This retrospective observational study was conducted using inpatient electronic health records (EHRs) collected from 6 local hospitals across 3 counties in Henan Province, China, covering the period from January 1, 2019, to December 31, 2023. The dataset comprised 300,241 inpatient visits, including inpatient identifiers, admission dates, discharge diagnoses, and ICD-10 (International Classification of Diseases, 10th Revision) codes.

Hypertension was treated as the index condition, given its recognized role as a central initiating disease in multimorbidity trajectories [1,12]. The present analysis was based on inpatient EHR data because these records provided patient-linked diagnostic information across repeated hospital admissions, allowing the temporal ordering of recorded chronic conditions to be reconstructed. Hypertension was used as a cohort-defining recorded diagnosis and was not necessarily the primary reason for hospitalization. Accordingly, the study was designed to characterize temporal multimorbidity patterns among hospitalized patients with a recorded hypertension diagnosis rather than to represent all individuals with hypertension in the underlying county populations. The study design followed a population-level, longitudinal disease trajectory framework, consistent with established approaches in multimorbidity trajectory research and temporal disease network construction [3-5]. This ensures that the evolution of comorbidity patterns analyzed thereafter accurately reflects the disease progression in hypertensive inpatients. Because this study used the complete set of eligible inpatient EHR records available from the 6 participating hospitals rather than data obtained through a probability-based survey design, survey sampling weights were not applicable. All analyses were therefore conducted using the unweighted analytical cohort. This study was reported in accordance with the STROBE (Strengthening the Reporting of Observational Studies in Epidemiology; Checklist 1) statement for cohort studies and the RECORD (Reporting of Studies Conducted Using Observational Routinely Collected Health Data; Checklist 2) statement.

This study was approved by the Ethics Committee of Research and New Technology/New Business of Fuwai Central China Cardiovascular Hospital (approval number 2025; ethics review number 25). Owing to the retrospective study design and the use of routinely collected inpatient EHR data, the requirement for individual informed consent was waived by the ethics committee.

The source data received by the research team contained direct inpatient identifiers, including names and ID numbers. Before the construction of the analytical dataset, these direct identifiers were removed and replaced with encrypted inpatient identifiers, which were used solely to link multiple inpatient admissions belonging to the same inpatient. All statistical analyses were conducted using the resulting deidentified analytical dataset, and no directly identifying information was included in the reported results.

Because this study involved the secondary analysis of existing inpatient records and no inpatients were directly recruited or contacted, participant compensation was not applicable.

Raw EHR data underwent standardized preprocessing procedures to ensure data integrity and analytical reproducibility. Specifically, this procedure included inpatient identity harmonization, diagnostic code standardization, temporal normalization, and record exclusion. Encrypted inpatient identifiers were harmonized across records to enable the longitudinal linkage of multiple admissions belonging to the same inpatient. All diagnoses were mapped to the National Clinical ICD-10 version 2.0 standard, following recommended practices for multimorbidity research [2,3]. Admission dates were converted to a unified format (“yyyy-mm-dd”) to enable precise temporal ordering. A total of 4300 inpatient records were excluded due to missing critical fields (patient ID, diagnosis code, or admission date), consistent with data quality control strategies in large-scale EHR studies [4]. No imputation was applied to missing diagnostic or temporal variables, as incomplete records could compromise disease sequencing accuracy.

Cohort selection was carried out according to predefined inclusion and exclusion criteria. To ensure that the comorbidity patterns analyzed would accurately reflect the disease progression of hypertensive inpatients, the inclusion criteria were (1) inpatients with at least 2 inpatient records during the study period, (2) each inpatient record must include at least 2 diagnoses, (3) at least one diagnosis classified as a CVD, and (4) at least one admission containing a hypertension diagnosis (ICD-10: I10.x). The exclusion criteria were (1) inpatients with only one inpatient record; (2) inpatients without any hypertension diagnosis; (3) records with missing key identifiers or temporal information such as inpatient ID, diagnosis, or date; and (4) duplicate inpatient records. After applying the predefined inclusion and exclusion criteria, 34,099 inpatients with at least one recorded hypertension diagnosis during the study period were included in the final analytical cohort.

To characterize coexisting disease patterns at cohort entry, chronic conditions recorded during each inpatient’s first admission in the study dataset were aggregated into inpatient-level transactions. The number of distinct chronic conditions was calculated after deduplicating diagnosis codes within the cohort-entry admission, with hypertension included in the condition count. Frequent disease combinations were identified using the Apriori algorithm with a minimum support threshold of 0.01. Itemset length was not restricted during the mining process; for readability, only the top-ranked 2-, 3-, and 4-itemsets are presented in Table 1.

To characterize statistically robust disease progression patterns, we constructed a temporal comorbidity network (TCN)—a weighted, directed network representing population-level disease evolution, following established disease trajectory modeling frameworks [5,6,12,13]. In the TCN, nodes represent distinct ICD-10 disease categories, and directed edges represent statistically significant temporal transitions between diseases. Across the cohort, 24,959 unique, temporally ordered disease diagnosis pairs were identified and used as candidate temporal comorbidity pairs for statistical inference.

For each inpatient, chronic disease records were chronologically ordered according to admission dates. Only the earliest recorded occurrence of each disease within the study dataset was retained, and subsequent records of the same disease were excluded to reduce repeated-counting bias associated with recurrent hospitalizations. The date assigned to each disease, therefore, represents its first recorded occurrence in the available inpatient EHR data rather than the true date of disease onset or initial clinical diagnosis.

The retained disease records were then grouped by inpatient admission to generate an inpatient-specific longitudinal sequence of effective admission time points. Conditions first recorded during the same admission were assigned to the same time point, and no temporal direction was inferred between them. For each pair of consecutive effective admissions, all eligible disease pairs were generated by linking a disease first recorded at admission t to a different disease first recorded at the immediately subsequent admission t+1. Thus, each inpatient contributed a sequence of temporally ordered disease pairs derived from successive inpatient admissions.

As a hypothetical example, consider an inpatient with hypertension (I10.X) and atherosclerotic heart disease (I25.1) first recorded during admission 1, type 2 diabetes (E11.9) first recorded during admission 2, and cerebral infarction (I63.9) first recorded during admission 3. Because hypertension and atherosclerotic heart disease were first recorded during the same admission, no temporal direction was assigned between them. The consecutive-admission procedure generated the pairs I10.X→E11.9 and I25.1→E11.9 between admissions 1 and 2, and E11.9→I63.9 between admissions 2 and 3. Repeated recordings of previously observed conditions during later admissions were ignored. This hypothetical example is provided solely to illustrate the sequence-generation procedure and does not represent an actual patient record.

Patient-level disease pairs were subsequently aggregated across the cohort. A chi-square test was first used to determine whether the observed co-occurrence frequency of each disease pair exceeded that expected under independence. A significance level of α=.05 was used, and pairs with P<.05 were retained. For each retained pair, a binomial test was then used to determine whether one temporal direction occurred more frequently than the reverse direction under a null probability of 0.5. Pairs with P<.05 in both tests were represented as directed edges in the population-level TCN, with edge weights corresponding to the observed frequency of the temporal transition among inpatients.

To evaluate the stability and reliability of the inferred TCN, 3 complementary robustness analyses were conducted, following best practices in network epidemiology [3,5]. First, bootstrap resampling with replacement was performed 1000 times. The TCN was reconstructed in each bootstrap sample, and the proportion of bootstrap networks in which each edge reappeared was calculated. Edges reproduced in at least 95% of the bootstrap samples were retained to form the Stable Network, representing the most reproducible core subnetwork of the original TCN.

Second, a case-dropping sensitivity analysis was conducted to assess the robustness of the network to changes in cohort composition. In each of the iterations, 3410 of 34,099 (10.0%) inpatients were randomly excluded, and the TCN was reconstructed. Similarity between the edge sets of each reduced network and the original TCN was quantified using the Jaccard similarity coefficient.

Third, a permutation-based null-network analysis was performed to determine whether the observed network topology exceeded that expected by chance. The temporal order of recorded disease diagnoses was randomly permuted for each inpatient, and the TCN was reconstructed using the same analytical procedure for each of the 1000 permutations. The number of nodes and edges, network density, and average clustering coefficient of the original TCN were compared with the corresponding null distributions.

For clarity, the original TCN refers to the full network constructed from all statistically significant temporal disease transitions identified in the observed cohort. The Stable Network is a bootstrap-derived subnetwork of the original TCN comprising edges reproduced in at least 95% of the bootstrap samples. The permutation-based null model refers to the distribution of randomized networks generated by disrupting the observed temporal ordering of diagnoses and serves as a chance-expectation benchmark for the original TCN.

TNA represents a substantive methodological advancement over traditional cross-sectional multimorbidity analyses by explicitly modeling directionality and the temporal ordering of diagnoses recorded across successive inpatient admissions. Most conventional multimorbidity studies have focused on static disease clusters or coprevalence matrices, which obscure the sequential nature of chronic disease accumulation [3,4]. Consistent with population-wide disease trajectory studies using Danish registry data, which demonstrated that disease pairs exhibit nonrandom and directional progression patterns over time [12], our TNA results reveal statistically significant and reproducible transition pathways rather than mere co-occurrence. In contrast to earlier hospital-based studies that relied on undirected comorbidity networks [13], our directed framework captures how diseases are temporally ordered in the available inpatient records, rather than only which conditions coexist. Notably, hypertension occupied a central upstream position in the network, with strong directed edges toward atherosclerotic heart disease and cerebral infarction. This finding corroborates large-scale temporal trajectory analyses, showing hypertension as a common upstream condition preceding ischemic heart disease and stroke [6,12]. By explicitly modeling directionality, our study extends prior work by identifying a densely interconnected temporal pattern involving hypertension, atherosclerotic heart disease, and cerebral infarction, which would remain indistinguishable in undirected network models. These directed associations reflect the temporal ordering of diagnoses in the available inpatient data and should not be interpreted as proof of causal disease progression.

Our subgroup analyses revealed pronounced sex-specific progression pathways. Male inpatients showed a more direct transition from hypertension to cerebral infarction, whereas female inpatients were more likely to progress first to atherosclerotic heart disease before subsequent cerebrovascular events. This pattern is consistent with epidemiological evidence, indicating that women experience adverse cardiovascular outcomes at lower blood pressure thresholds than men [14] and that hypertension-related vascular remodeling differs by sex [15]. Furthermore, estrogen-related mechanisms provide biological plausibility for our observed pathways. Sabbatini and Kararigas [16] demonstrated that postmenopausal estrogen decline exacerbates endothelial dysfunction, inflammation, and oxidative stress, accelerating coronary and vascular damage in women with hypertension. Our network-level finding of earlier coronary involvement among women empirically supports these mechanistic insights at the population level. In contrast to studies that report higher overall hypertension prevalence among men in early adulthood [14], our findings suggest that progression severity and pathway structure, rather than prevalence alone, differ substantially by sex—an aspect often overlooked in prior cross-sectional analyses.

Age-stratified analyses showed broader and more multisystem recorded pathways in the older age groups. Inpatients aged 40 to 49 years exhibited relatively constrained cardiovascular trajectories, whereas those aged 50 to 70 and particularly ≥70 years developed complex, multisystem disease networks involving metabolic, respiratory, and neurocognitive conditions. This observation is consistent with longitudinal evidence from the English Longitudinal Study of Aging, which showed that multimorbidity burden increases not only in prevalence but also in systemic complexity with advancing age [17]. Importantly, Bendayan et al [17] reported that specific multimorbidity patterns, including hypertension combined with diabetes and CVD, were associated with accelerated functional decline in older adults. Our findings extend this work by describing the temporal ordering of diagnoses within these patterns, rather than treating them as static states. Moreover, the prominence of diabetes as a transitional node from midlife onward aligns with MorbiNet findings, where type 2 diabetes was shown to connect CVD to multiple downstream complications despite moderate overall centrality [13]. This supports the interpretation that diabetes functions as a bridge disease within multimorbidity trajectories.

Our comparison across health insurance types revealed that the network for inpatients covered by Urban Resident Basic Medical Insurance contained a longer and more diverse dominant recorded pathway than that for inpatients covered by Urban Employee Basic Medical Insurance. The less extensive downstream patterns observed among inpatients covered by Urban Employee Basic Medical Insurance may reflect differences in population characteristics, continuity of care, preventive service use, or diagnostic opportunities across insurance groups. However, these factors were not directly measured in this study. More broadly, these insurance-related differences should be considered alongside nationwide evidence, showing that hypertensive inpatients with multimorbidity face significantly higher risks of catastrophic health expenditure, particularly among socioeconomically disadvantaged and nonemployee insurance groups [2]. Fu and Chen [2] demonstrated that multimorbidity increased the odds of catastrophic health expenditure by over 20% and that insurance protection was insufficient to offset this risk. Our temporal findings provide a disease-mechanistic explanation for these economic outcomes: more complex and accelerated multimorbidity trajectories likely translate into higher cumulative treatment costs and health care use. In contrast to studies focusing solely on expenditure outcomes [2], our work provides complementary information on the recorded disease pathways that may accompany such financial vulnerability.

County-level analyses revealed substantial heterogeneity in disease progression patterns, even within the same province. These findings are consistent with prior evidence that multimorbidity patterns vary by geographic and socioeconomic context [13,17] but extend the literature by showing that the order and speed of disease accumulation also differ across counties. For example, counties with higher progression toward peripheral vascular disease may reflect differential lifestyle exposures or health care accessibility, whereas elevated emphysema transitions following CVD may indicate environmental or occupational risks. Such place-based heterogeneity reinforces calls for localized public health strategies rather than uniform chronic disease management frameworks.

The temporal organization of the observed disease network has important implications for prevention-oriented hypertension management. The prominent transitions linking recorded hypertension with atherosclerotic heart disease, cerebral infarction, and heart failure suggest that hypertension management should be considered within a broader cardiovascular risk framework rather than focusing exclusively on blood pressure levels [18]. Rather than treating each subsequently recorded condition as an isolated event, clinicians and chronic disease management programs may benefit from considering these conditions as components of an evolving multimorbidity profile. The bridging position of type 2 diabetes between cardiovascular and multisystem disease clusters further supports integrated cardiometabolic and kidney risk assessment, including the appropriate evaluation of glycemic status and renal function [19,20]. Nevertheless, these network-defined transitions represent population-level temporal associations and should not be interpreted as deterministic or causal pathways for individual inpatients.

From a public health perspective, the heterogeneity observed across sex, age, insurance type, and county provides evidence for evaluating more risk-stratified and context-sensitive approaches to chronic disease prevention. Older inpatients, whose networks involved a broader range of metabolic, respiratory, renal, and neurological conditions, may benefit from more comprehensive, individualized care and improved coordination across clinical services [21]. County-level variation may also help identify communities in which screening, longitudinal follow-up, and integrated chronic disease services require further evaluation or prioritization. This interpretation is consistent with place-based chronic disease prevention frameworks that use local population data to inform priority setting, program planning, and resource allocation [22]. Differences across insurance groups may similarly identify populations facing greater challenges in continuity and coordination of care; however, insurance type may reflect multiple unmeasured socioeconomic and health care access factors and should not be interpreted as a causal determinant of disease progression. Therefore, the subgroup patterns identified in this study can inform local needs assessment and the development of targeted pilot strategies, but specific resource-allocation decisions would require external validation and evaluation of feasibility, effectiveness, and cost.

From a health data science perspective, this study demonstrates the value of integrating longitudinal EHR data with temporal network analytics to move beyond static descriptive analyses toward the characterization of population-level temporal patterns. Similar to prior trajectory-based frameworks [6,12], our approach enables early identification of high-risk transitions, supporting anticipatory intervention rather than reactive care. Importantly, by incorporating demographic, socioeconomic, and geographic stratification into the network analysis, our framework is consistent with precision public health approaches that use population-level data to identify groups that may benefit from more targeted and timely preventive strategies [23].

This study has several limitations. First, the retrospective observational design precludes causal inference. The directed edges in the TCN represent the temporal ordering of diagnoses recorded across successive inpatient admissions, rather than confirmed biological disease progression. Moreover, the earliest recorded occurrence of a condition in the study dataset may not correspond to its true onset or initial clinical diagnosis.

Second, the use of inpatient EHR data introduces selection considerations. Individuals whose hypertension and other chronic conditions were managed exclusively in outpatient or primary care settings were not captured, and people with limited access to hospital services may be underrepresented. Conversely, hospitalized patients may have greater morbidity or more intensive health care use and therefore more opportunities for additional conditions to be recorded. The observed temporal patterns should, consequently, be interpreted as patterns among inpatients with a recorded hypertension diagnosis at the participating hospitals and should not be generalized directly to the broader population of individuals with hypertension. Future studies integrating inpatient, outpatient, primary care, and community-based data would be valuable for assessing whether similar temporal patterns are observed across the broader hypertension population.

This study identified structured and heterogeneous temporal patterns of multimorbidity among inpatients with recorded hypertension across 3 counties in rural Henan Province. Cardiovascular and cerebrovascular conditions formed the principal recorded progression pathways, with atherosclerotic heart disease, cerebral infarction, and heart failure occupying prominent positions in the network. Type 2 diabetes appeared to connect cardiovascular pathways with broader metabolic, renal, and other multisystem complications. Differences by sex, age, insurance type, and county further indicated that recorded multimorbidity progression was not uniform across population groups.

These findings highlight the potential value of incorporating longitudinal disease patterns into hypertension-related chronic disease prevention and management. Population-specific network patterns may help identify priorities for risk assessment, continuity of care, and locally adapted service planning. However, the observed transitions represent population-level temporal associations in inpatient records rather than causal or deterministic disease pathways, and their application to specific prevention strategies requires external validation.