Multimodal Machine Learning Integrating Clinical and Proteomic Data for Early Prediction of Hypertensive Complications: A UKB Longitudinal Study Authors: Yuan Fei, Siwei Liu, Tianlang Tong, Xuemei Zhang, Hui Wang, Jun Liu, Xiaoqi Zheng Published: 2026 ABSTRACT BackgroundHypertension is a leading risk factor for cardiovascular, cerebrovascular, and renal diseases, significantly worsening prognosis and quality of life. We aimed to develop and validate multimodal machine learning models integrating clinical and proteomic features for early prediction of hypertensive complications.MethodsWe analyzed 502 166 participants from the UKB (UK Biobank). Proteomic profiling was performed using the Olink Explore platform. Clinical variables and complication outcomes were obtained from electronic health records. Features were selected using Cox proportional hazards models and light gradient boosting machine classifiers. Multimodal predictive models were constructed using random forest, with Shapley Additive Explanations applied for model interpretation.ResultsDuring follow‐up, 1232, 166, and 549 participants developed heart, brain, and kidney complications, respectively. Among 3244 candidate features, 774, 600, and 1227 were associated with these outcomes. The integrated models achieved an area under the curve of 0.73 (95% CI, 0.68–0.77) for heart disease, 0.83 (95% CI, 0.73–0.92) for brain disease, and 0.79 (95% CI, 0.73–0.85) for kidney disease. Growth/differentiation factor 15 (hazard ratio [HR], 2.16 [95% CI, 1.93–2.42]), adaptor protein 3 complex subunit σ‐2 (HR, 0.57 [95% CI, 0.42–0.78]), and tumor necrosis factor receptor superfamily member 10B (HR, 4.06 [95% CI, 3.40–4.85]) were significantly associated with their respective complications, effectively predicting the risk of clinical progression (all P<0.001).ConclusionsMultimodal machine learning models combining proteomic and clinical data enable early identification of hypertensive complications. Growth/differentiation factor 15, adaptor protein 3 complex subunit σ‐2, and tumor necrosis factor receptor superfamily member 10B may serve as potential biomarkers for risk prediction and early intervention. Nonstandard Abbreviations and AcronymsAP3S2adaptor protein 3 complex subunit σ‐2CCcystatin CCEACAM6carcinoembryonic antigen‐related cell adhesion molecule 6CXCL14chemokine ligand 14EDA2Rectodysplasin A2 receptorFCRLBFc Receptor Like BGDF15growth/differentiation factor 15HAVCR1hepatitis A virus cellular receptor 1HLSRhigh light scatter reticulocyteLGBMlight gradient boosting machineMEP1Ameprin A subunit αRNASE1ribonuclease A family member 1SHAPShapley Additive ExplanationsSORCS2VPS10 domain‐containing receptor 2TNFRSF10Btumor necrosis factor receptor superfamily member 10BUKBUK Biobank Clinical PerspectiveWhat Is New? The multimodal machine learning models, integrating baseline characteristics, lifestyle, blood assays, and proteomic data, facilitate the early identification of patients with hypertension at high risk for developing heart, brain, and kidney complications. What Are the Clinical Implications? Integrating proteomic profiles with clinical data significantly improves prediction, indicating that plasma proteins provide complementary pathophysiological information that can enhance risk prediction for future hypertensive complications.Incorporation of the hybrid models into clinical workflows allows for dynamic monitoring and stratified management of heart, brain, and kidney complications in patients with hypertension. Nonstandard Abbreviations and Acronyms adaptor protein 3 complex subunit σ‐2 cystatin C carcinoembryonic antigen‐related cell adhesion molecule 6 chemokine ligand 14 ectodysplasin A2 receptor Fc Receptor Like B growth/differentiation factor 15 hepatitis A virus cellular receptor 1 high light scatter reticulocyte light gradient boosting machine meprin A subunit α ribonuclease A family member 1 Shapley Additive Explanations VPS10 domain‐containing receptor 2 tumor necrosis factor receptor superfamily member 10B UK Biobank Clinical Perspective What Is New? The multimodal machine learning models, integrating baseline characteristics, lifestyle, blood assays, and proteomic data, facilitate the early identification of patients with hypertension at high risk for developing heart, brain, and kidney complications. The multimodal machine learning models, integrating baseline characteristics, lifestyle, blood assays, and proteomic data, facilitate the early identification of patients with hypertension at high risk for developing heart, brain, and kidney complications. The multimodal machine learning models, integrating baseline characteristics, lifestyle, blood assays, and proteomic data, facilitate the early identification of patients with hypertension at high risk for developing heart, brain, and kidney complications. What Are the Clinical Implications? Integrating proteomic profiles with clinical data significantly improves prediction, indicating that plasma proteins provide complementary pathophysiological information that can enhance risk prediction for future hypertensive complications.Incorporation of the hybrid models into clinical workflows allows for dynamic monitoring and stratified management of heart, brain, and kidney complications in patients with hypertension. Integrating proteomic profiles with clinical data significantly improves prediction, indicating that plasma proteins provide complementary pathophysiological information that can enhance risk prediction for future hypertensive complications. Integrating proteomic profiles with clinical data significantly improves prediction, indicating that plasma proteins provide complementary pathophysiological information that can enhance risk prediction for future hypertensive complications. Incorporation of the hybrid models into clinical workflows allows for dynamic monitoring and stratified management of heart, brain, and kidney complications in patients with hypertension. Incorporation of the hybrid models into clinical workflows allows for dynamic monitoring and stratified management of heart, brain, and kidney complications in patients with hypertension. Hypertension is a leading contributor to the global disease burden, affecting >1 billion adults globally. 1 , 2 According to the World Health Organization, nearly 80% of individuals with hypertension fail to manage their condition effectively, leading to significantly increased risk of serious complications. 3 Hypertension‐related complications, including stroke, renal failure, and heart disease, 4 are major causes of irreversible target‐organ damage and are closely linked to long‐term disability and death. 5 , 6 A 15‐year longitudinal study further demonstrated that diseases of the heart, brain, and kidneys accounted for >90% of all‐cause death among older patients with hypertension. 7 Therefore, the early identification of these complications is a critical public health priority. Over recent decades, several predictive models, such as the Framingham Heart Study model and KDpredict have been developed to predict cardiovascular and renal events. 8 , 9 However, these are primarily designed for the general or specific populations, potentially neglecting the specific risk profiles of individuals with hypertension. Additionally, they rely heavily on conventional clinical factors such as demographic information, lifestyle, family history, and laboratory tests, which may not fully capture the complex biological processes underlying disease progression. 10 This underscores an unmet need for models that integrate more informative biomarkers. Recent advances in proteomics and the decreasing costs of high‐throughput technologies have made plasma protein profiling increasingly accessible. 11 Several studies have used proteomic signatures as predictors of cerebral and renal complications in hypertensive populations. 12 , 13 Despite these advances, these studies are primarily retrospective and limited by relatively small sample sizes, which constrains their statistical power and generalizability in identifying complication‐associated biomarkers. Large‐scale, prospective research is needed to more effectively evaluate the predictive value of proteomic markers in complications of hypertension. In this study, we aimed to develop and validate multimodal machine learning models that integrate clinical and proteomic features to enable the early prediction of hypertensive complications. Using a large longitudinal cohort of >50 000 participants available from the UKB (UK Biobank), we first identified a set of key risk factors associated with the incidence of 3 major complications of patients with hypertension using Cox proportional hazards models and light gradient boosting machine (LGBM) classifiers. These features were then used to train and evaluate machine learning models designed to enhance risk stratification and guide early intervention. Methods Data Availability Due to the material transfer agreement with the UKB, full study data cannot be shared openly. Data from the UKB can be accessed by applying through the UKB Access Management System. This research has been conducted using the UKB Resource under Application Number 294934. Ethics Approval and Consent to Participate Each participant in the UKB studies completed a written informed consent form. The UKB study was approved by the National Information Governance Board for Health and Social Care in England and Wales, the Community Health Index Advisory Group in Scotland, and the Northwest Multi‐Center Research Ethics Committee (REC reference: 21/NW/0157). Study Participants The data for this study were obtained from the UKB, a global biomedical database that recruited >500 000 participants aged 37 to 73 years from 22 assessment centers across England, Scotland, and Wales between 2006 and 2010. 14 This study specifically included UKB data of patients with hypertension (n=152 546). Hypertension at baseline was ascertained according to the International Classification of Diseases, Tenth Revision (ICD‐10) codes relevant to hypertension (I10) or a self‐reported diagnosis by a physician. We also established the following exclusion criteria for participants: (1) complications of hypertension at baseline (n=23 831); (2) cancer at baseline (n=12 086); (3) diabetes at baseline (n=10 201); (4) no follow‐up and death (n=2260); and (5) missing baseline and lifestyle features or >20% missing blood assays or Olink protein features (n=97 775). Finally, a total of 6393 participants with hypertension were included, with a median follow‐up period of 16 years (Figure 1A and 1B). Flowchart of the participant selection. A, Individuals in the UKB cohort were excluded if they had diabetes, cancer, or hypertensive complications at baseline; were not diagnosed with hypertension at baseline; lacked follow‐up records; or had significant missing data in baseline, lifestyle, blood assay, or Olink protein features. The remaining participants were classified on the basis of their first reported years of heart disease, brain disease, and kidney disease after baseline. B, Distribution of observation times for the derived study population, whose median observation times were 16.0 years (IQR, 15.0–17.0). C, Distribution of first reported years of heart disease after baseline. D, Distribution of first reported years of brain disease after baseline. E, Distribution of first reported years of kidney disease after baseline. IQR indicates interquartile range; and UKB, UK Biobank. Blood Pressure Measurement Heart rate, brachial systolic blood pressure, and diastolic blood pressure were measured during the same imaging visit using an automated device (Omron 705, OMRON Healthcare Europe B.V., Hoofddorp, Netherlands). When the automated device could not be used, measurements were obtained with a manual mercury sphygmomanometer. The appropriate cuff size was selected according to the measured circumference at the midpoint of the upper arm. Two consecutive readings were taken on the left upper arm by a registered nurse while the participant was seated quietly, with the arm supported at heart level and free of constrictive clothing. The mean of the 2 readings was calculated and used for subsequent analyses. Clinical Variables Clinical variables were selected from the UKB database on the basis of their potential association with the risk of hypertensive complications. Baseline characteristics, including age, sex, and ethnicity, were recorded. Lifestyle factors such as smoking status, alcohol consumption, physical activity, and dietary history were also assessed. In addition, medication usage was collected to capture potential effects related to the management or treatment of hypertensive complications. Blood assays, including triglycerides, total cholesterol, low‐density lipoprotein cholesterol, high‐density lipoprotein cholesterol, and other related markers, comprehensively encompass the factors that can influence hypertensive complications. A complete list of clinical variables is provided in Tables S1 and S2. Plasma Proteomics Proteomic assay data for 54 306 participants were released by the UKB in April 2023. Detailed protocols for sample handling, plasma analysis using Olink proteomics detection methods, data processing, and quality control procedures have been described previously. 15 In brief, UKB conducted multiplexed proteomic assays on 54 306 plasma samples (primarily collected at baseline visits) using the Olink platform. This platform uses dual barcoded antibody technology to produce specific and semiquantitative readouts for 2923 protein assays in multiplex, selected to provide coverage across a range of systems and diseases. In this study, proteomic data obtained from baseline plasma samples were analyzed, and all 2923 available protein assays were incorporated into the analysis (Table S3). Complication Outcomes Diagnostic data were linked to UK electronic health records, where hypertensive complications were reported by professional clinicians in hospitals, family doctors in the primary care system, or staff in the UK death registration system. The outcomes were confirmed and classified according to the ICD‐10 codes and Data‐Coding 6 (Table S4), derived from self‐reported noncancer illness codes (field 20 002) and diagnoses (ICD‐10, field 41 270). The primary complications assessed in this study included heart disease, brain disease, and kidney disease. Heart disease includes conditions such as heart failure, myocardial infarction, heart arrhythmia, angina, other acute ischemic heart diseases, and chronic ischemic heart disease. Brain disease encompasses intracerebral hemorrhage, subarachnoid hemorrhage, cerebral infarction, and stroke. Kidney disease covers acute renal failure, chronic renal failure, and unspecified renal failure. Follow‐up visits commenced from the date of entry into the study at the assessment center and ended at the earliest date of diagnosis or death or the last date provided by the hospital or general practitioner, whichever came first. The last recorded and available date is August 2023. Outcomes were binary encoded; that is, participants who reached the endpoint during follow‐up were labeled as cases (1), whereas those remaining event‐free through censoring or the end of follow‐up were labeled as controls (0). Protein–Protein Interaction Analysis Protein–protein interaction analysis for the heart, brain, and kidney in hypertensive complications was constructed. Candidate proteins were selected through a mixed Top N strategy: the top 150 from the brain and the top 200 from the heart and kidney, ranked by Spearman correlations. A STRINGdb object was initialized with specified parameters (version 11.5, human species, score threshold of 400) and mapped candidate proteins to STRINGdb protein IDs. 16 Dynamic filtering was then applied to maintain organ balance, resulting in ≈300 to 350 high‐quality interactions. Interactions, including protein pairs, interaction scores, and organ sources, were saved as a comma‐separated values file. The data were visualized using Cytoscape version 3.10.2 software, displaying protein–protein interaction networks specific to each organ type. Feature Selection The feature selection process was performed in 2 stages. First, a Cox proportional hazards regression model was applied to identify variables significantly associated with time‐to‐event outcomes. Model assumptions were verified using Schoenfeld residual tests, and the functional forms of continuous predictors were inspected as part of model diagnostics. To mitigate the well‐known dependence of P values on sample size, candidate proteins were retained according to 3 criteria: P<0.05, hazard ratio (HR) >1.2 or HR <0.8, and a 95% CI excluding 1. In the second stage, to address class imbalance, the majority class (control) was randomly downsampled to achieve a balanced 1:1 ratio with the minority class (cases). Missing values were handled using median imputation, and features were standardized using Z score (0 mean, unit variance). Features were then grouped by type (eg, blood assays, proteomics, and other clinical variables) to ensure structured retention across modalities. Within each modality, an LGBM classifier was independently fitted using a 5‐fold cross‐validation framework to derive group‐specific feature importance rankings. The top 15 features per group (or all features for groups with <15 features) were retained. The final feature set was defined as the union of selected features across all groups. This 2‐step procedure refined the feature set by combining time‐to‐event association screening with modality‐stratified, model‐based importance ranking, yielding a set of relevant, standardized predictors for downstream analyses. Model Development We developed multimodal machine learning models using a multistep feature selection and optimization approach to improve the predictive accuracy across various hypertensive complications. The model estimates the probability of developing severe outcomes such as cardiovascular events, renal disorders, and cerebrovascular diseases, in comparison with remaining free of these complications. It was developed using UKB data that included both individuals with disease onset and corresponding control subjects. After applying the multistage feature selection framework, 80% of the data were designated as the training set, and 20% as the test set. The training cohort was further divided into 5 stratified folds using outcome‐aware splitting to preserve event rates. In each cross‐validation iteration, 4 folds were allocated for model training and 1 for validation. A random forest classifier was used, and model selection was based on maximizing the area under the curve (AUC) metric. The model achieving the highest validation AUC across the 5 folds was selected for subsequent optimization. To enhance the model's predictive performance, a feature optimization strategy was used, focusing on identifying the most informative features. All candidate features were specifically prioritized on the basis of their relative contribution to classification errors, quantified by the normalized mean difference between misclassified and correctly classified samples. The 7 most impactful proteomic features were then selected as the focus for targeted weighting and subset optimization, while all other features remained in the overall model. To further improve predictive performance, all possible subsets of these proteomic features were evaluated under uniform weighting, followed by stratified cross‐validation for model training and validation. The subset achieving the highest AUC on the independent test set was incorporated into the final weighted model. This optimized model demonstrated improved predictive performance relative to the original model, highlighting the effectiveness of error‐informed feature weighting in enhancing model accuracy and generalizability. The entire pipeline was implemented in Python version 3.8.19 (Python Software Foundation, Wilmington, DE) using key libraries including scikit‐learn version 1.3.2 for model development, pandas for data manipulation, numpy for numerical operations, imblearn for handling class imbalance, and shap for interpretability analysis. Missing values in the feature set were imputed using the median to ensure data completeness and consistency. During the model development process, the data set was split into 80% for training and 20% for testing. Within the training set, a 5‐fold stratified cross‐validation was used for feature selection and model training, ensuring balanced representation of cases and controls, while the test set remained independent and was used only for final performance evaluation. Model Validation Model performance was assessed by discrimination and calibration. Specifically, we evaluated discrimination by calculating the receiver operating characteristic AUC in the validation set, using either clinical features alone (baseline characteristics, lifestyle, medication use, and blood assays) or in combination with proteomic features, with 95% CIs calculated on the basis of 1000 bootstrap iterations. Model calibration used all the data, with bar plots indicating agreement between average predicted risk and the observed risk in each of the tertiles defined by the predicted risk score. The calibration of the multimodal model was evaluated in the validation set while retaining the original coefficients for each predictor variable. Statistical Analysis The Shapiro–Wilk test was used to assess the distribution of continuous variables. Continuous variables are presented as mean±SD, while categorical variables are described as number (percentage). We constructed Kaplan–Meier survival curves, delineating high and low baseline top feature levels to visualize the clinical progression of hypertensive complications over time. The optimal cutoff point for feature levels was established on the basis of the maximum Youden index. Cox proportional hazard models were used to estimate the differences in the prognostic value of dichotomized feature levels. The proportional hazards assumption was verified via Schoenfeld residual tests. All statistical analyses were performed using R version 4.4.1 (R Foundation for Statistical Computing, Vienna, Austria). A 2‐sided P value <0.05 indicated statistical significance for all analyses. Results Population Characteristics A total of 6393 participants were included in the study, of which 93.84% were White individuals, 53.50% were men (n=3420), and the mean age was 58.14 years (TableTable1). During a follow‐up period of ≈20 years, 1232 participants developed heart disease, 166 developed brain disease, and 549 developed kidney disease. The median time from cohort entry to the first report of disease was 9 years for heart disease, 10 years for brain disease, and 10 years for kidney disease (Figure 1C through 1E). Participants who developed heart disease were predominantly White individuals (95.21%), with a mean age of 61.33±6.310 years and 61.85% men. Similarly, brain disease cases comprised 92.77% White individuals, averaging 60.34±6.969 years, with 55.42% men. Among kidney disease cases, 94.54% were White participants, with a mean age of 61.64±6.640 years, and 58.83% were men. In contrast, complication‐free participants were younger (mean age 57.1±7.586 years) with a nearly balanced sex distribution (51.22% men). Table 1Baseline Characteristics of Participants With and Without Hypertensive Complications in the UKB CohortHeart disease (case)Brain disease (case)Kidney disease (case)No complications (control)TotalCount, n123216654947916393Age, y61.33±6.31060.34±6.96961.64±6.64057.14±7.58658.14±7.525SexFemale470 (38.15)74 (44.58)226 (41.17)2337 (48.78)2973 (46.50)Male762 (61.85)92 (55.42)323 (58.83)2454 (51.22)3420 (53.50)Race and ethnicityWhite1173 (95.21)154 (92.77)519 (94.54)4480 (93.51)5999 (93.84)Mixed10 (0.81)0 (0.00)2 (0.36)18 (0.38)29 (0.45)Asian21 (1.70)6 (3.61)10 (1.82)76 (1.59)106 (1.66)Black12 (0.97)4 (2.41)13 (2.37)158 (3.30)182 (2.85)Other13 (1.06)1 (0.60)5 (0.91)51 (1.06)65 (1.02)Unknown3 (0.24)1 (0.60)0 (0.00)8 (0.17)12 (0.19)Body mass indexObesity453 (36.77)53 (31.93)214 (38.98)1480 (30.89)2054 (32.13)Overweight560 (45.45)69 (41.57)244 (44.44)2164 (45.17)2887 (45.16)Normal weight219 (17.78)44 (26.51)90 (16.39)1138 (23.75)1442 (22.56)Underweight0 (0.00)0 (0.00)1 (0.18)9 (0.19)10 (0.16)SourceDiagnoses1177 (95.54)150 (90.36)544 (99.09)……Self‐reported7 (0.57)3 (1.81)3 (0.55)……Cause of death10 (0.81)3 (1.81)0 (0.00)……Diagnoses + self‐reported20 (1.62)6 (3.61)1 (0.18)……Diagnoses + cause of death18 (1.46)4 (2.41)1 (0.18)……Smoking statusCurrent112 (9.09)24 (14.46)53 (9.65)342 (7.14)490 (7.66)Previous494 (40.10)56 (33.73)222 (40.44)1701 (35.50)2329 (36.43)Never624 (50.65)85 (51.20)273 (49.73)2742 (57.23)3564 (55.75)Unknown2 (0.16)1 (0.60)1 (0.18)6 (0.13)10 (0.16)Hypertension medication statusNo medication720 (58.44)101 (60.84)295 (53.73)3105 (64.81)4032 (63.07)On medication512 (41.56)65 (39.16)254 (46.27)1686 (35.19)2361 (36.93)High‐density lipoprotein cholesterol, mmol/L1.29±0.0071.31±0.0211.29±0.0111.31±0.0031.30±0.003Total cholesterol, mmol/L4.56±0.0194.63±0.0604.50±0.0314.64±0.0104.62±0.009Systolic blood pressure, mm Hg154±1158±2152±1151±0152±0Follow‐up, y≤15436 (35.39)50 (30.12)178 (32.42)1805 (37.67)2365 (36.99)15–20796 (64.61)116 (69.88)371 (67.58)2986 (62.33)4028 (63.01)Continuous data are presented as mean±SD and categorical variables as n (%).UKB indicates UK Biobank. Baseline Characteristics of Participants With and Without Hypertensive Complications in the UKB Cohort Continuous data are presented as mean±SD and categorical variables as n (%). UKB indicates UK Biobank. Study Workflow Building on our analytical cohort, we developed the multimodal machine learning model following a structured workflow (Figure 2). The process begins with data acquisition, where the data set is collected on the basis of inclusion and exclusion criteria, focusing on 3 complications (heart disease, brain disease, and kidney disease) and a control group. Features are categorized into baseline characteristics, lifestyle, medication usage, blood assays, and Olink proteomic measurements (Figure 2A). In the feature‐selection stage, we first used a Cox proportional hazards model to screen candidate predictors associated with incident complications (time‐to‐event). To address class imbalance, the majority control class was randomly downsampled to achieve equality with incident cases, followed by Z score standardization. We then fitted an LGBM classifier independently within each group to rank predictors by importance and selected the top features from each predefined feature‐type group to form the final feature set for downstream modeling (Figure 2B). The workflow then proceeded to model development. Using the finalized feature set, we split the data into 80% training and 20% independent testing. We trained a random forest model with 5‐fold stratified cross‐validation on the training set and selected the model with the highest mean AUC. We then optimized this model by focusing on the 7 most impactful proteomic features, testing all subsets and incorporating the subset with the best test‐set AUC into the final weighted model. This procedure was performed separately for each complication end point. Ultimately, we retained the best‐performing model for each end point (Figure 2C). Finally, we evaluated the finalized models using multiple complementary analyses. We quantified discrimination by comparing AUCs and examined calibration and risk stratification by comparing the distribution of predicted risks with the estimated event risks across risk strata. To interpret model behavior, we performed SHAP‐based explanations to quantify each feature's contribution to the overall model output (Figure 2D). To ensure compliance with guidelines for machine learning–based prediction model studies, the Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis + Artificial Intelligence (TRIPOD+AI) checklist was completed. Workflow of the established multimodal model using the UKB and study design. A, Participants were selected using predefined inclusion and exclusion criteria, including 3 complication outcomes (heart, brain, and kidney disease) and a complication‐free control group. Predictors included baseline characteristics, lifestyle, medication use, blood assays, and Olink proteomics. B, Candidate predictors associated with incident complications (time to event) were screened using Cox models. After random undersampling and Z score standardization, predictors were grouped into 5 modalities, and LGBM was applied within each group to rank features. The top 15 predictors from each group were retained, and the final feature set was defined as their union. C, The final feature set was evaluated using an 80/20 train‐test split. Random forest models were trained with 5‐fold stratified CV, and the model with the highest validation AUC was selected. Further optimization focused on the 7 most important proteomic features, and the subset with the best test‐set AUC was incorporated into the final weighted model. This procedure was performed separately for each complication outcome. D, Model performance was evaluated by AUC comparison, calibration, risk stratification, and SHAP‐based interpretation. AUC indicates area under the curve; Base, baseline characteristics; CV, cross‐validation; LGBM, light gradient boosting machine; Life, lifestyle; Med, medication use; OLINK, Olink protein features; ROC, receiver operating characteristic; SHAP, Shapley Additive Explanations; and UKB, UK Biobank. Identifying Predictors Associated With Incident Hypertensive Complications From the initial pool of 3244 candidate predictors, a preliminary selection process was conducted to identify variables potentially associated with incident heart, brain, and kidney diseases. This process yielded 774 predictors for heart disease, 600 for brain disease, and 1227 for kidney disease, reflecting the unique risk profiles and overlapping factors among these diseases. A total of 317 predictors were shared across all 3 disease types, suggesting common biological pathways and risk factors. Beyond the shared predictors, 53 were unique to heart disease, 217 to brain disease, and 452 to kidney disease, indicating potential disease‐specific mechanisms (Figure 3A). These predictor categories include baseline characteristics, lifestyle, medication usage, blood assays, and Olink protein (Figure 3B). Figure 3C illustrates the distribution of Lithostathine‐1‐α expression values across health and disease states. Figures 3D through 3F depict forest plots of hazard ratios (HRs) for 15 key predictors. For heart disease, insulin‐like growth factor‐binding protein 4 and SORCS2 (VPS10 domain‐containing receptor 2) were the most significant predictors, with HRs of 2.14 (95% CI, 1.91–2.40) and 2.14 (95% CI, 1.87–2.46), respectively (Figure 3D). In brain disease, markers such as heart‐type fatty acid–binding protein and HAVCR1 (hepatitis A virus cellular receptor 1) emerge as potential predictors, with HRs of 1.99 (95% CI, 1.60–2.47) and 1.93 (95% CI, 1.58–2.36), respectively (Figure 3E). For kidney disease, leukocyte‐associated immunoglobulin‐like receptor 1 and smooth muscle α‐actin 2 (ACTA2) exhibited HRs of 2.21 (95% CI, 2.00–2.43) and 2.20 (95% CI, 1.99–2.44), respectively, suggesting a strong association with disease outcomes (Figure 3F). Predictor selection using Cox proportional hazards regression model. A, Venn diagram of predictors associated with heart disease, brain disease, and kidney disease. B, Distribution of predictor categories for heart disease, brain disease, and kidney disease. C, Box plot of Lithostathine‐1‐α expression values across different populations. D, Forest plot of several predictors related to heart disease. E, Forest plot of several predictors related to brain disease. F, Forest plot of several predictors related to kidney disease. Gene names are shown instead of the full‐length protein names. HLSR% indicates high light scatter reticulocyte percentage; HR, hazard ratio; LC, lymphocyte count; SS, smoking status; and UAHM, use of antihypertensive medications. Protein–protein interaction network analysis of the heart, brain, and kidney in hypertensive complications revealed significant protein interactions within and across organs (Figure 4). The network contained 116 proteins and 171 interactions, of which cross‐organ interactions account for 60.8% (104 edges). Hub proteins included epidermal growth factor receptor (degree 27), CD4, interleukin‐6, mast/stem cell growth factor receptor (KIT), TNF (tumor necrosis factor; each degree 11), disks large homolog 4, and Fas ligand (each degree 8), involved in proliferation, inflammation, and apoptosis. 17 , 18 , 19 , 20 , 21 , 22 , 23 These insights aid in understanding protein interactions among the heart, brain, and kidney in hypertensive complications and identifying potential therapeutic targets. Protein–protein interaction network of the brain–heart–kidney system in hypertensive complications. Nodes represent proteins colored by their dominant organ association, defined by the lowest P value. Heart‐related proteins are shown in deep blue (n=16), brain‐related in light green (n=18), and kidney‐related in dark green (n=82). Edges denote experimentally validated protein interactions, and their thickness corresponds to confidence scores. Interaction types are distinguished by color: orange lines mark cross‐organ interactions (104 edges: heart–kidney, 65; brain–kidney, 35; heart–brain, 4), while light gray lines represent intraorgan interactions (67 edges). Gene names are shown instead of the full‐length protein names. Features Importance Ranking Following the initial identification of complication‐associated predictors via Cox proportional hazards regression model, we applied a second‐stage feature selection using the LGBM classifier to assess and rank the predictive importance of the retained variables. As shown in the bar charts (Figures 5A through 5C), GDF15 (growth/differentiation factor 15) emerged as the most important predictor for heart disease, followed by NTproBNP (N‐terminal pro‐B‐type natriuretic peptide) and EDA2R (ectodysplasin A2 receptor). For brain disease, AP3S2 (adaptor protein 3 complex subunit σ‐2) ranked the highest, with HAVCR1 and CEACAM6 (carcinoembryonic antigen‐related cell adhesion molecule 6) as subsequent key predictors. In kidney disease, TNFRSF10B (tumor necrosis factor receptor superfamily member 10B) demonstrated the strongest predictive value, followed by CC (cystatin C) and RNASE1 (ribonuclease A family member 1). Feature importance, cumulative AUC analysis, and SHAP visualization plot of heart, brain, and kidney disease. A–C, Feature importance and cumulative AUC analysis of heart, brain, and kidney disease. The bar chart represents the importance of each predictor, ranked by its individual contribution to model classification performance based on feature importance scores derived from the trained model. The selection process follows a sequential addition of predictors in descending order of importance. The line chart (right axis) illustrates the cumulative AUC values as predictors are sequentially incorporated into the model, with shaded regions representing standard errors derived from cross‐validation. D–F, SHAP visualization plot of heart, brain, and kidney disease. The length of each horizontal bar represents the predictor's impact on model predictions; a longer bar indicates a greater impact. The color gradient of the bars, from blue (low) to red (high), reflects the magnitude of each feature, as shown by the color bar on the right. Movement along the x axis indicates the likelihood of developing disease (right) vs remaining healthy (left). Gene names are shown instead of the full‐length protein names. ApoA indicates apolipoprotein A; ApoB, apolipoprotein B; AUC, area under the curve; CC, cystatin C; CHOL, cholesterol; EC, eosinophil count; HDL‐C, high‐density lipoprotein cholesterol; HLSR%, high light scatter reticulocyte percentage; HLSRC, high light scatter reticulocyte count; IRF, immature reticulocyte fraction; LC, lymphocyte count; LDL‐d, LDL direct; MC, monocyte count; PC, platelet crit; Phe, phenylalanine; RBC, red blood cell count; RC, remnant cholesterol; RDW, red blood cell distribution width; SHAP, Shapley Additive Explanations; SS, smoking status; and UAHM, use of antihypertensive medications. For heart disease, 32 predictors were retained, including 1 baseline characteristic (sex), 1 medication usage (use of antihypertensive medications), 15 blood assays (CC, apolipoprotein A, high‐density lipoprotein cholesterol, calcium, plateletcrit, monocyte count, acetate, remnant cholesterol, glycine, high light scatter reticulocyte [HLSR] percentage, tyrosine, creatinine, immature reticulocyte fraction, phenylalanine, and HLSR count), and 15 proteins (GDF15, NTproBNP, EDA2R, natriuretic peptide B, CXCL14 [chemokine ligand 14], FCRLB [Fc Receptor Like B], matrix metalloproteinase‐12, interleukin‐18R1, netrin receptor DCC, BRICK1 Subunit of SCAR/WAVE Actin Nucleating Complex, tenascin‐R, C‐C motif chemokine 11, histidine‐rich calcium‐binding protein, spliceosome associated factor 1, and interleukin‐1R1). For brain disease, 19 variables were selected, including 1 lifestyle factor (smoking status), 3 blood assays (lymphocyte count, CC, and acetate), and 15 proteins (AP3S2, HAVCR1, CEACAM6, arginase type II, neurogenic locus notch homolog protein 1, integrin subunit β like 1, vacuolar protein sorting 37 homolog A, interleukin‐31RA, triggering receptor expressed on myeloid cells 2, desmocollin‐2, FCRLB, adenosine deaminase 2, γ‐interferon‐inducible lysosomal thiol reductase, MEP1A [meprin A subunit α], and Deltex E3 ubiquitin ligase 3). For kidney disease, 33 variables were retained, including 1 baseline characteristic (sex), 1 lifestyle factor (smoking status), 1 medication usage (use of antihypertensive medications), 15 blood assays (CC, urea, apolipoprotein B, creatinine, low‐density lipoprotein direct, cholesterol, red blood cell count, high‐density lipoprotein cholesterol, apolipoprotein A, HLSR percentage, immature reticulocyte fraction, HLSR count, eosinophil count, red blood cell distribution width, and monocyte count), and 15 proteins (TNFRSF10B, RNASE1, SORCS2, GDF15, osteoglycin, disintegrin and metalloproteinase domain–containing protein 9, angiopoietin‐2, CXCL14, CD5, tenascin‐R, apolipoprotein L1, MEP1A, FCRLB, methionine aminopeptidase 1, and nitric oxide synthase 3). More thorough feature information is presented in Table S5. SHAP plots were used to intuitively interpret the effect of each selected feature based on its value magnitude (coded by a gradient of colors) and tendency direction on the horizontal axis (the likelihood of developing hypertensive complications) (Figure 5D–5F). Predictive Performance of the Multimodal Machine Learning Model Based on the features selected and ranked in the previous 2 steps, we developed 3 machine learning models for each specific hypertension‐related complication by incorporating different sets of predictive variables. Model 1 (base model) included baseline characteristics, lifestyle factors, and medication usage. Model 2 (extended model) expanded upon this by adding blood assay data. Model 3 (complete model) further integrated proteomic features. Figure 6A through 6C presents the receiver operating characteristic curves illustrating the performance of different predictive models for heart, brain, and kidney disease. For heart disease (Figure 6A), the base model achieves an AUC of 0.58 (95% CI, 0.53–0.63), while the extended model by adding blood assays enhances the AUC to 0.63 (95% CI, 0.58–0.68). The most comprehensive model 3 yields the highest AUC of 0.73 (95% CI, 0.68–0.77), indicating substantial improvement in predictive performance. In brain disease prediction (Figure 6B), the base model yields an AUC of 0.56 (95% CI, 0.43–0.67), which increases to 0.60 (95% CI, 0.46–0.73) with the inclusion of blood assays. The complete model incorporating all available data sources, including proteomics, achieves the highest AUC of 0.83 (95% CI, 0.73–0.92), suggesting that proteomic data significantly improve model accuracy. For kidney disease (Figure 6C), the base model achieves an AUC of 0.58 (95% CI, 0.51–0.66), which improves to 0.76 (95% CI, 0.70–0.82) with the addition of blood assays. The complete model, incorporating all features, reaches an AUC of 0.79 (95% CI, 0.73–0.85), demonstrating incremental gains in predictive power with proteomics integration. Model performance and risk distribution for heart, brain, and kidney disease prediction. A–C, ROC curves of models for heart, brain, and kidney disease prediction. D–F, Predicted and estimated actual risk distribution for heart, brain, and kidney disease prediction. AUC indicates area under the curve; Base, baseline characteristics; Life, lifestyle; Med, medication use; OLINK, Olink protein features; and ROC, receiver operating characteristic. We next compared predicted versus observed risk across different tertile groups (33.33% quantiles each) for heart, brain, and kidney disease, based on the comprehensive model that incorporates clinical and proteomic features (Figure 6D–6F). For heart disease, predicted and actual risks were closely aligned in the lowest tertile (29% versus 22%; Figure 6D). In the middle tertile, predicted risk (54%) slightly exceeded the actual risk (45%), while in the highest tertile, predicted risk (68%) underestimated the actual 100%. For brain disease (Figure 6E), predicted risk in the lowest tertile (27%) notably overestimated the actual risk (8%). In the middle tertile, the predicted risk (49%) showed better alignment with observed value (57%), but in the highest tertile, prediction (69%) falls short of the actual risk (100%). For kidney disease (Figure 6F), predicted and actual risks are close in the lowest tertile (27% versus 25%). In the middle tertile, predicted risk (46%) is lower than the actual 66%, and in the highest tertile, predicted risk (68%) underestimates the actual 100%. Top Features and the Risk of Clinical Progression We next evaluated the prognostic value of the top features for the development of hypertensive complications. Baseline values of the top features were dichotomized into high and low groups based on the cutoff determined by the maximum Youden index. Patients with hypertension with elevated baseline levels of GDF15 (HR, 2.16 [95% CI, 1.93–2.42]; P<0.001) and TNFRSF10B (HR, 4.06 [95% CI, 3.40–4.85]; P<0.001) demonstrated significantly increased risks of developing heart and kidney diseases, respectively. In contrast, higher AP3S2 expression was associated with a lower risk of brain diseases (HR, 0.57 [95% CI, 0.42–0.78]; P<0.001). These results indicated that GDF15, AP3S2, and TNFRSF10B levels can effectively predict the risk of clinical progression to hypertensive complications (Figure 7). No substantial violation of the proportional hazards assumption was detected for the Cox models (Table S6). Predictive performance of baseline top feature levels on the risk of clinical progression. A–C, The Kaplan–Meier curves demonstrate clinical progression over time for heart (A), brain (B), and kidney (C) complications, visualizing hypertensive patients with low (blue lines) and high (red lines) baseline top features levels of GDF15, AP3S2, and TNFRSF10B. Baseline top features were divided into high and low groups, with the cutoff determined by the largest Youden index. Cox proportional hazard models estimate the association between the dichotomized baseline top feature and disease risk, where HRs and P values are calculated. Shaded regions represent standard errors derived from survival proportions. AP3S2 indicates adaptor protein 3 complex subunit σ‐2; GDF15, growth/differentiation factor 15; HR, hazard ratio; and TNFRSF10B, tumor necrosis factor receptor superfamily member 10B. Discussion In this study, we developed novel multimodal machine learning models using data from the large‐scale UKB cohort to predict the risk of hypertensive complications. By integrating clinical and proteomic features, our models achieved superior predictive performance compared with models based solely on clinical data. Notably, our findings emphasize the importance of top features, particularly plasma proteins GDF15, AP3S2, and TNFRSF10B, in predicting the incidence of heart, brain, and kidney diseases, respectively. Previous efforts to predict hypertensive complications have largely relied on either clinical or proteomic data. 13 , 24 , 25 , 26 For example, models have been developed to estimate stroke risk using demographic and lifestyle data, 25 or coronary heart disease using electronic health records. 26 Others have explored plasma proteomic data to predict the onset of micro‐ or macroalbuminuria in patients with hypertension. 13 However, to our knowledge, this is the first study to integrate multimodal clinical and proteomic features to simultaneously predict multiple complications of hypertensive complications. While such integrated models remain absent for hypertensive complications, similar integrative approaches have already shown promise in other disease domains, such as osteoarthritis risk 27 and hyperuricemia. 28 Our findings underscore the potential of such multimodal strategies to improve risk stratification and enable early intervention. Our multimodal model consistently outperformed the clinical‐only risk model for predicting incident heart and brain diseases in patients with hypertension, with AUC improvements of 0.10 for heart disease and 0.23 for brain disease. For renal outcomes, the improvement was modest (AUC increasing from 0.76 to 0.79) compared with using clinical risk factors alone, possibly due to overlapping information captured by both proteomic and standard blood assay features. Importantly, individuals in the top tertile of predicted risk showed an observed incident hypertensive complication rate of 100% over 10 years. Accurately identifying these high‐risk individuals, who have no prior history of hypertensive complications, is of significant interest to patients, health care providers, and health care systems. Compared with the lowest tertiles, the risk in the top tertile was 2.3‐fold higher in validation set, highlighting the model's ability to distinguish between different risk levels. However, risk calibration was suboptimal, particularly in high‐risk subgroups. This phenomenon is mainly due to the low proportion of high‐risk subgroups in the training data. Despite the large overall cohort size, the number of patients who developed each specific complication, particularly those with multiple risk factors placing them at the highest risk, remains relatively small for model training. The imbalance limits the model's exposure to the complex risk patterns characteristic of this specific population. This underscores a need for further refinement before clinical deployment. While GDF15 is an established biomarker for treatment response and prognosis in cardiovascular diseases, 29 , 30 its prognostic role in patients with hypertension has not been elucidated. Here, we extended this evidence by identifying GDF15 as the strongest predictor of incident cardiac events in this population. GDF15 is a stress‐responsive cytokine belonging to the transforming growth factor‐β superfamily, and is markedly upregulated in cardiomyocytes and cardiac tissue under conditions of inflammation, oxidative stress, and mitochondrial dysfunction. 31 In the context of hypertension, persistent pressure overload and neurohormonal activation can directly induce myocardial and vascular wall stress, thereby elevating GDF15 expression and activating Akt, ERK, and SMAD2 pathways in both endothelial cells and cardiomyocytes. Activation of these pathways promotes cell migration, nitric oxide production, cell proliferation, and possibly cardiac hypertrophy. 32 Our study also indicated the protective effect of AP3S2 in developing brain diseases for patients with hypertension. AP3S2 is a key subunit of the tetrameric adaptor protein 3, which is critical for the transport of proteins to lysosomes and lysosome‐related organelles. 33 The underlying causes and mechanisms of cerebrovascular damage caused by hypertension include the complex interaction between edema, inflammation, iron‐induced injury, and oxidative stress. 34 The lysosomal pathway mediates the process of autophagy, which is essential for neurons and neuroglial cells to break down macromolecules such as misfolded proteins in the cytoplasm and organelles. 35 We hypothesize that higher levels of AP3S2 may enhance neuronal resilience by promoting lysosomal biogenesis and function, thereby mitigating the accumulation of toxic protein aggregates and cellular debris under hypertensive stress. This proposed mechanism aligns with emerging evidence linking lysosomal dysfunction to neurodegenerative and neurovascular diseases. 36 , 37 Our study confirmed that TNFRSF10B is a pivotal predictor for renal complications in patients with hypertension. Unlike previous studies in general chronic kidney disease populations, 38 our findings specifically implicate TNFRSF10B in the context of hypertensive nephropathy. The prominent role of TNFRSF10B in kidney disease may be attributed to its ability to activate caspase‐mediated podocyte apoptosis after binding to its ligand (tumor necrosis factor–related apoptosis‐inducing ligand), leading to podocyte loss and glomerular injury, a critical event in the initiation and progression of glomerulosclerosis. 39 , 40 Clinically, the significant predictive improvement achieved by integrating plasma proteins with demographic data underscores the unique and complementary pathophysiological information carried by proteomic biomarkers. This finding may largely translate into the potential clinical utility of proteins as an additional source of discriminatory information to refine future hypertensive complication prediction. Integration of the hybrid models into clinical workflows through electronic health records and clinical decision support systems allows for dynamic monitoring and stratified management of heart, brain, and kidney complications in patients with hypertension during routine health examinations. For high‐risk groups, strengthening patient education, improving disease awareness, and promoting lifestyle improvement can potentially prevent the occurrence of heart failure, stroke, and other adverse outcomes. Meanwhile, identification of top features associated with future risk of hypertensive complications informs new prognostic biomarkers. Protein–protein interaction network analysis identified hub complication‐related proteins and their interactions, providing crucial mediators of hypertensive complication risk and has the potential to guide future investigations aimed at developing heart, brain, and kidney protective therapies in patients with hypertension. This proactive, data‐driven approach holds significant promise to mitigate the overall burden of hypertensive complications. While our models show encouraging predictive performance, several factors currently limit their direct applicability in clinical settings. First, this study is based solely on the UKB, which, despite its high‐quality longitudinal data, may limit generalizability to populations with different demographic or genetic backgrounds. As the UKB is predominantly composed of White individuals, future work should include fairness metrics and external validation in more diverse cohorts to ensure broader applicability and ethical rigor. Second, although a broad range of clinical variables was considered during feature selection, the inclusion of additional unmeasured or emerging variables such as fruit intake, lipid intake, and sleep duration may further improve the model's predictive performance. As is common with many machine learning approaches, our model prioritizes performance over mechanistic interpretability, which could limit clinical application in contexts where understanding specific causal pathways is essential. Third, the current study did not account for the heterogeneity of hypertensive complications within each target‐organ system. For instance, heart complications can include diverse conditions such as myocardial infarction, heart failure, or arrhythmias, each with distinct pathogeneses and clinical trajectories. Similarly, brain complications may encompass ischemic stroke, hemorrhagic stroke, or vascular dementia. By grouping these heterogeneous conditions under broad categories, important differences in pathophysiology and risk profiles may have been overlooked. Future work should consider stratifying complications into more granular subtypes to enhance the biological interpretability and clinical utility of predictive models. Finally, although this study used a large prospective data set, the stringent inclusion and exclusion criteria resulted in relatively few incident cases, particularly brain disease and high‐risk subgroups, which reduced statistical power and calibration ability. Despite these limitations, this study remains, to our knowledge, the largest prospective investigation to date that integrates clinical and proteomic data for the prediction of complications associated with hypertension. Conclusions Overall, reliable and practical multimodal machine learning models integrating clinical and proteomic features were developed and validated for the timely identification of hypertensive complications. Incorporating the models into clinical workflows through electronic health records could enable more accurate screening and prompt intervention for individuals at increased risk of such complications. Additionally, the analysis further highlighted the importance of top features, particularly plasma GDF15, AP3S2, and TNFRSF10B, as important predictors of heart, brain, and kidney diseases, respectively, suggesting their potential as prognostic or monitoring biomarkers in the preclinical phase of these complications. These findings may alleviate the burden of diseases associated with hypertensive complications to a certain extent. Sources of Funding This work was supported by the National Key R&D Program of China (2024YFC2309600 to X. Zheng); National Natural Science Foundation of China (62372286 to X. Zheng); Science and Technology Innovation Plan of Shanghai (23JC1403200 to X. Zheng); Guizhou Provincial Natural Science Foundation (QKHJCZK[2022]YB598 and QKHJCZK[2023]YB503 to J.L.); Zunyi Science and Technology Innovation Team for Nutrition and Health (Zun KCTD [2025] 89 to J.L.), and Science and Technology Program of Guizhou Province (QKHHBZ[2020]3002, QKHPTRC‐GCC[2022]039–1, and QKHPTRC‐CXTD[2022]014 to J.L.). Disclosures None. Supporting information Tables S1–S6 STROBE Checklist TRIPOD AI Checklist