategies.^12,34-36

However, caution is needed when interpreting clustering outcomes, because these outcomes are not only data-driven³⁷ and are also influenced by methodological decisions and researcher input. Our analysis, as shown in the case of the identification of cluster 3 (circulatory), demonstrates the significance of such considerations. We used multiple clustering methods and ensured clinical relevance through team review, underscoring the meticulous approach adopted. The association among multimorbidity clusters, death, and, particularly, sociodemographic characteristics guided the subsequent detection of subpopulations.

Although our analytical approach may diverge from others due to methodological nuances, it provides valuable insights and contributes to the ongoing discourse on establishing standardized methods for analyzing multimorbidity patterns.³⁷ We recommend considering these methodological nuances when comparing our study with literature using similar approaches, such as unsupervised machine learning. Awareness of these nuances will enhance the interpretation of our study findings within the broader context of multimorbidity research.

Comparison with the literature

The existing literature highlights a spectrum of methodologies used to evaluate multimorbidity,³⁸ ranging from simple disease counts^39-41 to more nuanced approaches considering mortality risk⁴² and affected body systems.⁴³ Although disease count remains prevalent, its limitations in capturing the complexity of disease combinations are increasingly recognized. In contrast, clustering methods offer a more nuanced understanding by grouping conditions based on shared characteristics. Notably, our data-driven analysis revealed clusters aligned with affected body systems, such as respiratory and cardiovascular, enhancing clinical interpretation compared with generic disease counting.

Our study takes a longitudinal approach by using an 8-year follow-up of a robust outcome, such as death, to examine multimorbidity patterns, thus departing from the more common cross-sectional studies prevalent in the literature. This approach provides a deeper understanding of how chronic conditions interact within populations over time. Importantly, our focus on LMICs addresses a notable gap in the literature, because these settings are often underrepresented in longitudinal multimorbidity studies.⁴⁴ By including a range of LMIC-specific settings in our analysis, we aimed to capture the unique challenges and complexities faced by populations in these regions, thereby enriching our understanding of multimorbidity on a global scale.

Studies across different populations and using varied analytical techniques identify clusters predominantly comprising cardiovascular and respiratory conditions. For instance, studies in the United Kingdom,^45,46 China,⁴⁷ and Chile,^36,47 have consistently reported clusters featuring cardiovascular and respiratory diseases, with diabetes and hypertension often clustered together.^37,48 Despite differences in population demographics and analysis methods, the consistency of these clusters underscores their relevance across diverse settings. Notably, our approach differed from traditional methods by first identifying disease clusters and subsequently characterizing subgroups of individuals based on sociodemographic factors. This innovative approach allowed for a more nuanced understanding of multimorbidity dynamics within the target population, revealing distinct subgroups with differential mortality risks.

Our findings for subpopulation II align with evidence showing that rural and socioeconomically disadvantaged groups in LMICs face elevated health risks due to limited health care access and a higher prevalence of chronic diseases.^49,50 Respiratory multimorbidity (cluster 1) emerged as a major contributor to death in this group, emphasizing the need for interventions targeting respiratory conditions in rural and high-altitude areas where poor indoor air quality and limited health care infrastructure may worsen these issues.⁵¹ In contrast, the risk profile of subpopulation I likely reflects the effects of urbanization and lifestyle factors like dietary patterns and physical inactivity, both linked to cardiovascular disease in more affluent LMIC settings.^52,53 These findings highlight the need for context-specific multimorbidity management that addresses sociodemographic and environmental influences on health outcomes in varied populations.

Relevance for public health

As our study revealed, multimorbidity significantly affects mortality rates and health care utilization.^1,54 Individuals with multimorbidity face higher risks of premature death and more frequent hospital admissions than individuals with single chronic conditions,^55,56 which aligns with our findings of distinct clusters of chronic conditions associated with varying mortality risks. Moreover, managing multimorbidity often involves complex treatment regimens and increased health care costs,^57,58 which burden individuals and health care systems.^59-61 Our identification of specific clusters of chronic conditions, such as respiratory and circulatory diseases, underscores the importance of tailored priority settings to address the unique needs of different subpopulations, as highlighted in our study.

The economic burden of multimorbidity extends beyond health care costs, affecting households through increased caregiving responsibilities and potential loss of income.^34,62-64 These challenges are particularly pronounced in LMICs, where social determinants of health exacerbate the impact of multimorbidity. In light of these findings, accurately identifying multimorbidity clusters is crucial for optimizing health care delivery and resource allocation. By understanding how clusters form within populations, health care providers can develop targeted interventions to improve outcomes such as mortality rates, hospitalizations, and overall health care costs, ultimately enhancing public health efforts in managing multimorbidity.

Strengths and limitations

Our study leveraged longitudinal data from diverse socioeconomic and environmental settings, highlighting the broad spectrum of populations represented and underscoring the robustness of our findings, which reflect the multifaceted nature of multimorbidity across different contexts. Additionally, our emphasis on death as a primary outcome provides a robust measure of disease burden and underscores the severity of multimorbidity's impact on health outcomes. Focusing on death ensured a comprehensive assessment of the risks associated with different multimorbidity clusters, contributing valuable insights for health care decision-making and resource allocation. Also, our two-stage analytical approach involving health conditions and individuals yielded valuable insights. Moreover, identifying subpopulations based on sociodemographic characteristics further enriches our understanding of how multimorbidity patterns manifest across different groups. Ultimately, our study provides valuable insights into these patterns and their association with death, and it also contributes to advancing methodological approaches in this field, thus facilitating more comprehensive analyses in future research endeavors.

However, it is essential to acknowledge several limitations in our study. First, the inclusion of self-reported conditions, such as depression assessed through validated questionnaires, introduces a degree of uncertainty regarding the accuracy of prevalence estimates. Second, our analysis was constrained by the availability of only all-cause mortality data, precluding a detailed examination of specific causes of death associated with different multimorbidity clusters. Although our findings provide valuable insights into overall mortality risks, the lack of granularity limits our ability to discern the relative contributions of various chronic conditions to mortality outcomes. Additionally, our multimorbidity definition encompassed a limited number of diseases comprising 12 prevalent chronic conditions. However, this approach may only partially capture the evolving landscape of chronic diseases, particularly those experiencing increasing prevalence rates, such as non–alcoholic fatty liver disease and certain neoplasms, warranting consideration in investigations seeking to provide a more holistic assessment of multimorbidity patterns. Our population was relatively young at baseline, which posed challenges, such as a low incidence of death. However, as we initially noted, there is growing recognition that multimorbidity also affects younger adults, making it essential to generate this new evidence for that population. Last, although our study was conducted within a single country, namely Peru, it is essential to acknowledge that Peru presents a unique testing ground for examining multimorbidity dynamics. The country's diverse populations, languages, settings, and socioeconomic differentials contribute to a rich tapestry of health determinants that offset the limitation of a single-country focus. Indeed, the participants in this study reflect populations residing across variations in urbanization, air pollution levels, and altitude.

In conclusion, our study elucidated three distinct clusters of chronic physical and mental conditions, with only the cluster of circulatory diseases demonstrating a significant association with all-cause mortality in the entire population. We used a two-stage data-driven analytical approach to identify two latent subpopulations with different sociodemographic profiles, each exhibiting differential risks associated with specific multimorbidity clusters. Specifically, individuals in subpopulation II, characterized by older age, lower wealth, and residence in high-altitude settings, faced a higher mortality risk linked to respiratory multimorbidity. Conversely, individuals in subpopulation I, primarily younger, wealthier, and residing at sea level, faced a higher mortality risk associated with circulatory multimorbidity.

Acknowledgments

The authors thank everyone who took part in the study. The views expressed in this publication are those of the author(s) and not necessarily those of the National Institute for Health and Care Research or the Department of Health and Social Care. J.J.M. acknowledges having received support from the Academy of Medical Sciences (NGR2\1210), Alliance for Health Policy and Systems Research (2009/32034, 2012/253750), Bloomberg Philanthropies (grant 46129, via University of North Carolina at Chapel Hill School of Public Health), FONDECYT via CIENCIACTIVA/CONCYTEC, British Council, British Embassy and the Newton-Paulet Fund (223-2018, 224-2018), DFID/MRC/Wellcome Global Health Trials (MR/M007405/1), Fogarty International Center (R21TW009982, D71TW010877, R21TW011740, K01TW011478), Grand Challenges Canada (GMH-POC-0335-04), International Development Research Center Canada (IDRC 106887, 108167), Inter-American Institute for Global Change Research (IAI CRN3036), National Cancer Institute (NCI 1P20CA217231), National Council for Scientific and Technological Development (CNPq Brasil 408523/2023-9), National Health and Medical Research Council (NHMRC 2022036, 2022566, 2044237, 2044850), National Heart, Lung and Blood Institute (NHLBI HHSN268200900033C, 5U01HL114180, 1UM1HL134590), National Institute on Aging (NIA R01AG057531), National Institute for Health and Care Research (NIHR150261, NIHR150287, NIHR303125, NIHR306208), National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK K23DK135798), National Institute of Mental Health (NIMH 1U19MH098780), NSW Health (H23/37663, DG23/7050), Medical Research Future Fund (MRRF 204223), Swiss National Science Foundation (40P740-160366), UKRI BBSRC (BB/T009004/1), UKRI EPSRC (EP/V043102/1), UKRI MRC (MR/P008984/1, MR/P024408/1, MR/P02386X/1, MR/X004163/1, MR/X020851/1), Wellcome (074833/Z/04/Z, 093541/Z/10/Z, 103994/Z/14/Z, 107435/Z/15/Z, 205177/Z/16/Z, 214185/Z/18/Z, 218743/Z/19/Z), World Diabetes Foundation (WDF15-1224) and the World Health Organization (2021/1189041, 2022/1249357).

Supplementary material

Supplementary material is available at American Journal of Epidemiology online.

Funding

The establishment of the CRONICAS Cohort Study was supported by the National Heart, Lung, and Blood Institute Global Health Initiative under the contract Global Health Activities in Developing Countries to Combat Non-Communicable Chronic Diseases (project no. 268200900033C-1-0-1). J.C.B.-A. was supported by the National Institute for Health and Care Research (NIHR) and the Three Schools Mental Health Programme (grant MHF012). Additional sources of support over time are listed in the acknowledgement section.

Conflict of interest

The authors declare no conflict of interest.

Data availability

All data used for this study are available from the corresponding author upon reasonable request.