Multimorbidity, commonly defined as the coexistence of multiple chronic diseases and/or conditions within one individual, is prevalent among older populations (Salive, 2013). Multimorbidity has been associated with functional limitations (Kadambi et al., 2020), poor quality of life (Kanesarajah et al., 2018), and mortality (Nunes et al., 2016). With a rapidly aging global population, multimorbidity poses a great economic burden on both individuals and health care systems (Larkin et al., 2021; Soley-Bori et al., 2021). Identifying older patients with multimorbidity at high risks of adverse health outcomes in the community may inform clinicians and public health policymakers to prioritize those groups of people and enable early, effective, and targeted interventions to prevent premature death and reduce health costs (Charlson et al., 2014). Although previous research has made considerable efforts to assess multimorbidity (Nicholson et al., 2019), an international consensus regarding the standard measurement of multimorbidity has yet to be reached (Johnston et al., 2019). Therefore, further explorations for tools to measure multimorbidity are needed for patient care, resource allocation, and the prevention of multimorbidity progression and complications (Wei et al., 2016).

Among all multimorbidity assessments, weighted-multimorbidity indices have been widely used (Stirland et al., 2020). The majority of weighted indices for older adults were developed from hospital patients based on in-patient medical records (Diederichs et al., 2011). Moreover, current multimorbidity indices for community-dwelling adults were mostly developed from young or middle-aged populations in western countries (Lee et al., 2006; Wister et al., 2015; Wei et al., 2016). Given disparities in the spectrum of diseases in different regions and populations worldwide, these indices might not be generalizable to older Chinese adults (GBD 2017 Causes of Death Collaborators, 2018; Zhou et al., 2019).

Although current multimorbidity indices are able to distinguish and depict the influences of chronic conditions on mortality, little research considers the interaction of those conditions in the indices. Previous studies showed that disease combinations had discordant effects on mortality, but most of them identified disease combinations through traditional methods (Caughey et al., 2010; Ferrer et al., 2017). Association rule mining (ARM) is a data-driven approach that has been applied previously to discover significant disease combinations (Held et al., 2016; Yao et al., 2020). Therefore, a multimorbidity index considering the effects of important disease combinations derived from ARM might better capture the whole impact of multiple chronic conditions on mortality.

Developing an explicit and validated measurement tool is significant to help assess the health risks based on diagnoses of chronic diseases and classify older adults into different risk groups for targeted clinical treatment and health management. Moreover, identifying important combinations of diseases may facilitate the evidence-based co-treatment and further investigation into underlying mechanisms (Brown and Thorsteinsson, 2020). Therefore, this study aimed to investigate disease combinations significantly associated with 5-year mortality among community-dwelling older Chinese adults aged 65–84 years from the Chinese Longitudinal Healthy Longevity Survey (CLHLS), as well as develop a multimorbidity index incorporating disease combinations to predict 5-year mortality.

Baseline characteristics of all participants are summarized in Table 1. Among 11,853 participants, the median age was 76.0 (IQR: 69.0, 81.0) years and 53.0% were men. Hypertension, depressive symptom, and sensory impairment were the most common conditions, with the prevalence rates of 55.6, 36.0, and 23.7%, respectively.

Table 2 presents the results of restrictive ARM. Among all the participants, a total of 13 disease combinations were identified. The disease combination with the highest lift was hypertension and cognitive impairment. Older patients with hypertension and depressive symptoms, who died during a 5-year follow-up accounted for 6.6% among all participants. The prevalence rates of the co-occurrence of hypertension and depressive symptoms were 20.4% among all participants and 32.6% among participants who died during a 5-year follow-up.

Figure 1 shows the HRs and weights of each condition or disease combination. Cognitive impairment, diabetes, cancer, and bedridden status were the conditions with a weight of 3, while hypertension and arthritis with HR < 1 were excluded for the calculation of both MI and MIDC. Each of the six disease combinations derived from restrictive ARM and with HR > 1 was assigned a weight of 1 or 2. The total ranges of MI and MIDC were 0–23 and 0–31, respectively. A higher score indicates a greater burden of multimorbidity. The distributions of these two indices in the validation set were shown in Supplementary Figure 4. The Cronbach’s α coefficient of MIDC in the validation set was 0.65, indicating acceptable internal consistency reliability. For validity, Pearson’s correlation coefficients (Supplementary Table 4) showed that the MIDC was significantly related to condition count (r = 0.88, p < 0.001) and MI (r = 0.95, p < 0.001).

The results of predictive validity analyses are presented in Table 3. The C-statistics of condition count, MI, and MIDC were significantly higher for mortality prediction than that of the base model with only age and sex (all p < 0.001). Moreover, the reclassification measures of discrimination showed that all models, such as condition count, MI, and MIDC performed better in predicting 5-year mortality than the base model (IDI: more than 0, all p < 0.001; NRI: more than 0, all p < 0.001). Compared with condition count, MIDC showed significantly better discriminative ability (C-statistic: p = 0.016; IDI: 0.005, p < 0.001; NRI: 0.038, p = 0.478). Compared with MI, the C-statistic of the model using MIDC was significantly higher (p = 0.031), while the IDI was more than 0 but not statistically significant (IDI: 0.003, p = 0.090).

To the best of our knowledge, this is one of the first studies to develop and validate the MIDC to predict 5-year mortality in community-dwelling older Chinese adults aged 65–84 years. The MIDC showed greater predictive performance than simple condition count and MI in predicting mortality. This suggests that it is of great importance to consider the effect of specific disease combinations when assessing the impact of coexisting chronic conditions on mortality in older adults.

In our study, we found that 13 disease combinations and 5-year mortality were co-occurring more frequently than expected in older adults. Among these combinations, dyads of hypertension and depressive symptoms, as well as hypertension and sensory impairment, were most prevalent, which was in line with a large body of literature and could be explained by underlying pathophysiological mechanisms (de Moraes Marchiori et al., 2006; Bhargava et al., 2012; Maatouk et al., 2016; Crews et al., 2017; Jin et al., 2019). It has been reported that the co-occurrence between hypertension and depressive symptoms could be because of neurobiological changes caused by vascular diseases and psychosocial stressors due to the diagnosis of hypertension (Maatouk et al., 2016; Jin et al., 2019). Microvascular damage in the retina and cochlea associated with elevated blood pressure, such as vascular stenosis and hemorrhage, has been proposed as a potential mechanism of hearing impairment and vision loss (de Moraes Marchiori et al., 2006; Bhargava et al., 2012). These findings highlighted that somatic diseases associated with comorbid mental health problems and age-related sensory changes in older people required more attention in primary care and clinical settings.

According to the cox regression results, higher mortality risks of the disease pairs, such as hypertension and tooth loss, and cognitive impairment and depressive symptoms, were observed. There has been little research comparing the effect of coexistence of hypertension and tooth loss on mortality to that of other disease pairs. However, a previous study reported that tooth loss increased the risk of hypertension due to insufficient intake of fiber and chronic systemic inflammation caused by periodontal disease (Da et al., 2019). Another cohort study of 7,674 Sweden adults also showed that tooth loss was significantly associated with all-cause mortality and cardiovascular diseases mortality (Holmlund et al., 2010). Furthermore, periodontitis has been shown to significantly increase the magnitude of endothelial dysfunction in patients with hypertension, which may accelerate the progression of carotid atherosclerosis and incidence of stroke, myocardial infarction, and cardiovascular diseases death (Higashi et al., 2008; Desvarieux et al., 2010). These results may imply that older people with both hypertension and tooth loss may be more likely to have cardiovascular diseases and further lead to premature mortality. Moreover, previous research has revealed that older patients with dementia and depression had higher mortality risks than those with most of the other chronic conditions (Schafer et al., 2018; Zheng et al., 2021). Late-life depression could interact with cognitive impairment by sharing common underlying pathogenetic mechanisms related to ischemic brain lesions (Linnemann and Lang, 2020). The co-occurrence of cognitive impairment and depression may indicate severe vascular dysfunction, leading to a high cardiovascular mortality risk (Georgakis et al., 2016). As a result, the combinations of these diseases may provide additional valuable information on predicting mortality in older patients with multimorbidity.

Our study compared the performance of condition count, MI, and MIDC in predicting mortality. As expected, based on C-statistics, we found that multimorbidity, measured by condition count, MI, and MIDC, showed significantly better discriminative ability in predicting mortality than age and sex. Although age has been found to be a strong predictor of mortality in several multimorbidity indices, multimorbidity should be additionally considered when assessing the mortality risks of older adults, as previously validated (Charlson et al., 1994; Lee et al., 2006; Nunes et al., 2016). Furthermore, in accordance with the study conducted in older adults aged ≥65 years in Canada, our finding showed that condition count, which was easy to use and understand in clinical settings, has an acceptable prediction performance for mortality among older adults (Quail et al., 2011).

In this study, MIDC performed better than condition count in predicting 5-year mortality. As numerous studies have noted that the association between different chronic conditions and mortality varies among older adults, considering the types and severity of diseases might capture the heterogeneity of their impacts on health (Wei et al., 2016; Stanley and Sarfati, 2017). In addition, we also found that the C-statistic of MIDC was significantly higher than that of MI. Prior research has indicated that specific combinations of diseases, especially cardiovascular and neuropsychiatric disease patterns, may have a synergistic effect on disability or mortality, which partially supports our findings of better performance of MIDC (Jackson et al., 2015; Zheng et al., 2021). Therefore, measuring multimorbidity through multimorbidity index including the effects of specific disease combinations provides a more nuanced risk classification of older patients with multimorbidity and a qualitative dimension that can be useful in clinical practice and research (Johnston et al., 2019; Stirland et al., 2020).

One of the strengths of this study is that we developed a multimorbidity index for community-dwelling older Chinese adults using the data from a multi-province longitudinal study. This approach can support that our index has good generalizability among this population. Moreover, our study included 13 chronic conditions covering most of the conditions prevalent among older Chinese adults and widely used in previous multimorbidity indices (Diederichs et al., 2011; Wang et al., 2020). In addition, compared to previous multimorbidity indices using only individual diseases, we conducted the restrictive ARM to obtain disease combinations important to predict mortality and considered the joint effects of multiple chronic diseases in our index. However, our study has several limitations. First, since our index focused on community-dwelling older adults, it is likely that our index will not be applicable to a nursing home and other institutional populations. Second, most of the chronic conditions were assessed by self-reported questionnaires rather than clinical records, which might lead to recall bias, especially diabetes. However, previous studies also emphasized the importance of self-reported information on a population level since the accuracy of self-reported diseases may reflect the participants’ health literacy (e.g., the awareness of diseases), which could improve the generalization of our index in the community (Smith et al., 2008; van den Akker et al., 2015). Third, the MMSE is not a sensitive measurement of cognitive impairment, and future studies need to assess cognitive function by better tools such as the Montreal Cognitive Assessment (MoCA) (Nasreddine et al., 2005; Arevalo-Rodriguez et al., 2015). Last, we did not have information on the severity, and treatment of each condition, and their influence on mortality needs to be considered in further research.

Multimorbidity index incorporating disease combinations, followed by multimorbidity index with individual diseases, improved the accuracy of 5-year mortality prediction. This may provide a tool for overall risk stratification, care management, and healthcare resource allocation among community-dwelling older Chinese adults. Moreover, our findings may also imply to researchers that considering significant disease combinations to capture synergistic effects is extremely valuable in predicting mortality among older adults with multimorbidity. Further studies are needed to investigate the association of the MIDC with other health outcomes and validate the MIDC in other populations. In addition, condition count may also be a good choice for measuring multimorbidity for its simplicity and the ease of data ascertainment.

Publicly available datasets were analyzed in this study. This data can be found here: Peking University Open Research Data Platform, https://opendata.pku.edu.cn/dataset.xhtml?persistentId=doi: 10.18170/DVN/WBO7LK.

The studies involving human participants were reviewed and approved by the Duke University Medical Health System’s Institutional Review Board (IRB), the National Bureau of Statistics of China, and the Ethical Committee of Social Science Division of Peking University. The patients/participants provided their written informed consent to participate in this study.

YL, HL, and BX designed the study. YL acquired the data, performed the statistical analyses, and drafted the manuscript. YL, ZH, HL, HX, HS, YC, YH, and BX revised the manuscript. All authors approved the final version of the manuscript.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.