Machine Learning Approach for Subtype Prediction in Primary Aldosteronism: A Comprehensive Analysis of Models and Features

Machine Learning Approach for Subtype Prediction in Primary Aldosteronism: A Comprehensive Analysis of Models and Features | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Machine Learning Approach for Subtype Prediction in Primary Aldosteronism: A Comprehensive Analysis of Models and Features Yosuke Mizutani, Kazutoshi Miyashita, Toshifumi Nakamura, Jumpei Nagase, and 7 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8268064/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 20 Feb, 2026 Read the published version in Scientific Reports → Version 1 posted 12 You are reading this latest preprint version Abstract Primary aldosteronism (PA) has two major subtypes: unilateral (uPA) and bilateral (bPA). Although several diagnostic models for subtype classification have been reported, the optimal combination of algorithms and clinical features remains unclear. This study aimed to identify machine learning models and clinical features that contribute to PA subtype prediction. A total of 274 PA patients who underwent successful adrenal venous sampling (AVS) at a single center were analyzed. Overall, 196 endocrine features were comprehensively collected and classified into four categories: A, PA-related features; B, challenge tests; C, general biochemistry; and D, urinary steroid profile. Five machine learning algorithms were applied; predictive performance of the models as well as predictive contribution of features and categories were evaluated. Among the models, the random forest model achieved the highest predictive accuracy (91.3%). The most contributing feature in the RF model was plasma aldosterone concentration after the captopril challenge test (CCT90-PAC). Category B showed the greatest contribution to RF, followed by Categories A, D, and C. Combining Categories A and B improved predictive performance. These findings indicate that machine learning models, particularly RF, are effective for PA subtype prediction, with challenge test–related features in Category B making a major contribution. Health sciences/Biomarkers Biological sciences/Computational biology and bioinformatics Health sciences/Diseases Health sciences/Endocrinology Health sciences/Medical research Health sciences/Risk factors primary aldosteronism steroid profile subtype prediction machine learning predictive performance predictive contribution Figures Figure 1 Figure 2 Figure 3 Introduction Primary aldosteronism (PA) is a common cause of secondary hypertension, accounting for 5–10% of hypertensive patients [ 1 , 2 ]. It results from autonomous and excessive secretion of aldosterone from the adrenal glands and is associated with a higher risk of cardiovascular disease than essential hypertension (EHT) [ 3 ]. There are two major subtypes of PA: unilateral PA (uPA), which is curable by adrenalectomy, and bilateral PA (bPA), which requires medical treatment [ 4 ]. Given that treatment strategies differ by subtype, accurate diagnosis of uPA and bPA is critical for selecting the appropriate treatment strategy. Although adrenal venous sampling (AVS) remains the gold standard for subtype diagnosis of primary aldosteronism (PA), the invasive nature, technical difficulty and limited availability restrict clinical utility [ 5 , 6 ]. In addition, PA is often underrecognized in primary care [ 7 ]. To address these challenges, several clinical algorithms have been developed to distinguish PA subtype without relying on AVS, using readily available clinical features. Machine learning models are considered suitable for accurate prediction, as they can integrate complex combinations of clinical features [ 8 , 9 ]. However, the optimal models and feature sets remain unclear, and prior studies have shown limited precision and generalizability [ 10 ]. Accordingly, a comprehensive assessment of suitable models and feature sets for PA prediction was warranted. This study examined five machine learning models and 196 clinical features useful for PA subtype prediction, under the hypothesis that machine learning approaches combined with comprehensive feature sets are beneficial for prediction. A total of 196 features were collected from 274 cases and classified into four categories: A, PA-related features; B, challenge tests and diurnal hormone sampling; C, general biochemistry and blood counts; and D, urinary steroid profile. The number of features in this study is unprecedented, and the inclusion of Categories C and D is especially notable. The comprehensive approach permitted exploration of previously unrecognized features that may have high predictive contribution. PA subtype classification typically relies on PA-related features in Category A, such as plasma aldosterone concentration (PAC) and serum potassium (K) level. In addition, the use of challenge tests such as the saline infusion test (SIT), captopril challenge test (CCT), furosemide upright test (FUT), and rapid ACTH stimulation test (AST) has been reported [ 13 – 17 ]. We incorporated features in Category B with the intention of enhancing predictive performance beyond that of Category A alone. To support early outpatient prediction in primary care, we included Category C. Urinary steroid profile features (Category D) were also added because of reported utility in subtype classification [ 18 , 19 ]. Accordingly, this study evaluated predictive performance of five machine learning models and predictive contribution of 196 individual features and four categories for PA subtype prediction under the hypothesis that comprehensive feature sets and machine learning approaches are beneficial for investigating optimal models for predicting PA subtype without relying on AVS. Methods Overview of the study flow in accordance with the TRIPOD statement. This study investigated machine learning models, features, and feature categories that are useful for subtype prediction in PA. All the methodological steps, including study design, participant selection, feature handling, modeling procedures, and internal validation, were predefined to ensure clarity, reproducibility, and transparency in accordance with the TRIPOD (Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis) statement. Predictive performance of the machine learning models, predictive contribution of individual features, and that of feature categories were evaluated. The inclusion process for the 274 PA patients is shown in Supplementary Figure 1. Study participants were identified from 526 consecutive PA patients admitted to a single institution (Keio University Hospital) in Japan between January 2013 and December 2020. PA was diagnosed in patients who met the ARR screening criteria and had a positive CCT result following Japanese or U.S. guidelines [11,12]. Among them, 411 underwent successful AVS, and the subtype was determined according to the AVS result. A total of 274 patients who completed the endocrine evaluation, including challenge tests and urinary steroid profiling, were included. This sample size is among the largest reported from a single center in Japan. A comprehensive feature collection strategy was adopted, yielding 196 predictors, including secondary variables such as ratios. This unprecedented number of features enabled broad and exploratory analyses with minimal selection bias. Missing data were handled by stratifying patients into eight groups on the basis of quartiles of age and sex. Missing values were imputed by the mean of the corresponding group. The 196 features were grouped into four categories, and their predictive contributions were analyzed. To examine useful models, five machine learning algorithms were evaluated: logistic regression (LR), support vector machine (SVM), random forest (RF), gradient boosting decision tree (GBDT), and multilayer perceptron (MLP). The 274 cases were randomly divided into a training dataset (n=205) and a test dataset (n=69). All models were trained on the training dataset, and hyperparameters were optimized via a grid search. Predictive performance of the models was analyzed on the test dataset using logarithmic loss (LogLoss), area under the receiver operating characteristic curve (ROC-AUC), accuracy, sensitivity, and specificity. To explore useful features, predictive contribution of individual features and feature categories was evaluated using model-appropriate importance metrics: regression coefficient for LR, feature importance for RF, and permutation importance for SVM, GBDT, and MLP. Several models for PA subtype prediction have been reported. In the present study, potentially more appropriate machine learning approaches and predictive features were explored as comprehensively as possible. The findings may support the development of predictive models that inform treatment selection with reduced reliance on AVS. Major limitations include the absence of residual analyses for misclassified cases and the lack of external validation, both of which warrant further investigation. Assessment of patients and collection of features The complete inpatient endocrine evaluation included challenge tests (CCT, FUT, AST), diurnal hormone sampling (8:00, 16:00, 24:00), 24-hour urine collection and AVS. uPA was defined as lateralized ratio ≧4.0 or lateralized ratio ≧2.5 with contralateral ratio <1 after ACTH stimulation. Urinary steroid profile (Category D) of 63 metabolites was analyzed, as shown in Supplementary Figure 2. A total of 196 features were used to develop prediction models. Details of the features (abbreviations, full descriptions, units of measurements and index numbers) are provided in Supplementary Table 1. Medians, interquartile ranges, and p-values for all 196 features are shown in Supplementary Table 2. The 196 features were grouped into four categories and 21 subcategories (Table 1). Briefly, PA-related features were assigned to Category A (A1–A23), challenge tests to Category B (B1–B47), general biochemistry data to Category C (C1–C37), and urinary steroid profile to Category D (D1–D89). This grouping was expected to reduce multicollinearity among features and thereby enhance the reliability and interpretability of the models. Features in Categories A and C were collected at the initial outpatient visit, while those in Categories B and D were measured during hospitalization. Antihypertensive medications were not restricted at initial visit. Before hospitalization, all antihypertensive agents except calcium channel blockers and α-blockers were discontinued for at least two weeks to avoid affecting PAC, ARC, and aldosterone renin ratio (ARR). Development of machine learning-based prediction models Prediction models for PA subtype were constructed using five machine learning algorithms: LR, SVM, RF, GBDT, and MLP with Python 3.11, scikit-learn 1.5.0, and LightGBM 4.4.0 [20]. LR and SVM are regression models, RF and GBDT are decision tree models, and MLP is a deep learning model. All the models were trained on the training dataset, and the hyperparameters were optimized via a grid search using four-fold cross-validation [21]. Following model development, the models were used to predict PA subtype in the test dataset. The models generated a prediction score (0–1) of uPA for each case, and cases with scores≧0.5 were classified as uPA. Predictive performance of machine learning models and predictive contribution of features and categories To examine the models and features useful for predicting PA subtype, predictive performance of five machine learning models and predictive contributions of 196 features and four feature categories were analyzed. Predictive performance was estimated on the test dataset (n=69) using the following standard metrics: logarithmic loss (LogLoss), area under the receiver operating characteristic curve (ROC-AUC), accuracy, sensitivity, and specificity. Models that achieved higher ROC-AUC values or accuracies were considered clinically useful. Predictive contribution of features was assessed using model-appropriate importance metrics: regression coefficient for LR, feature importance for RF, and permutation importance for SVM, GBDT, and MLP; all metrics were normalized so that the sum of 196 features equaled 1. Predictive contribution of categories and subcategories was calculated by summing the metrics of the respective features. The features were sorted by the importance metrics in descending order and presented as indicators of the predictive contribution. Ethical statement All methods were carried out in accordance with relevant guidelines and regulations, including the Ethical Guidelines for Medical and Health Research Involving Human Subjects in Japan. The study protocol was reviewed and approved by the Ethics Committee of Keio University School of Medicine (Approval No. 20180135). Informed consent was obtained from all participants who were able to provide consent, after oral explanation using a written information sheet. In addition, an opt-out approach was implemented on the institutional website in accordance with institutional and national regulations. Results Statistical analysis of features between bPA and uPA To explore features relevant to PA subtype prediction, we conducted statistical comparisons between the bPA and uPA groups using 196 features from a single-center cohort of 274 patients. P-values were calculated for each feature, and the top 40 features with the lowest p-values were visualized (Figure 1A). The top 10 features from each of the four Categories (A–D) are shown separately (Figure 1B), with a detailed list of the results (Table 1B). The three most statistically significant features were plasma aldosterone concentration at 90 minutes after captopril challenge (CCT90-PAC; Category B, p = 1.74×10⁻¹⁹), PAC at 8:00 (Category B, p = 3.55×10⁻¹⁶), and tetrahydro-aldosterone (THAld) (Category D, p = 6.55×10⁻¹⁶). Among the top 40 features, 20 belonged to Category B (47 total), 12 to Category D (89 total), and 8 to Category A (23 total); no features from Category C (37 total) were included. Representative features with the lowest p-values in each category were as follows: serum K, 24-hour urinary aldosterone, and tumor size (Category A); PAC after CCT or AST, and at 8:00 (Category B); mean corpuscular volume (MCV), mean corpuscular hemoglobin concentration (MCHC), and creatinine (Category C); and THAld and 18-hydroxy-tetrahydro-11-dehydrocorticosterone (18OHTHA) (Category D). The utility of the feature with the lowest p-value in each category (K, CCT90-PAC, MCV, and THAld) for PA subtype classification was analyzed. As shown in Figure 1C and Table 1C, CCT90-PAC had the highest ROC-AUC (0.836) and classification accuracy (0.818). Figure 1C and Table 1C present the diagnostic performance when each feature was applied to all included cases (n=274), whereas Figures 2 and 3 show predictive performance of machine learning models with multiple features evaluated on the test dataset (n=69). Given that individual features alone might lack diagnostic accuracy, machine learning models using multiple features were expected to improve the performance of subtype prediction. Predictive performance of machine learning models Based on the hypothesis that the machine learning approach is beneficial for prediction using a large number of features, we developed subtype prediction models using five algorithms: LR, SVM, RF, GBDT, and MLP algorithms. The prediction performance of each model was evaluated by five metrics: LogLoss, ROC-AUC, accuracy, sensitivity, and specificity. The total dataset (n=274) was pre-divided into a training dataset (n=205; 75%) and a test dataset (n=69; 25%). Predictive models were developed through training on the training dataset. The classification accuracy of PA subtype within the internal training dataset (n=205) and the hyperparameters are shown in Supplementary Table 3A. Following model development, the models were used for predicting PA subtype in the test dataset (n=69). Dot plots and performance metrics for the prediction are shown in Figure 2A and Table 2A. The ROC-AUC scores were 0.920 (LR), 0.932 (SVM), 0.972 (RF), 0.929 (GBDT), and 0.926 (MLP). Predictive accuracy followed a similar trend: 0.797 (LR), 0.870 (SVM), 0.913 (RF), 0.884 (GBDT), and 0.884 (MLP). RF demonstrated the best overall performance, indicating the strength of decision tree-based approaches for subtype prediction. Predictive contribution of categories and subcategories Among the five machine learning models, the RF model achieved the highest predictive accuracy (0.913). To evaluate the utility of each feature category for subtype prediction, we constructed RF models using features from each category (A, B, C, or D) and from combinations of Category A with other categories (A&B, A&C, A&D), in addition to those from all categories. Hyperparameters were optimized for each configuration (Supplementary Table 3B). Performance metrics for the RF models and the ROC curves, which were constructed using single-category or the combinations, are shown in Table 2B and Figure 2B. The accuracy was highest for Category B, followed by Categories A, D, and C. Combining Categories B and A improved predictive performance in the test set, whereas adding Category C reduced the accuracy. Although the A&D combination achieved an excellent classification accuracy of 0.995 in the training dataset, predictive accuracy in the test dataset was 0.884 (Table 2B), suggesting that adding more features did not necessarily enhance the predictive performance of the RF model. Predictive contribution of each category, which is based on model-appropriate importance metrics, that is, regression coefficient (LR), feature importance (RF), and permutation importance (SVM, GBDT, and MLP), is presented in Table 2C. The categories that most significantly contributed to PA subtype prediction in each predictive model were Category A in LR (regression coefficient 0.611), Category B in SVM, RF, and GBDT (RF feature importance 0.475), and Category D in MLP (permutation importance 0.411). Subcategory-level contributions are summarized in Table 2D. The subcategories that contributed the most were tumor size (Subcategory Af) in LR and SVM, CCT (Bc) in RF, and urinary steroid metabolites (Da) in GBDT and MLP. Among the subcategories of challenge tests within Category B, those that showed the highest contribution to PA subtype prediction were in the order of CCT, AST, and FUT, in LR, RF, and GBDT; whereas AST was the highest in SVM and MLP. Predictive contribution of features To identify contribution of individual features to PA subtype prediction, we evaluated the importance metrics, namely, the regression coefficient (LR), feature importance (RF), and permutation importance (MLP), of individual features as indices of predictive contribution. First, we analyzed feature importance in the best-performing RF model. The top 40 features are visualized in Figure 3A, and the top 10 features in each category are shown in Figure 3B. A complete list of the top 40 features is shown in Table 3A. The top three predictors in the RF model were CCT90-PAC (Category B, 0.212), tumor size (Category A, 0.145), and aldosterone/cortisol ratio (A/C ratio) at 8:00 (Category B, 0.050). Some Category D features also contributed substantially. Overall, features with small p-values (Table 1B) still contributed to the prediction of PA subtype at least in the RF model. This accounts for the similarity in feature lists between Tables 1B and 3A. In addition, features such as tumor size were more important in the RF model, likely because they provided unique information not captured by other variables. A full list of feature importance values of all 196 features is provided in Supplementary Table 2. Predictive contribution to the LR model (regression coefficient) is shown in Figure 3C and Table 3B, and that to the MLP model (permutation importance) in Figure 3D and Table 3C. The LR model automatically selected and used 9 features, due to regularization to prevent feature redundancy, most of which overlapped with the low p-value features in Table 1B. In contrast, the MLP model showed relatively uniform permutation importance across features in the presence of multicollinearity among similar features, and even those with relatively high p-values still demonstrated substantial predictive contributions. These findings suggest that while LR and RF emphasize statistically significant features, MLP assigned relatively uniform contribution to each feature. Discussion On the basis of the hypothesis that machine learning models using a comprehensive feature set are beneficial for PA subtype prediction, this study investigated the predictive performance of five models (based on LR, SVM, RF, GBDT, and MLP) and the predictive contribution of 196 features across four categories. Decision tree models (RF, GBDT) and the deep learning model (MLP) outperformed regression models (LR, SVM), with RF showing the highest performance. Category B (challenge tests and diurnal hormone sampling) was identified as the category that contributed the most to PA subtype prediction in the RF model, whereas Category C (general biochemistry) was found to lack predictive value and may adversely affect model performance. Category D (steroid profile) played a significant role in the MLP model. Although the contributions of individual features varied across models, CCT90-PAC was a useful feature in RF. The academic significance of this study lies in 1) evaluation of the predictive performance of multiple machine learning models, 2) collection of a comprehensive set of 196 clinical features to explore unknown predictive features, and 3) systematic categorization of features to quantify category- and subcategory-level predictive contributions. To our knowledge, this is the first study to evaluate machine learning models using a broad feature set that includes challenge tests and diurnal hormonal sampling (Category B), general biochemical parameters and blood counts (Category C), and urinary steroid profile (Category D), along with PA-related features (Category A). Previous studies have constructed models to differentiate PA subtype, reporting sensitivity of 32–95% and specificity of 46–100% [22–27], including studies using machine learning-based prediction of PA subtype [28–29]. Kaneko et al. applied four machine learning methods (LR, SVM, RF, GBDT) to 229 PA cases, using 80% for training and 20% for testing, and reported that RF achieved the highest predictive performance (ROC-AUC 0.990, accuracy 95.7%). Regarding features that are useful for prediction, the utility of CCT and SIT has been documented [30]. Eisenhofer et al. reported a model using urinary steroid profile that achieved 100% sensitivity and 98% specificity for diagnosing uPA with KCNJ5 variants, whereas the ROC-AUC was 0.716 in wild-type patients [19]. This study originally aimed to explore features that are useful for PA subtype prediction through a comprehensive feature collection approach. Machine learning-based models were introduced because they can handle extensive feature sets effectively. To address multicollinearity arising from the large number of features, we organized potentially redundant features into categories and subcategories and evaluated the relative contributions of each category. Category A features, including tumor size, K, PAC and ARR, have conventionally been identified as essential for subtype classification [31–36]. PA is more strongly associated with cardiovascular disease than essential hypertension. In particular, uPA tends to be more severe than bPA and is characterized by higher PAC and lower serum K [37–39]. These findings underscore the importance of Category A features in subtype . Among other biochemical markers, Na, Ca, intact PTH, and BNP also reflected aldosterone’s impact on electrolyte balance and contributed to subtype prediction in the RF model, which is consistent with prior reports. In Category B, we assessed CCT, AST, and FUT; however, SIT was not included, as it was not routinely performed at our facility. In this study, CCT and AST emerged as key predictors, which is consistent with previously reported findings for subtype prediction [40,41]. Notably, CCT90-PAC was the top contributor in the RF model, surpassing all Category A features. The superiority of CCT90-PAC over baseline PAC aligns with the well-recognized fact that PAC is often insufficiently suppressed by captopril in patients with APA. For CCT, absolute PAC values were more predictive than ARR or PAC fold increases were. Moreover, AST120 A/C ratio ranked higher than AST120-PAC. Category C included general clinical parameters such as liver and renal function, blood counts, and markers of glucose and lipid metabolism, excluding electrolytes in Category A. Category C features showed limited predictive utility in the RF model, and adding them to Category A reduced performance, indicating that Category C introduced noise and attenuated the predictive contribution of Category A. Although lower cholesterol and triglycerides in uPA have been reported [42], only urinary C-peptide emerged as a notable contributor in the present RF model. Subcategory analysis indicated minimal added value from blood counts, liver function, and metabolic markers. Category D comprises inpatient urinary steroid profiles. In previous reports, hybrid steroids which possess both glucocorticoid- and mineralocorticoid-like structures (17α-hydroxyl and 18-hydroxyl or -oxo groups; 18OHF, etc.) have been reported to associate with uPA and may contribute to subtype classification [43–47]. In this study, 18OHTHA, a urinary metabolite of an intermediate in aldosterone biosynthesis, was the most contributory Category D feature in the RF model, surpassing urinary aldosterone or other metabolites, including 18OHF. Elevation of 18OHTHA may suggest increased CYP11B2 (aldosterone synthase) activity in the adrenal cortex. In addition to individual metabolites, we included metabolite sums and ratios reflecting enzyme activity. Among these ratios, the 18OHF/(THE+THF) ratio had the highest predictive contribution. These findings support the utility of urinary steroid metabolites, particularly 18OHTHA and 18OHF, as predictors of PA subtype. Feature category and subcategory-level contributions to subtype prediction were evaluated. In the RF model, Category B contributed the most, followed by Categories A, D, and C. Adding Category B to Category A improved predictive performance, and inclusion of all categories yielded further gains. Category C showed limited predictive contribution. Subcategory analysis within Category A indicated that tumor size was the largest contributor. The usefulness of the challenge test features within Category B for subtype classification has been reported in previous studies. However, it remains unclear which test (CCT, FUT or AST) is useful. In this study, parameters from each test were subcategorized to facilitate comparisons. In the RF model, CCT and AST showed greater predictive contribution than FUT. Within Category D, several individual metabolites demonstrated high contributions, whereas metabolite sums and ratios contributed less, possibly reflecting the inclusion of non-informative components. With respect to effective machine learning models, predictive accuracy on the test dataset ranged from 0.797 to 0.913, indicating relatively high performance. Decision tree models (RF, GBDT) and MLP outperformed regression-based models (LR, SVM). The LR model, limited by its linear nature, could not capture feature interactions; however, it offered high interpretability, emphasizing significant features such as serum K, CCT90-PAC, and tumor size. Lasso regularization reduced the LR model to nine features, most of which overlapped with variables showing low p-values in Table 1B. The relatively poorer performance of LR may partly reflect overreliance on the limited features. The nonlinear regression model, SVM, outperformed LR and achieved 100% specificity for uPA prediction, although the sensitivity was low. Decision-tree models capture feature interactions effectively, resulting in higher accuracy. The RF model exceeded 90% accuracy and integrated steroid-related features effectively to enhance prediction. While GBDT often achieves greater performance than RF does, the limited data in this study may have led to overfitting. The deep learning model MLP achieved high predictive accuracy, however, the contributory features differed substantially from those in the other algorithms. Several features with limited contributions in other models (e.g., MCHC) were utilized more effectively in MLP, likely because it can leverage complex interactions among variables. Nevertheless, permutation importance values in MLP may be configuration-dependent under multicollinearity. Interpretation should focus on broader feature categories rather than individual features. Consequently, predictive contributions of features in MLP were relatively uniform, individual feature-level contributions were small, and category level contributions were proportional to the number of features in each category. Features from urinary steroid profile (Category D) were effectively utilized in MLP, with several individual metabolites showing higher contributions. Notably, despite the higher accuracy achieved by MLP, it exhibited distinct outliers in both bPA and uPA predictions (dot plots in Figure 2A), indicating a tendency toward confident misclassification. False positive predictions in MLP may reflect clinically meaningful uPA-like features, as observed in bilateral APAs. Moreover, the sensitivity for uPA remained below 0.765 across models, suggesting that some cases of AVS-confirmed uPA may exhibit bPA-like features. In contrast, the specificity was high: 0.904 (LR), 0.923 (MLP), 0.962 (RF), 0.962 (GBDT), and 1.000 (SVM). Thus, despite similar overall accuracies, the models differed in sensitivity–specificity profiles, suggesting model-specific clinical implications. Several limitations should be noted. First, the sample size (n=274) might be insufficient for reliable development of machine learning models, although it represents one of the largest single-center AVS cohorts in Japan. Second, multicollinearity among redundant features was present and might distort estimates of predictive contribution. Third, the contribution of features differed among models, complicating identification of consistently useful predictors. Fourth, the saline infusion test (SIT) was not performed, limiting direct comparisons among challenge tests. Finally, urinary steroid profile is not widely available and AVS criteria for subtype classification are not fully standardized [48]; thus, the results may vary in other settings. The aim of this study was not to present a single best model or a fixed ranking of predictors, but to compare the performance of five machine learning algorithms and assess the relative usefulness of features and predefined feature categories. Despite these limitations, the study has notable strengths: 1) five distinct machine learning models were examined and the potential utility of RF for predicting PA subtype was demonstrated; 2) a broad set of 196 clinical features including urinary steroid metabolites was collected, with CCT90-PAC identified as a useful predictor in the RF model; and 3) feature categories and subcategories were defined to reduce the impact of multicollinearity among features and improve interpretability, thereby clarifying the utility of challenge tests (Category B) for subtype prediction. In summary, machine learning-based models using 196 features could effectively predict PA subtype. The RF model achieved ≧90% predictive accuracy on the test dataset, with Category B, which includes challenge tests, contributing most to the prediction. These findings suggest that predictive models based primarily on Category B features can achieve superior predictive performance with a small feature set, supporting clinically feasible treatment selection with reduced reliance on AVS. External validation and misclassification analysis remain essential to refine models and achieve reliable PA subtype prediction. Declarations Acknowledgements As host members of ISH 2022 Kyoto (International Society of Hypertension Scientific Meeting, Kyoto, Japan), the authors gratefully acknowledge the meeting participants for their academic insights that informed the conception of this study. Author contributions Y.M., K.M., To.N. and J.N. conceived the study, designed the analysis and data collection, and drafted the manuscript Ke.H., Te.N. and H.M. contributed to data curation. M.N., K.K., H.I. and Ka.H. critically reviewed the manuscript and provided important interpretation of the data. All authors reviewed and approved the final version of the manuscript. Data availability statement The datasets are available from the corresponding author on reasonable request. The Python code used in this study is publicly available on GitHub. Additional Information COI statement: The authors declare no conflicts of interest. Sources of Funding: This study was conducted as a voluntary activity of authors with no external funding. References Young WF. Primary aldosteronism: renaissance of a syndrome. Clin Endocrinol (Oxf). 2007;66:607–618. Monticone S, et al. Prevalence and clinical manifestations of primary aldosteronism encountered in primary care practice. J Am Coll Cardiol. 2017;69:1811–1820. Ohno Y; Nagahama Study / JPAS Study Group. Prevalence of cardiovascular disease and its risk factors in primary aldosteronism: a multicenter study in Japan. Hypertension. 2018;71:530–537. Funder JW, et al. The management of primary aldosteronism: case detection, diagnosis, and treatment. J Clin Endocrinol Metab. 2016;101:1889–1916. Vonend O, et al. Adrenal venous sampling: evaluation of the German Conn’s registry. Hypertension. 2011;57:990–995. Monticone S, et al. Clinical management and outcomes of adrenal hemorrhage following adrenal vein sampling in primary aldosteronism. Hypertension. 2016;67:146–152. Mulatero P, et al. Guidelines for primary aldosteronism: uptake by primary care physicians in Europe. J Hypertens. 2016;34:2253–2257. Baştanlar Y, Özuysal M. Introduction to machine learning. In: Yousef M, Allmer J, eds. Methods Mol Biol. 2014;1107:105–128. Cruz JA, Wishart DS. Applications of machine learning in cancer prediction and prognosis. Cancer Inform. 2007;2:59–77. Sam D, et al. External validation of clinical prediction models in unilateral primary aldosteronism. Am J Hypertens. 2022;35:365–373. Nishikawa T, et al. Guidelines for the diagnosis and treatment of primary aldosteronism — The Japan Endocrine Society 2009. Endocr J. 2009;58:711–721. Young WF Jr, et al. Role for adrenal venous sampling in primary aldosteronism. Surgery. 2004;136:1227–1235. Küpers EM, et al. A clinical prediction score to diagnose unilateral primary aldosteronism. J Clin Endocrinol Metab. 2012;97:3530–3537. Ying Z, et al. Identifying unilateral disease in Chinese patients with primary aldosteronism by using a modified prediction score. J Hypertens. 2017;35:2486–2492. Nanba K, et al. Shortened saline infusion test for subtype prediction in primary aldosteronism. Endocrine. 2015;50:802–806. Rossi GP, et al. Comparison of the captopril and the saline infusion test for excluding aldosterone-producing adenoma. Hypertension. 2007;50:424–431. Sonoyama T, et al. Significance of adrenocorticotropin stimulation test in the diagnosis of aldosterone-producing adenoma. J Clin Endocrinol Metab. 2011;96:2771–2778. Eisenhofer G, et al. Mass spectrometry–based adrenal and peripheral venous steroid profiling for subtyping primary aldosteronism. Clin Chem. 2016;62:514–524. Eisenhofer G, et al. Use of steroid profiling combined with machine learning for identification and subtype classification in primary aldosteronism. JAMA Netw Open. 2020;3:e2016209. Pedregosa F, et al. Scikit-learn: machine learning in Python. J Mach Learn Res. 2011;12:2825–2830. Bergstra J, Bengio Y. Random search for hyper-parameter optimization. J Mach Learn Res. 2012;13:281–305. Kobayashi H, et al. Scoring system for the diagnosis of bilateral primary aldosteronism in the outpatient setting before adrenal venous sampling. Clin Endocrinol (Oxf). 2017;86:467–472. Nanba K, et al. A subtype prediction score for primary aldosteronism. J Hum Hypertens. 2014;28:716–720. Buffolo F, et al. Subtype diagnosis of primary aldosteronism: is adrenal vein sampling always necessary? Int J Mol Sci. 2017;18:848. Chen S, et al. Computed tomography combined with confirmatory tests for the diagnosis of aldosterone-producing adenoma. Endocr J. 2021;68:299–306. He K, et al. A clinical radiomic nomogram based on unenhanced computed tomography for predicting the risk of aldosterone-producing adenoma. Front Oncol. 2021;11:634879. Burrello J, et al. Development and validation of prediction models for subtype diagnosis of patients with primary aldosteronism. J Clin Endocrinol Metab. 2020;105:e3706–e3717. Kaneko H, et al. Machine learning based models for prediction of subtype diagnosis of primary aldosteronism using blood test. Sci Rep. 2021;11:9140. Tamaru S, et al. Machine learning approach to predict subtypes of primary aldosteronism is helpful to estimate indication of adrenal vein sampling. High Blood Press Cardiovasc Prev. 2022;29:375–383. Byun JM, et al. A case of primary aldosteronism presenting as non–ST elevation myocardial infarction. Korean J Intern Med. 2013;28:739–742. Zhang Y, et al. Is primary aldosteronism a potential risk factor for aortic dissection? A case report and literature review. BMC Endocr Disord. 2020;20:115. Fava C, et al. Unusual presentation of primary aldosteronism with advanced target organ damage: a case report. Radiol Case Rep. 2019;14:814–818. Kamemura K, et al. Significance of adrenal computed tomography in predicting laterality and indicating adrenal vein sampling in primary aldosteronism. J Hum Hypertens. 2017;31:195–199. Kobayashi H, et al.; JPAS Study Group. Development and validation of subtype prediction scores for the workup of primary aldosteronism. J Hypertens. 2018;36:2269–2276. Kocjan T, et al. A new clinical prediction criterion accurately determines a subset of patients with bilateral primary aldosteronism before adrenal venous sampling. Endocr Pract. 2016;22:587–594. Umakoshi H, et al. Significance of computed tomography and serum potassium in predicting subtype diagnosis of primary aldosteronism. J Clin Endocrinol Metab. 2018;103:900–908. Rossi GP, et al. Hyperparathyroidism can be useful in the identification of primary aldosteronism due to aldosterone-producing adenoma. Hypertension. 2012;60:431–436. Kato J, et al. Atrial and brain natriuretic peptides as markers of cardiac load and volume retention in primary aldosteronism. Am J Hypertens. 2005;18:354–357. Jakubik P, et al. Impact of essential hypertension and primary aldosteronism on plasma brain natriuretic peptide concentration. Blood Press. 2006;15:302–307. Xiao M, et al. Evaluation of the saline infusion test and the captopril challenge test in Chinese patients with primary aldosteronism. J Clin Endocrinol Metab. 2018;103:853–860. Moriya A, et al. ACTH stimulation test and computed tomography are useful for differentiating the subtype of primary aldosteronism. Endocr J. 2017;64:65–73. Ohno Y; JPAS Study Group. Obesity as a key factor underlying idiopathic hyperaldosteronism. J Clin Endocrinol Metab. 2018;103:4456–4464. Zhou Y, et al. Diagnostic accuracy of adrenal imaging for subtype diagnosis in primary aldosteronism: systematic review and meta-analysis. BMJ Open. 2020;10:e038489. Mulatero P, et al. 18-hydroxycorticosterone, 18-hydroxycortisol, and 18-oxocortisol in the diagnosis of primary aldosteronism and its subtype. J Clin Endocrinol Metab. 2012;97:881–889. Satoh F, et al. Measurement of peripheral plasma 18-oxocortisol can discriminate unilateral adenoma from bilateral disease in patients with primary aldosteronism. Hypertension. 2015;65:1096–1102. Lenders JWM, et al. Diagnosis of endocrine disease: 18-oxocortisol and 18-hydroxycortisol — is there clinical utility of these steroids? Eur J Endocrinol. 2018;178:517–563. Monticone S, et al. Immunohistochemical, genetic and clinical characterization of sporadic aldosterone-producing adenomas. Mol Cell Endocrinol. 2015;411:146–154. Rossi GP, et al. An expert consensus statement on use of adrenal vein sampling for the subtyping of primary aldosteronism. Hypertension. 2014;63:151–160. Table Table 1. Overview of four Categories and 21 Subcategories covering the 196 features Category A: PA-related features in the outpatient setting 23 features Aa general status 5 Ab circulatory system 4 Ac renin-aldosterone system 3 Ad Electrolytes 6 Ae 24h urine 4 Af tumor size 1 Category B: Challenge tests and diurnal profiles in the inpatient setting 47 features Ba daily profile of aldosterone 7 Bb daily profile of cortisol 7 Bc captopril challenge test (CCT) 6 Bd furosemide upright test (FUT) 12 Be ACTH stimulation test (AST) 15 Category C: General biochemical tests and blood counts 37 features Ca blood counts 13 Cb hepatic function 4 Cc renal function 2 Cd lipid profile 3 Ce glucose metabolism 8 Cf Others 4 Cg 24h urine others 3 Category D: Urinary steroid profile 89 features Da urinary metabolite 63 Db total of metabolites 13 Dc ratio of metabolite 13 Total: 196 features The 196 features used in the analysis, grouped into four Categories and 21 Subcategories are summarized. The number of features in each Category or Subcategory is shown. Category A (A1–A23, blue) represents PA-related features examined in the outpatient setting. Category B (B1–B47, red) includes endocrine challenge tests and diurnal hormonal profiles in the inpatient setting. Category C (C1–C37, green) consists of outpatient-based general biochemical tests and blood counts. Category D (D1–D89, gold) covers the urinary steroid profile, as further detailed in Supplementary Figure 2. The predictive contribution of each Category and Subcategory was investigated in machine learning-based subtype prediction models. Additional Declarations No competing interests reported. Supplementary Files SupplementaryMaterial.docx Supplementary Information is available Supplementary Material: Extended Methods, Supplementary Tables 1 to 3, and Legends for Supplementary Figures SupplementaryFigures.pptx Supplementary Figures TableData.xlsx Cite Share Download PDF Status: Published Journal Publication published 20 Feb, 2026 Read the published version in Scientific Reports → Version 1 posted Editorial decision: Revision requested 20 Jan, 2026 Reviews received at journal 19 Jan, 2026 Reviewers agreed at journal 19 Jan, 2026 Reviews received at journal 31 Dec, 2025 Reviewers agreed at journal 12 Dec, 2025 Reviewers agreed at journal 11 Dec, 2025 Reviewers agreed at journal 10 Dec, 2025 Reviewers invited by journal 09 Dec, 2025 Editor assigned by journal 09 Dec, 2025 Editor invited by journal 08 Dec, 2025 Submission checks completed at journal 06 Dec, 2025 First submitted to journal 06 Dec, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8268064","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":560185001,"identity":"bb84cab5-ca0b-4c58-99da-86b8ff868eb6","order_by":0,"name":"Yosuke Mizutani","email":"","orcid":"","institution":"Keio University School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Yosuke","middleName":"","lastName":"Mizutani","suffix":""},{"id":560185002,"identity":"0818e0c9-d136-4f8c-aec5-ddaf74279437","order_by":1,"name":"Kazutoshi Miyashita","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAtklEQVRIiWNgGAWjYDACZgYGxgYGBjkJMMVDghZjErQwQLQkziDaXbrtPIYfZ9TUpc9sb258wCBzh7AWs8M8xpIbjrHlzuY52GzAwPOMKC0Gkg/YeHLnSSS2STDwHCbOlp8P/kmky8k/JF6LmeTGNoMEaQlGorWwlVnO7EswnNmT2GyQQJRfzh/efLPnW528xPHjDx987CEixBgYOAwQ7MSeA8RoYX+AxPlBlJZRMApGwSgYYQAAKMI56sSjspEAAAAASUVORK5CYII=","orcid":"","institution":"Keio University School of Medicine","correspondingAuthor":true,"prefix":"","firstName":"Kazutoshi","middleName":"","lastName":"Miyashita","suffix":""},{"id":560185003,"identity":"b0c38f1e-0068-4dc1-955b-d4f9846ae094","order_by":2,"name":"Toshifumi Nakamura","email":"","orcid":"","institution":"Keio University School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Toshifumi","middleName":"","lastName":"Nakamura","suffix":""},{"id":560185004,"identity":"8e573b16-4b0c-4a7d-9cfa-a793e0b02cad","order_by":3,"name":"Jumpei Nagase","email":"","orcid":"","institution":"University of Electro-Communications","correspondingAuthor":false,"prefix":"","firstName":"Jumpei","middleName":"","lastName":"Nagase","suffix":""},{"id":560185005,"identity":"5738d00d-8c63-46e8-a7bf-37df9555acb6","order_by":4,"name":"Masato Nishihara","email":"","orcid":"","institution":"Cliffhanger Inc.","correspondingAuthor":false,"prefix":"","firstName":"Masato","middleName":"","lastName":"Nishihara","suffix":""},{"id":560185006,"identity":"ed8092c8-8aa8-4310-9a0a-a7ba31083f8c","order_by":5,"name":"Keiko Honma","email":"","orcid":"","institution":"Keio University School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Keiko","middleName":"","lastName":"Honma","suffix":""},{"id":560185007,"identity":"6e022366-a165-4135-9100-f4b27f20fe78","order_by":6,"name":"Terumichi Nakagawa","email":"","orcid":"","institution":"Keio University School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Terumichi","middleName":"","lastName":"Nakagawa","suffix":""},{"id":560185008,"identity":"4a74f4c2-bce8-414e-a251-6fc22cc3bb45","order_by":7,"name":"Hiromichi Matsushita","email":"","orcid":"","institution":"Keio University School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Hiromichi","middleName":"","lastName":"Matsushita","suffix":""},{"id":560185009,"identity":"a499cf75-7b64-4447-ae8c-1d42bdbb574a","order_by":8,"name":"Kenichiro Kinouchi","email":"","orcid":"","institution":"Keio University School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Kenichiro","middleName":"","lastName":"Kinouchi","suffix":""},{"id":560185010,"identity":"eab7dd93-18e2-4c6a-a79e-6a0ce8ab3578","order_by":9,"name":"Hiroshi Itoh","email":"","orcid":"","institution":"Keio University School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Hiroshi","middleName":"","lastName":"Itoh","suffix":""},{"id":560185011,"identity":"9797bf04-d8c7-4fff-bbc0-278f4ea6282d","order_by":10,"name":"Kaori Hayashi","email":"","orcid":"","institution":"Keio University School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Kaori","middleName":"","lastName":"Hayashi","suffix":""}],"badges":[],"createdAt":"2025-12-03 08:53:23","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8268064/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8268064/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1038/s41598-026-41005-4","type":"published","date":"2026-02-20T15:57:09+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":98433138,"identity":"88e5616d-2fec-479e-9b1d-74d8dbc270d9","added_by":"auto","created_at":"2025-12-17 16:50:20","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":83563,"visible":true,"origin":"","legend":"","description":"","filename":"Manuscript251206r.docx","url":"https://assets-eu.researchsquare.com/files/rs-8268064/v1/496d1626b1ec4c4025cec0a5.docx"},{"id":98249781,"identity":"8a0087e3-c980-4e06-80b5-7b4790fc4742","added_by":"auto","created_at":"2025-12-15 16:44:28","extension":"json","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":11269,"visible":true,"origin":"","legend":"","description":"","filename":"e5966cd8f2164e9abd52b223769cc33d.json","url":"https://assets-eu.researchsquare.com/files/rs-8268064/v1/97d44850281c28f261858012.json"},{"id":98434025,"identity":"1dba16ca-ee89-4463-ac2d-f185e6cc27f3","added_by":"auto","created_at":"2025-12-17 16:51:25","extension":"pptx","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":162931,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryFigures.pptx","url":"https://assets-eu.researchsquare.com/files/rs-8268064/v1/e12dc739b0bf344aa5655317.pptx"},{"id":98433694,"identity":"e70aa79a-e6fa-43af-9115-b6896ff1125c","added_by":"auto","created_at":"2025-12-17 16:51:02","extension":"docx","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":48818,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryMaterial.docx","url":"https://assets-eu.researchsquare.com/files/rs-8268064/v1/3e523ac50eaee4bd328821ae.docx"},{"id":98249786,"identity":"63f1ec7c-810c-4edb-91e1-0fa7af7c6d15","added_by":"auto","created_at":"2025-12-15 16:44:28","extension":"xlsx","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":66721,"visible":true,"origin":"","legend":"","description":"","filename":"TableData.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-8268064/v1/5c90de8f9f280c941fae6224.xlsx"},{"id":98433143,"identity":"3a329563-22e9-439d-87f2-e7b5ab0b896a","added_by":"auto","created_at":"2025-12-17 16:50:20","extension":"xml","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":111658,"visible":true,"origin":"","legend":"","description":"","filename":"e5966cd8f2164e9abd52b223769cc33d1enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-8268064/v1/f43f3c0e74cbfe060ea41207.xml"},{"id":98433722,"identity":"cee4795c-e114-491e-a009-18faef6d29ad","added_by":"auto","created_at":"2025-12-17 16:51:04","extension":"pptx","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1690064,"visible":true,"origin":"","legend":"","description":"","filename":"MainFigures.pptx","url":"https://assets-eu.researchsquare.com/files/rs-8268064/v1/0da7007c0f36c3cd8b418433.pptx"},{"id":98249790,"identity":"84435033-2203-4dd2-a718-f42641c0036c","added_by":"auto","created_at":"2025-12-15 16:44:28","extension":"xml","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":108463,"visible":true,"origin":"","legend":"","description":"","filename":"e5966cd8f2164e9abd52b223769cc33d1structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-8268064/v1/a4f32f1f20756bc7540ccfd1.xml"},{"id":98434994,"identity":"27e978ae-f35a-4f75-a5dd-1fb58b4f3ca1","added_by":"auto","created_at":"2025-12-17 16:52:53","extension":"html","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":124615,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-8268064/v1/b37553372f6ef3510ecfbdec.html"},{"id":98249776,"identity":"82770035-31ee-490d-a9af-c58395a6d981","added_by":"auto","created_at":"2025-12-15 16:44:28","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":141596,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eStatistical analysis of all 196 clinical features for PA subtype classification\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTwo-sample t-tests were performed on 196 features to obtain p-values as indicators for candidate predictors of PA subtype. Figures 1A, B, and Table 1B show p-values of all 196 clinical features, and Figure 1C and Table 1C present ROC-AUC and other performances for subtype classification of each individual feature when applied to all cases (n=274). Supplementary Table 1 presents details of all 196 features, including index IDs, abbreviations, full descriptions, and units of measurements. Supplementary Table 2 provides full data for all 196 features.\u003c/p\u003e\n\u003cp\u003eA) Top 40 features with lower p-values in differentiating bPA and uPA. The colors of the circles represent feature Categories: A (blue), B (red), C (green), and D (gold). The red line indicates a p-value of 0.05.\u003c/p\u003e\n\u003cp\u003eB) Top 10 features from each Category.\u003c/p\u003e\n\u003cp\u003eC) ROC curves for the lowest p-value feature of each Category: serum K (Category A), CCT90-PAC (B), MCV (C), and THAld (D).\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-8268064/v1/1683c51487f53ff6e0e2a70a.png"},{"id":98432916,"identity":"586e5ac2-2f37-494f-a2d8-511c4baf0698","added_by":"auto","created_at":"2025-12-17 16:50:07","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":174521,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ePredictive performance of machine learning models and contribution of feature categories\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo predict PA subtype, we developed five machine learning models. The models output a probability score (0–1) for uPA, and cases were classified as uPA if the score exceeded 0.5. Predictive performance was evaluated by LogLoss, ROC-AUC, accuracy, sensitivity, and specificity. Figures 2A and Table 2A show the predictive performance of each model constructed using all features and applied to the test dataset (n=69). Figures 2B and Table 2B present the predictive performance of the RF model trained on features from each category or combinations, highlighting the predictive utility of each category. Tables 2C and D presentthe predictive contribution of individual feature categories and subcategories in each model constructed using all 196 features.\u003c/p\u003e\n\u003cp\u003eA) ROC curves and dot plots of prediction scores of five machine learning-based models (LR, SVM, RF, GBDT, MLP). Dot plots indicate true/false predictions for bPA and uPA. The vertical axis represents probability scores. bPA cases with scores \u0026lt;0.5 and uPA cases with scores ≧0.5 were considered correctly classified (true predictions). The red line indicates the prediction threshold of 0.5. True predictions are indicated by dark-colored circles, whereas false predictions are shown by light-colored circles.\u003c/p\u003e\n\u003cp\u003eB) ROC curves in the RF models constructed using features from each category or in combination with Category A. Models incorporating Categories B and A improved ROC-AUC, whereas the addition of Category C decreased it.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-8268064/v1/059ae85b318409258ce2a63b.png"},{"id":98249777,"identity":"fb770f97-7bdd-49fd-95b9-1fa8466ac7b1","added_by":"auto","created_at":"2025-12-15 16:44:28","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":221131,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ePredictive contributions of features in the RF, LR, and MLP models\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003ePredictive contributions of individual features in the RF, LR, and MLP models were evaluated using model-appropriate importance metrics: regression coefficient (LR), feature importance (RF), and permutation importance (MLP). Figures 3A, B, and Table 3A show feature importance in the RF model. Figure 3C and Table 3B present regression coefficient in the LR model. Figure 3D and Table 3C present the permutation importance in the MLP model.\u003c/p\u003e\n\u003cp\u003eA) Top 40 features with the highest feature importance in the RF model. Bar colors represent feature Categories: A (blue), B (red), C (green), and D (gold). The top three features with the highest feature importance for subtype prediction with RF were CCT90-PAC (Category B, 0.212), tumor size (A, 0.145) and A/C ratio at 8:00 (B, 0.050).\u003c/p\u003e\n\u003cp\u003eB) Top 10 features from each category.\u003c/p\u003e\n\u003cp\u003eC) The nine predictive features based on regression coefficient in the LR model with Lasso regularization. All other features were assigned zero coefficients.\u003c/p\u003e\n\u003cp\u003eD) Top 40 most predictive features in the MLP model based on permutation importance.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-8268064/v1/4d51dd5e8b13f735e6972b66.png"},{"id":103252704,"identity":"3dd0a73f-b12d-4fb8-a655-7157d845dd19","added_by":"auto","created_at":"2026-02-23 16:15:43","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1597448,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8268064/v1/a3a8cf1e-8207-494a-9a26-41ce2d2b2097.pdf"},{"id":98434066,"identity":"9ea98450-a4cb-4005-a98f-ecddd8398797","added_by":"auto","created_at":"2025-12-17 16:51:28","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":48818,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSupplementary Information is available\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e1) Supplementary Material:\u003c/p\u003e\n\u003cp\u003eExtended Methods, Supplementary Tables 1 to 3, and Legends for Supplementary Figures\u003c/p\u003e","description":"","filename":"SupplementaryMaterial.docx","url":"https://assets-eu.researchsquare.com/files/rs-8268064/v1/a53f9ea1e14af04b40f64c19.docx"},{"id":98249782,"identity":"6f8fb6ee-9ded-4ca1-b330-fee6ac750d39","added_by":"auto","created_at":"2025-12-15 16:44:28","extension":"pptx","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":162931,"visible":true,"origin":"","legend":"\u003cp\u003e2) Supplementary Figures\u003c/p\u003e","description":"","filename":"SupplementaryFigures.pptx","url":"https://assets-eu.researchsquare.com/files/rs-8268064/v1/ceba9d57762aaf69e89e065d.pptx"},{"id":98433700,"identity":"9ec12e2a-345f-4495-9b62-905a0a2bb58d","added_by":"auto","created_at":"2025-12-17 16:51:03","extension":"xlsx","order_by":3,"title":"","display":"","copyAsset":false,"role":"supplement","size":66721,"visible":true,"origin":"","legend":"","description":"","filename":"TableData.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-8268064/v1/8c8e7d753da5b603046fbb26.xlsx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Machine Learning Approach for Subtype Prediction in Primary Aldosteronism: A Comprehensive Analysis of Models and Features","fulltext":[{"header":"Introduction","content":"\u003cp\u003ePrimary aldosteronism (PA) is a common cause of secondary hypertension, accounting for 5\u0026ndash;10% of hypertensive patients [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. It results from autonomous and excessive secretion of aldosterone from the adrenal glands and is associated with a higher risk of cardiovascular disease than essential hypertension (EHT) [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. There are two major subtypes of PA: unilateral PA (uPA), which is curable by adrenalectomy, and bilateral PA (bPA), which requires medical treatment [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. Given that treatment strategies differ by subtype, accurate diagnosis of uPA and bPA is critical for selecting the appropriate treatment strategy. Although adrenal venous sampling (AVS) remains the gold standard for subtype diagnosis of primary aldosteronism (PA), the invasive nature, technical difficulty and limited availability restrict clinical utility [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e, \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. In addition, PA is often underrecognized in primary care [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eTo address these challenges, several clinical algorithms have been developed to distinguish PA subtype without relying on AVS, using readily available clinical features. Machine learning models are considered suitable for accurate prediction, as they can integrate complex combinations of clinical features [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. However, the optimal models and feature sets remain unclear, and prior studies have shown limited precision and generalizability [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. Accordingly, a comprehensive assessment of suitable models and feature sets for PA prediction was warranted.\u003c/p\u003e\u003cp\u003eThis study examined five machine learning models and 196 clinical features useful for PA subtype prediction, under the hypothesis that machine learning approaches combined with comprehensive feature sets are beneficial for prediction. A total of 196 features were collected from 274 cases and classified into four categories: A, PA-related features; B, challenge tests and diurnal hormone sampling; C, general biochemistry and blood counts; and D, urinary steroid profile. The number of features in this study is unprecedented, and the inclusion of Categories C and D is especially notable. The comprehensive approach permitted exploration of previously unrecognized features that may have high predictive contribution.\u003c/p\u003e\u003cp\u003ePA subtype classification typically relies on PA-related features in Category A, such as plasma aldosterone concentration (PAC) and serum potassium (K) level. In addition, the use of challenge tests such as the saline infusion test (SIT), captopril challenge test (CCT), furosemide upright test (FUT), and rapid ACTH stimulation test (AST) has been reported [\u003cspan additionalcitationids=\"CR14 CR15 CR16\" citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. We incorporated features in Category B with the intention of enhancing predictive performance beyond that of Category A alone. To support early outpatient prediction in primary care, we included Category C. Urinary steroid profile features (Category D) were also added because of reported utility in subtype classification [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eAccordingly, this study evaluated predictive performance of five machine learning models and predictive contribution of 196 individual features and four categories for PA subtype prediction under the hypothesis that comprehensive feature sets and machine learning approaches are beneficial for investigating optimal models for predicting PA subtype without relying on AVS.\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003e\u003cstrong\u003eOverview of the study flow in accordance with the TRIPOD statement.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study investigated machine learning models, features, and feature categories that are useful for subtype prediction in PA. All the methodological steps, including study design, participant selection, feature handling, modeling procedures, and internal validation, were predefined to ensure clarity, reproducibility, and transparency in accordance with the TRIPOD (Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis) statement. Predictive performance of the machine learning models, predictive contribution of individual features, and that of feature categories were evaluated.\u003c/p\u003e\n\u003cp\u003eThe inclusion process for the 274 PA patients is shown in Supplementary Figure 1. Study participants were identified from 526 consecutive PA patients admitted to a single institution (Keio University Hospital) in Japan between January 2013 and December 2020. PA was diagnosed in patients who met the ARR screening criteria and had a positive CCT result following Japanese or U.S. guidelines [11,12]. Among them, 411 underwent successful AVS, and the subtype was determined according to the AVS result. A total of 274 patients who completed the endocrine evaluation, including challenge tests and urinary steroid profiling, were included. This sample size is among the largest reported from a single center in Japan. A comprehensive feature collection strategy was adopted, yielding 196 predictors, including secondary variables such as ratios. This unprecedented number of features enabled broad and exploratory analyses with minimal selection bias. Missing data were handled by stratifying patients into eight groups on the basis of quartiles of age and sex. Missing values were imputed by the mean of the corresponding group. The 196 features were grouped into four categories, and their predictive contributions were analyzed.\u003c/p\u003e\n\u003cp\u003eTo examine useful models, five machine learning algorithms were evaluated: logistic regression (LR), support vector machine (SVM), random forest (RF), gradient boosting decision tree (GBDT), and multilayer perceptron (MLP). The 274 cases were randomly divided into a training dataset (n=205) and a test dataset (n=69). All models were trained on the training dataset, and hyperparameters were optimized via a grid search. Predictive performance of the models was analyzed on the test dataset using logarithmic loss (LogLoss), area under the receiver operating characteristic curve (ROC-AUC), accuracy, sensitivity, and specificity. To explore useful features, predictive contribution of individual features and feature categories was evaluated using model-appropriate importance metrics: regression coefficient for LR, feature importance for RF, and permutation importance for SVM, GBDT, and MLP.\u003c/p\u003e\n\u003cp\u003eSeveral models for PA subtype prediction have been reported. In the present study, potentially more appropriate machine learning approaches and predictive features were explored as comprehensively as possible. The findings may support the development of predictive models that inform treatment selection with reduced reliance on AVS. Major limitations include the absence of residual analyses for misclassified cases and the lack of external validation, both of which warrant further investigation.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAssessment of patients and collection of features\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe complete inpatient endocrine evaluation included challenge tests (CCT, FUT, AST), diurnal hormone sampling (8:00, 16:00, 24:00), 24-hour urine collection and AVS. uPA was defined as lateralized ratio ≧4.0 or lateralized ratio ≧2.5 with contralateral ratio \u0026lt;1 after ACTH stimulation. Urinary steroid profile (Category D) of 63 metabolites was analyzed, as shown in Supplementary Figure 2.\u003c/p\u003e\n\u003cp\u003eA total of 196 features were used to develop prediction models. Details of the features (abbreviations, full descriptions, units of measurements and index numbers) are provided in Supplementary Table 1. Medians, interquartile ranges, and p-values for all 196 features are shown in Supplementary Table 2. The 196 features were grouped into four categories and 21 subcategories (Table 1). Briefly, PA-related features were assigned to Category A (A1–A23), challenge tests to Category B (B1–B47), general biochemistry data to Category C (C1–C37), and urinary steroid profile to Category D (D1–D89). This grouping was expected to reduce multicollinearity among features and thereby enhance the reliability and interpretability of the models. Features in Categories A and C were collected at the initial outpatient visit, while those in Categories B and D were measured during hospitalization. Antihypertensive medications were not restricted at initial visit. Before hospitalization, all antihypertensive agents except calcium channel blockers and α-blockers were discontinued for at least two weeks to avoid affecting PAC, ARC, and\u0026nbsp;aldosterone renin ratio\u0026nbsp;(ARR).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDevelopment of machine learning-based prediction models\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003ePrediction models for PA subtype were constructed using five machine learning algorithms: LR, SVM, RF, GBDT, and MLP with Python 3.11, scikit-learn 1.5.0, and LightGBM 4.4.0 [20]. LR and SVM are regression models, RF and GBDT are decision tree models, and MLP is a deep learning model. All the models were trained on the training dataset, and the hyperparameters were optimized via a grid search using four-fold cross-validation [21]. Following model development, the models were used to predict PA subtype in the test dataset. The models generated a prediction score (0–1) of uPA for each case, and cases with scores≧0.5 were classified as uPA.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ePredictive performance of machine learning models and predictive contribution of features and categories\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo examine the models and features useful for predicting PA subtype, predictive performance of five machine learning models and predictive contributions of 196 features and four feature categories were analyzed. Predictive performance was estimated on the test dataset (n=69) using the following standard metrics: logarithmic loss (LogLoss), area under the receiver operating characteristic curve (ROC-AUC), accuracy, sensitivity, and specificity. Models that achieved higher ROC-AUC values or accuracies were considered clinically useful.\u003c/p\u003e\n\u003cp\u003ePredictive contribution of features was assessed using model-appropriate importance metrics: regression coefficient for LR, feature importance for RF, and permutation importance for SVM, GBDT, and MLP; all metrics were normalized so that the sum of 196 features equaled 1. Predictive contribution of categories and subcategories was calculated by summing the metrics of the respective features. The features were sorted by the importance metrics in descending order and presented as indicators of the predictive contribution.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthical statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll methods were carried out in accordance with relevant guidelines and regulations, including the Ethical Guidelines for Medical and Health Research Involving Human Subjects in Japan. The study protocol was reviewed and approved by the Ethics Committee of Keio University School of Medicine (Approval No. 20180135). Informed consent was obtained from all participants who were able to provide consent, after oral explanation using a written information sheet. In addition, an opt-out approach was implemented on the institutional website in accordance with institutional and national regulations.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003e\u003cstrong\u003eStatistical analysis of features between bPA and uPA\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo explore features relevant to PA subtype prediction, we conducted statistical comparisons between the bPA and uPA groups using 196 features from a single-center cohort of 274 patients. P-values were calculated for each feature, and the top 40 features with the lowest p-values were visualized (Figure 1A). The top 10 features from each of the four Categories (A\u0026ndash;D) are shown separately (Figure 1B), with a detailed list of the results (Table 1B). The three most statistically significant features were plasma aldosterone concentration at 90 minutes after captopril challenge (CCT90-PAC; Category B, p = 1.74\u0026times;10⁻\u0026sup1;⁹), PAC at 8:00 (Category B, p = 3.55\u0026times;10⁻\u0026sup1;⁶), and tetrahydro-aldosterone (THAld) (Category D, p = 6.55\u0026times;10⁻\u0026sup1;⁶). Among the top 40 features, 20 belonged to Category B (47 total), 12 to Category D (89 total), and 8 to Category A (23 total); no features from Category C (37 total) were included.\u003c/p\u003e\n\u003cp\u003eRepresentative features with the lowest p-values in each category were as follows: serum K, 24-hour urinary aldosterone, and tumor size (Category A); PAC after CCT or AST, and at 8:00 (Category B); mean corpuscular volume (MCV), mean corpuscular hemoglobin concentration (MCHC), and creatinine (Category C); and THAld and 18-hydroxy-tetrahydro-11-dehydrocorticosterone (18OHTHA) (Category D). The utility of the feature with the lowest p-value in each category (K, CCT90-PAC, MCV, and THAld) for PA subtype classification was analyzed. As shown in Figure 1C and Table 1C, CCT90-PAC had the highest ROC-AUC (0.836) and classification accuracy (0.818). Figure 1C and Table 1C present the diagnostic performance when each feature was applied to all included cases (n=274), whereas Figures 2 and 3 show predictive performance of machine learning models with multiple features evaluated on the test dataset (n=69). Given that individual features alone might lack diagnostic accuracy, machine learning models using multiple features were expected to improve the performance of subtype prediction.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ePredictive performance of machine learning models\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eBased on the hypothesis that the machine learning approach is beneficial for prediction using a large number of features, we developed subtype prediction models using five algorithms: LR, SVM, RF, GBDT, and MLP\u0026nbsp;algorithms. The prediction performance of each model was evaluated by five metrics: LogLoss, ROC-AUC, accuracy, sensitivity, and specificity. The total dataset (n=274) was\u0026nbsp;pre-divided\u0026nbsp;into a training dataset (n=205; 75%) and a test dataset (n=69; 25%). Predictive models were developed through training on the training dataset. The classification accuracy of PA subtype within the internal training dataset (n=205) and the hyperparameters\u0026nbsp;are\u0026nbsp;shown in Supplementary Table 3A. Following model development, the models were used for predicting PA subtype in the test dataset (n=69). Dot plots and performance metrics for the prediction are shown in Figure 2A and Table 2A.\u0026nbsp;The\u0026nbsp;ROC-AUC scores were 0.920 (LR), 0.932 (SVM), 0.972 (RF), 0.929 (GBDT), and 0.926 (MLP). Predictive accuracy followed a similar trend: 0.797 (LR), 0.870 (SVM), 0.913 (RF), 0.884 (GBDT), and 0.884 (MLP). RF demonstrated the best overall performance, indicating the strength of decision tree-based approaches for subtype prediction.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ePredictive contribution of categories and subcategories\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAmong the five machine learning models, the RF model achieved the highest predictive accuracy (0.913). To evaluate the utility of each feature category for subtype prediction,\u0026nbsp;we constructed RF models using features from each category (A, B, C, or D) and from combinations of Category A with other categories (A\u0026amp;B, A\u0026amp;C, A\u0026amp;D), in addition to those from all categories.\u0026nbsp;Hyperparameters\u0026nbsp;were optimized for each configuration (Supplementary Table 3B). Performance metrics for\u0026nbsp;the RF models\u0026nbsp;and the ROC curves, which were constructed using single-category or the combinations, are shown in Table 2B and Figure 2B.\u0026nbsp;The accuracy\u0026nbsp;was highest for Category B, followed by\u0026nbsp;Categories\u0026nbsp;A, D, and C. Combining\u0026nbsp;Categories\u0026nbsp;B and A improved predictive performance in the test set,\u0026nbsp;whereas\u0026nbsp;adding Category C reduced\u0026nbsp;the\u0026nbsp;accuracy. Although the A\u0026amp;D combination achieved an excellent classification accuracy of 0.995 in the training dataset, predictive accuracy in the test dataset was 0.884 (Table 2B), suggesting that adding more features did not necessarily enhance the predictive performance of the RF model.\u003c/p\u003e\n\u003cp\u003ePredictive contribution of each category, which is based on model-appropriate importance metrics, that is, regression coefficient (LR), feature importance (RF), and permutation importance (SVM, GBDT, and MLP), is presented in Table 2C. The categories that most significantly contributed to PA subtype prediction in each predictive model were Category A in LR (regression coefficient 0.611), Category B in SVM, RF, and GBDT (RF feature importance 0.475), and Category D in MLP (permutation importance 0.411). Subcategory-level contributions are summarized in Table 2D. The subcategories that contributed the most were tumor size (Subcategory Af) in LR and SVM, CCT (Bc) in RF, and urinary steroid metabolites (Da) in GBDT and MLP. Among the subcategories of challenge tests within Category B, those that showed the highest contribution to PA subtype prediction were in the order of CCT, AST, and FUT, in LR, RF, and GBDT; whereas AST was the highest in SVM and MLP.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ePredictive contribution of features\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo identify contribution of individual features to PA subtype prediction, we evaluated the importance metrics, namely, the regression coefficient (LR), feature importance (RF), and permutation importance (MLP), of individual features as indices of predictive contribution. First, we analyzed feature importance in the best-performing RF model. The top 40 features are visualized in Figure 3A, and the top 10 features in each category are shown in Figure 3B. A complete list of the top 40 features is shown in Table 3A. The top three predictors in the RF model were CCT90-PAC (Category B, 0.212), tumor size (Category A, 0.145), and aldosterone/cortisol ratio (A/C ratio) at 8:00 (Category B, 0.050). Some Category D features also contributed substantially. Overall, features with small p-values (Table 1B) still contributed to the prediction of PA subtype at least in the RF model. This accounts for the similarity in feature lists between Tables 1B and 3A. In addition, features such as tumor size were more important in the RF model, likely because they provided unique information not captured by other variables. A full list of feature importance values of all 196 features is provided in Supplementary Table 2.\u003c/p\u003e\n\u003cp\u003ePredictive contribution to the LR model (regression coefficient) is shown in Figure 3C and Table 3B, and that to the MLP model (permutation importance) in Figure 3D and Table 3C. The LR model automatically selected and used 9 features, due to regularization to prevent feature redundancy, most of which overlapped with\u0026nbsp;the\u0026nbsp;low p-value features in Table 1B. In contrast,\u0026nbsp;the\u0026nbsp;MLP model showed relatively uniform permutation importance across features\u0026nbsp;in\u0026nbsp;the presence of multicollinearity among similar features, and even those with relatively high p-values\u0026nbsp;still demonstrated substantial predictive\u0026nbsp;contributions. These findings suggest that while LR and RF emphasize statistically significant features, MLP assigned relatively uniform contribution to each feature.\u003cstrong\u003e\u003cbr\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eOn the basis of the hypothesis that machine learning models using a comprehensive feature set are beneficial for PA subtype prediction, this study investigated the predictive performance of five models (based on LR, SVM, RF, GBDT, and MLP) and the predictive contribution of 196 features across four categories. Decision tree models (RF, GBDT) and the deep learning model (MLP) outperformed regression models (LR, SVM), with RF showing the highest performance. Category B (challenge tests and diurnal hormone sampling) was identified as the category that contributed the most to PA subtype prediction in the RF model, whereas Category C (general biochemistry) was found to lack predictive value and may adversely affect model performance. Category D (steroid profile) played a significant role in the MLP model. Although the contributions of individual features varied across models, CCT90-PAC was a useful feature in RF. The academic significance of this study lies in 1) evaluation of the predictive performance of multiple machine learning models, 2) collection of a comprehensive set of 196 clinical features to explore unknown predictive features, and 3) systematic categorization of features to quantify category- and subcategory-level predictive contributions.\u003c/p\u003e\n\u003cp\u003eTo our knowledge, this is the first study to evaluate machine learning models using a broad feature set that includes challenge tests and diurnal hormonal sampling (Category B), general biochemical parameters and blood counts (Category C), and urinary steroid profile (Category D), along with\u0026nbsp;PA-related features (Category A). Previous studies have constructed models to differentiate PA subtype, reporting sensitivity of 32–95% and specificity of 46–100% [22–27], including studies using machine learning-based prediction of PA subtype [28–29]. Kaneko et al. applied four machine learning methods (LR, SVM, RF, GBDT) to 229 PA cases, using 80% for training and 20% for testing, and reported that RF achieved the highest predictive performance (ROC-AUC 0.990, accuracy 95.7%). Regarding features that are useful for prediction, the utility of CCT and SIT has been documented [30]. Eisenhofer et al. reported a model using urinary steroid profile that achieved 100% sensitivity and 98% specificity for diagnosing uPA with KCNJ5 variants, whereas the ROC-AUC was 0.716 in wild-type patients [19].\u003c/p\u003e\n\u003cp\u003eThis study originally aimed to explore features that are useful for PA subtype prediction through a comprehensive feature collection approach. Machine learning-based models were introduced because they can handle extensive feature sets effectively. To address multicollinearity arising from the large number of features, we organized potentially redundant features into categories and subcategories and evaluated the relative contributions of each category. Category A features, including tumor size, K, PAC and ARR, have conventionally been identified as essential for subtype classification [31–36]. PA is more strongly associated with cardiovascular disease than essential hypertension. In particular, uPA tends to be more severe than bPA and is characterized by higher PAC and lower serum K [37–39]. These findings underscore the importance of Category A features in subtype . Among other biochemical markers, Na, Ca, intact PTH, and BNP also reflected aldosterone’s impact on electrolyte balance and contributed to subtype prediction in the RF model, which is consistent with prior reports.\u003c/p\u003e\n\u003cp\u003eIn Category B, we assessed CCT, AST, and FUT; however, SIT was not included, as it was not routinely performed at our facility. In this study, CCT and AST emerged as key predictors, which is consistent with previously reported findings for subtype prediction [40,41]. Notably, CCT90-PAC was the top contributor in the RF model, surpassing all Category A features. The superiority of CCT90-PAC over baseline PAC aligns with the well-recognized fact that PAC is often insufficiently suppressed by captopril in patients with APA. For CCT, absolute PAC values were more predictive than ARR or PAC fold increases were. Moreover, AST120 A/C ratio ranked higher than AST120-PAC.\u003c/p\u003e\n\u003cp\u003eCategory C included general clinical parameters such as liver and renal function, blood counts, and markers of glucose and lipid metabolism, excluding electrolytes in Category A. Category C features showed limited predictive utility in the RF model, and adding them to Category A reduced performance, indicating that Category C introduced noise and attenuated the predictive contribution of Category A. Although lower cholesterol and triglycerides in uPA have been reported [42], only urinary C-peptide emerged as a notable contributor in the present RF model. Subcategory analysis indicated minimal added value from blood counts, liver function, and metabolic markers.\u003c/p\u003e\n\u003cp\u003eCategory D comprises inpatient urinary steroid profiles. In previous reports, hybrid steroids which possess both glucocorticoid- and mineralocorticoid-like structures (17α-hydroxyl and 18-hydroxyl or -oxo groups; 18OHF, etc.) have been reported to associate with uPA and may contribute to subtype classification [43–47]. In this study, 18OHTHA, a urinary metabolite of an intermediate in aldosterone biosynthesis, was the most contributory Category D feature in the RF model, surpassing urinary aldosterone or other metabolites, including 18OHF. Elevation of 18OHTHA may suggest increased CYP11B2 (aldosterone synthase) activity in the adrenal cortex. In addition to individual metabolites, we included metabolite sums and ratios reflecting enzyme activity. Among these ratios, the 18OHF/(THE+THF) ratio had the highest predictive contribution. These findings support the utility of urinary steroid metabolites, particularly 18OHTHA and 18OHF, as predictors of PA subtype.\u003c/p\u003e\n\u003cp\u003eFeature category and subcategory-level contributions to subtype prediction were evaluated. In the RF model, Category B contributed the most, followed by Categories A, D, and C. Adding Category B to Category A improved predictive performance, and inclusion of all categories yielded further gains. Category C showed limited predictive contribution. Subcategory analysis within Category A indicated that tumor size was the largest contributor. The usefulness of the challenge test features within Category B for subtype classification has been reported in previous studies. However, it remains unclear which test (CCT, FUT or AST) is useful. In this study, parameters from each test were subcategorized to facilitate comparisons. In the RF model, CCT and AST showed greater predictive contribution than FUT. Within Category D, several individual metabolites demonstrated high contributions, whereas metabolite sums and ratios contributed less, possibly reflecting the inclusion of non-informative components.\u003c/p\u003e\n\u003cp\u003eWith respect to effective machine learning models, predictive accuracy on the test dataset ranged from 0.797 to 0.913, indicating relatively high performance. Decision tree models (RF, GBDT) and MLP outperformed regression-based models (LR, SVM). The LR model, limited by its linear nature, could not capture feature interactions; however, it offered high interpretability, emphasizing significant features such as serum K, CCT90-PAC, and tumor size. Lasso regularization reduced the LR model to nine features, most of which overlapped with variables showing low p-values in Table 1B. The relatively poorer performance of LR may partly reflect overreliance on the limited features. The nonlinear regression model, SVM, outperformed LR and achieved 100% specificity for uPA prediction, although the sensitivity was low. Decision-tree models capture feature interactions effectively, resulting in higher accuracy. The RF model exceeded 90% accuracy and integrated steroid-related features effectively to enhance prediction. While GBDT often achieves greater performance than RF does, the limited data in this study may have led to overfitting.\u003c/p\u003e\n\u003cp\u003eThe deep learning model MLP achieved high predictive accuracy, however, the contributory features differed substantially from those in the other algorithms. Several features with limited contributions in other models (e.g., MCHC) were utilized more effectively in MLP, likely because it can leverage complex interactions among variables. Nevertheless, permutation importance values in MLP may be configuration-dependent under multicollinearity. Interpretation should focus on broader feature categories rather than individual features. Consequently, predictive contributions of features in MLP were relatively uniform, individual feature-level contributions were small, and category level contributions were proportional to the number of features in each category. Features from urinary steroid profile (Category D) were effectively utilized in MLP, with several individual metabolites showing higher contributions. Notably, despite the higher accuracy achieved by MLP, it exhibited distinct outliers in both bPA and uPA predictions (dot plots in Figure 2A), indicating a tendency toward confident misclassification. False positive predictions in MLP may reflect clinically meaningful uPA-like features, as\u0026nbsp;observed\u0026nbsp;in bilateral\u0026nbsp;APAs. Moreover, the\u0026nbsp;sensitivity for uPA remained below 0.765 across models, suggesting that some cases of AVS-confirmed uPA may exhibit bPA-like features.\u0026nbsp;In\u0026nbsp;contrast,\u0026nbsp;the\u0026nbsp;specificity was high: 0.904 (LR), 0.923 (MLP), 0.962 (RF), 0.962 (GBDT), and 1.000 (SVM). Thus, despite similar overall\u0026nbsp;accuracies, the\u0026nbsp;models differed in sensitivity–specificity profiles, suggesting model-specific clinical implications.\u003c/p\u003e\n\u003cp\u003eSeveral limitations should be noted. First, the sample size (n=274) might be insufficient for reliable development of machine learning models, although it represents one of the largest single-center AVS cohorts in Japan. Second, multicollinearity among redundant features was present and might distort estimates of predictive contribution. Third, the contribution of features differed among models, complicating identification of consistently useful predictors. Fourth, the saline infusion test (SIT) was not performed, limiting direct comparisons among challenge tests. Finally, urinary steroid profile is not widely available and AVS criteria for subtype classification are not fully standardized [48]; thus, the results may vary in other settings. The aim of this study was not to present a single best model or a fixed ranking of predictors, but to compare the performance of five machine learning algorithms and assess the relative usefulness of features and predefined feature categories. Despite these limitations, the study has notable strengths: 1) five distinct machine learning models were examined and the potential utility of RF for predicting PA subtype was demonstrated; 2) a broad set of 196 clinical features including urinary steroid metabolites was collected, with CCT90-PAC identified as a useful predictor in the RF model; and 3) feature categories and subcategories were defined to reduce the impact of multicollinearity among features and improve interpretability, thereby clarifying the utility of challenge tests (Category B) for subtype prediction.\u003c/p\u003e\n\u003cp\u003eIn summary, machine learning-based models using 196 features could effectively predict PA subtype. The RF model achieved ≧90% predictive accuracy on the test dataset, with Category B,\u0026nbsp;which includes challenge tests,\u0026nbsp;contributing most to the prediction. These findings suggest that predictive models based primarily on Category B features can achieve superior predictive performance with a small feature set, supporting clinically feasible treatment selection with reduced reliance on AVS. External validation and misclassification analysis remain essential to refine models and achieve reliable PA subtype prediction.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAs host members of ISH 2022 Kyoto (International Society of Hypertension Scientific Meeting, Kyoto, Japan), the authors gratefully acknowledge the meeting participants for their academic insights that informed the conception of this study.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eY.M., K.M., To.N. and J.N. conceived the study, designed the analysis and data collection, and drafted the manuscript Ke.H., Te.N. and H.M. contributed to data curation. M.N., K.K., H.I. and Ka.H. critically reviewed the manuscript and provided important interpretation of the data. All authors reviewed and approved the final version of the manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData availability statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe datasets are available from the corresponding author on reasonable request.\u003c/p\u003e\n\u003cp\u003eThe Python code used in this study is publicly available on GitHub.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAdditional Information\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCOI statement:\u0026nbsp;\u003c/strong\u003eThe authors declare no conflicts of interest.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eSources of Funding:\u003c/strong\u003e This study was conducted as a voluntary activity of authors with no external funding.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eYoung WF. Primary aldosteronism: renaissance of a syndrome. Clin Endocrinol (Oxf). 2007;66:607\u0026ndash;618.\u003c/li\u003e\n\u003cli\u003eMonticone S, et al. Prevalence and clinical manifestations of primary aldosteronism encountered in primary care practice. J Am Coll Cardiol. 2017;69:1811\u0026ndash;1820.\u003c/li\u003e\n\u003cli\u003eOhno Y; Nagahama Study / JPAS Study Group. Prevalence of cardiovascular disease and its risk factors in primary aldosteronism: a multicenter study in Japan. Hypertension. 2018;71:530\u0026ndash;537.\u003c/li\u003e\n\u003cli\u003eFunder JW, et al. The management of primary aldosteronism: case detection, diagnosis, and treatment. J Clin Endocrinol Metab. 2016;101:1889\u0026ndash;1916.\u003c/li\u003e\n\u003cli\u003eVonend O, et al. Adrenal venous sampling: evaluation of the German Conn\u0026rsquo;s registry. Hypertension. 2011;57:990\u0026ndash;995.\u003c/li\u003e\n\u003cli\u003eMonticone S, et al. Clinical management and outcomes of adrenal hemorrhage following adrenal vein sampling in primary aldosteronism. Hypertension. 2016;67:146\u0026ndash;152.\u003c/li\u003e\n\u003cli\u003eMulatero P, et al. Guidelines for primary aldosteronism: uptake by primary care physicians in Europe. J Hypertens. 2016;34:2253\u0026ndash;2257.\u003c/li\u003e\n\u003cli\u003eBaştanlar Y, \u0026Ouml;zuysal M. Introduction to machine learning. In: Yousef M, Allmer J, eds. Methods Mol Biol. 2014;1107:105\u0026ndash;128.\u003c/li\u003e\n\u003cli\u003eCruz JA, Wishart DS. Applications of machine learning in cancer prediction and prognosis. Cancer Inform. 2007;2:59\u0026ndash;77.\u003c/li\u003e\n\u003cli\u003eSam D, et al. External validation of clinical prediction models in unilateral primary aldosteronism. Am J Hypertens. 2022;35:365\u0026ndash;373.\u003c/li\u003e\n\u003cli\u003eNishikawa T, et al. Guidelines for the diagnosis and treatment of primary aldosteronism \u0026mdash; The Japan Endocrine Society 2009. Endocr J. 2009;58:711\u0026ndash;721.\u003c/li\u003e\n\u003cli\u003eYoung WF Jr, et al. Role for adrenal venous sampling in primary aldosteronism. Surgery. 2004;136:1227\u0026ndash;1235.\u003c/li\u003e\n\u003cli\u003eK\u0026uuml;pers EM, et al. A clinical prediction score to diagnose unilateral primary aldosteronism. J Clin Endocrinol Metab. 2012;97:3530\u0026ndash;3537.\u003c/li\u003e\n\u003cli\u003eYing Z, et al. Identifying unilateral disease in Chinese patients with primary aldosteronism by using a modified prediction score. J Hypertens. 2017;35:2486\u0026ndash;2492.\u003c/li\u003e\n\u003cli\u003eNanba K, et al. Shortened saline infusion test for subtype prediction in primary aldosteronism. Endocrine. 2015;50:802\u0026ndash;806.\u003c/li\u003e\n\u003cli\u003eRossi GP, et al. Comparison of the captopril and the saline infusion test for excluding aldosterone-producing adenoma. Hypertension. 2007;50:424\u0026ndash;431.\u003c/li\u003e\n\u003cli\u003eSonoyama T, et al. Significance of adrenocorticotropin stimulation test in the diagnosis of aldosterone-producing adenoma. J Clin Endocrinol Metab. 2011;96:2771\u0026ndash;2778.\u003c/li\u003e\n\u003cli\u003eEisenhofer G, et al. Mass spectrometry\u0026ndash;based adrenal and peripheral venous steroid profiling for subtyping primary aldosteronism. Clin Chem. 2016;62:514\u0026ndash;524.\u003c/li\u003e\n\u003cli\u003eEisenhofer G, et al. Use of steroid profiling combined with machine learning for identification and subtype classification in primary aldosteronism. JAMA Netw Open. 2020;3:e2016209.\u003c/li\u003e\n\u003cli\u003ePedregosa F, et al. Scikit-learn: machine learning in Python. J Mach Learn Res. 2011;12:2825\u0026ndash;2830.\u003c/li\u003e\n\u003cli\u003eBergstra J, Bengio Y. Random search for hyper-parameter optimization. J Mach Learn Res. 2012;13:281\u0026ndash;305.\u003c/li\u003e\n\u003cli\u003eKobayashi H, et al. Scoring system for the diagnosis of bilateral primary aldosteronism in the outpatient setting before adrenal venous sampling. Clin Endocrinol (Oxf). 2017;86:467\u0026ndash;472.\u003c/li\u003e\n\u003cli\u003eNanba K, et al. A subtype prediction score for primary aldosteronism. J Hum Hypertens. 2014;28:716\u0026ndash;720.\u003c/li\u003e\n\u003cli\u003eBuffolo F, et al. Subtype diagnosis of primary aldosteronism: is adrenal vein sampling always necessary? Int J Mol Sci. 2017;18:848.\u003c/li\u003e\n\u003cli\u003eChen S, et al. Computed tomography combined with confirmatory tests for the diagnosis of aldosterone-producing adenoma. Endocr J. 2021;68:299\u0026ndash;306.\u003c/li\u003e\n\u003cli\u003eHe K, et al. A clinical radiomic nomogram based on unenhanced computed tomography for predicting the risk of aldosterone-producing adenoma. Front Oncol. 2021;11:634879.\u003c/li\u003e\n\u003cli\u003eBurrello J, et al. Development and validation of prediction models for subtype diagnosis of patients with primary aldosteronism. J Clin Endocrinol Metab. 2020;105:e3706\u0026ndash;e3717.\u003c/li\u003e\n\u003cli\u003eKaneko H, et al. Machine learning based models for prediction of subtype diagnosis of primary aldosteronism using blood test. Sci Rep. 2021;11:9140.\u003c/li\u003e\n\u003cli\u003eTamaru S, et al. Machine learning approach to predict subtypes of primary aldosteronism is helpful to estimate indication of adrenal vein sampling. High Blood Press Cardiovasc Prev. 2022;29:375\u0026ndash;383.\u003c/li\u003e\n\u003cli\u003eByun JM, et al. A case of primary aldosteronism presenting as non\u0026ndash;ST elevation myocardial infarction. Korean J Intern Med. 2013;28:739\u0026ndash;742.\u003c/li\u003e\n\u003cli\u003eZhang Y, et al. Is primary aldosteronism a potential risk factor for aortic dissection? A case report and literature review. BMC Endocr Disord. 2020;20:115.\u003c/li\u003e\n\u003cli\u003eFava C, et al. Unusual presentation of primary aldosteronism with advanced target organ damage: a case report. Radiol Case Rep. 2019;14:814\u0026ndash;818.\u003c/li\u003e\n\u003cli\u003eKamemura K, et al. Significance of adrenal computed tomography in predicting laterality and indicating adrenal vein sampling in primary aldosteronism. J Hum Hypertens. 2017;31:195\u0026ndash;199.\u003c/li\u003e\n\u003cli\u003eKobayashi H, et al.; JPAS Study Group. Development and validation of subtype prediction scores for the workup of primary aldosteronism. J Hypertens. 2018;36:2269\u0026ndash;2276.\u003c/li\u003e\n\u003cli\u003eKocjan T, et al. A new clinical prediction criterion accurately determines a subset of patients with bilateral primary aldosteronism before adrenal venous sampling. Endocr Pract. 2016;22:587\u0026ndash;594.\u003c/li\u003e\n\u003cli\u003eUmakoshi H, et al. Significance of computed tomography and serum potassium in predicting subtype diagnosis of primary aldosteronism. J Clin Endocrinol Metab. 2018;103:900\u0026ndash;908.\u003c/li\u003e\n\u003cli\u003eRossi GP, et al. Hyperparathyroidism can be useful in the identification of primary aldosteronism due to aldosterone-producing adenoma. Hypertension. 2012;60:431\u0026ndash;436.\u003c/li\u003e\n\u003cli\u003eKato J, et al. Atrial and brain natriuretic peptides as markers of cardiac load and volume retention in primary aldosteronism. Am J Hypertens. 2005;18:354\u0026ndash;357.\u003c/li\u003e\n\u003cli\u003eJakubik P, et al. Impact of essential hypertension and primary aldosteronism on plasma brain natriuretic peptide concentration. Blood Press. 2006;15:302\u0026ndash;307.\u003c/li\u003e\n\u003cli\u003eXiao M, et al. Evaluation of the saline infusion test and the captopril challenge test in Chinese patients with primary aldosteronism. J Clin Endocrinol Metab. 2018;103:853\u0026ndash;860.\u003c/li\u003e\n\u003cli\u003eMoriya A, et al. ACTH stimulation test and computed tomography are useful for differentiating the subtype of primary aldosteronism. Endocr J. 2017;64:65\u0026ndash;73.\u003c/li\u003e\n\u003cli\u003eOhno Y; JPAS Study Group. Obesity as a key factor underlying idiopathic hyperaldosteronism. J Clin Endocrinol Metab. 2018;103:4456\u0026ndash;4464.\u003c/li\u003e\n\u003cli\u003eZhou Y, et al. Diagnostic accuracy of adrenal imaging for subtype diagnosis in primary aldosteronism: systematic review and meta-analysis. BMJ Open. 2020;10:e038489.\u003c/li\u003e\n\u003cli\u003eMulatero P, et al. 18-hydroxycorticosterone, 18-hydroxycortisol, and 18-oxocortisol in the diagnosis of primary aldosteronism and its subtype. J Clin Endocrinol Metab. 2012;97:881\u0026ndash;889.\u003c/li\u003e\n\u003cli\u003eSatoh F, et al. Measurement of peripheral plasma 18-oxocortisol can discriminate unilateral adenoma from bilateral disease in patients with primary aldosteronism. Hypertension. 2015;65:1096\u0026ndash;1102.\u003c/li\u003e\n\u003cli\u003eLenders JWM, et al. Diagnosis of endocrine disease: 18-oxocortisol and 18-hydroxycortisol \u0026mdash; is there clinical utility of these steroids? Eur J Endocrinol. 2018;178:517\u0026ndash;563.\u003c/li\u003e\n\u003cli\u003eMonticone S, et al. Immunohistochemical, genetic and clinical characterization of sporadic aldosterone-producing adenomas. Mol Cell Endocrinol. 2015;411:146\u0026ndash;154.\u003c/li\u003e\n\u003cli\u003eRossi GP, et al. An expert consensus statement on use of adrenal vein sampling for the subtyping of primary aldosteronism. Hypertension. 2014;63:151\u0026ndash;160.\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"Table","content":"\u003cp\u003e\u003cstrong\u003eTable 1.\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003eOverview of four Categories and 21 Subcategories covering the 196 features\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" width=\"631\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"2\" style=\"width: 83.3861%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eCategory A: PA-related features in the outpatient setting\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 16.6139%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e23 features\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 7.75316%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAa\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 75.6329%;\"\u003e\n \u003cp\u003e\u003cstrong\u003egeneral status\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 16.6139%;\"\u003e\n \u003cp\u003e5\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 7.75316%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAb\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 75.6329%;\"\u003e\n \u003cp\u003e\u003cstrong\u003ecirculatory system\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 16.6139%;\"\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 7.75316%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAc\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 75.6329%;\"\u003e\n \u003cp\u003e\u003cstrong\u003erenin-aldosterone system\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 16.6139%;\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 7.75316%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAd\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 75.6329%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eElectrolytes\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 16.6139%;\"\u003e\n \u003cp\u003e6\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 7.75316%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAe\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 75.6329%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e24h urine\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 16.6139%;\"\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 7.75316%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAf\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 75.6329%;\"\u003e\n \u003cp\u003e\u003cstrong\u003etumor size\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 16.6139%;\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"2\" style=\"width: 83.3861%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eCategory B: Challenge tests and diurnal profiles in the inpatient setting\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 16.6139%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e47 features\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 7.75316%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eBa\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 75.6329%;\"\u003e\n \u003cp\u003e\u003cstrong\u003edaily profile of aldosterone\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 16.6139%;\"\u003e\n \u003cp\u003e7\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 7.75316%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eBb\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 75.6329%;\"\u003e\n \u003cp\u003e\u003cstrong\u003edaily profile of cortisol\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 16.6139%;\"\u003e\n \u003cp\u003e7\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 7.75316%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eBc\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 75.6329%;\"\u003e\n \u003cp\u003e\u003cstrong\u003ecaptopril challenge test (CCT)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 16.6139%;\"\u003e\n \u003cp\u003e6\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 7.75316%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eBd\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 75.6329%;\"\u003e\n \u003cp\u003e\u003cstrong\u003efurosemide upright test (FUT)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 16.6139%;\"\u003e\n \u003cp\u003e12\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 7.75316%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eBe\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 75.6329%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eACTH stimulation test (AST)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 16.6139%;\"\u003e\n \u003cp\u003e15\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"2\" style=\"width: 83.3861%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eCategory C: General biochemical tests and blood counts\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 16.6139%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e37 features\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 7.75316%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eCa\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 75.6329%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eblood counts\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 16.6139%;\"\u003e\n \u003cp\u003e13\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 7.75316%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eCb\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 75.6329%;\"\u003e\n \u003cp\u003e\u003cstrong\u003ehepatic function\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 16.6139%;\"\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 7.75316%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eCc\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 75.6329%;\"\u003e\n \u003cp\u003e\u003cstrong\u003erenal function\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 16.6139%;\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 7.75316%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eCd\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 75.6329%;\"\u003e\n \u003cp\u003e\u003cstrong\u003elipid profile\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 16.6139%;\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 7.75316%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eCe\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 75.6329%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eglucose metabolism\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 16.6139%;\"\u003e\n \u003cp\u003e8\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 7.75316%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eCf\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 75.6329%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eOthers\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 16.6139%;\"\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 7.75316%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eCg\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 75.6329%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e24h urine others\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 16.6139%;\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"2\" style=\"width: 83.3861%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eCategory D: Urinary steroid profile\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 16.6139%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e89 features\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 7.75316%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eDa\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 75.6329%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eurinary metabolite\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 16.6139%;\"\u003e\n \u003cp\u003e63\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 7.75316%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eDb\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 75.6329%;\"\u003e\n \u003cp\u003e\u003cstrong\u003etotal of metabolites\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 16.6139%;\"\u003e\n \u003cp\u003e13\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 7.75316%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eDc\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 75.6329%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eratio of metabolite\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 16.6139%;\"\u003e\n \u003cp\u003e13\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 7.75316%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e \u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 75.6329%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTotal:\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 16.6139%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e196 features\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eThe 196 features used in the analysis, grouped into four Categories and 21 Subcategories are summarized. The number of features in each Category or Subcategory is shown. Category A (A1\u0026ndash;A23, blue) represents PA-related features examined in the outpatient setting. Category B (B1\u0026ndash;B47, red) includes endocrine challenge tests and diurnal hormonal profiles in the inpatient setting. Category C (C1\u0026ndash;C37, green) consists of outpatient-based general biochemical tests and blood counts. Category D (D1\u0026ndash;D89, gold) covers the urinary steroid profile, as further detailed in Supplementary Figure 2. The predictive contribution of each Category and Subcategory was investigated in machine learning-based subtype prediction models.\u003c/p\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"primary aldosteronism, steroid profile, subtype prediction, machine learning, predictive performance, predictive contribution ","lastPublishedDoi":"10.21203/rs.3.rs-8268064/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8268064/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003ePrimary aldosteronism (PA) has two major subtypes: unilateral (uPA) and bilateral (bPA). Although several diagnostic models for subtype classification have been reported, the optimal combination of algorithms and clinical features remains unclear. This study aimed to identify machine learning models and clinical features that contribute to PA subtype prediction. A total of 274 PA patients who underwent successful adrenal venous sampling (AVS) at a single center were analyzed. Overall, 196 endocrine features were comprehensively collected and classified into four categories: A, PA-related features; B, challenge tests; C, general biochemistry; and D, urinary steroid profile. Five machine learning algorithms were applied; predictive performance of the models as well as predictive contribution of features and categories were evaluated. Among the models, the random forest model achieved the highest predictive accuracy (91.3%). The most contributing feature in the RF model was plasma aldosterone concentration after the captopril challenge test (CCT90-PAC). Category B showed the greatest contribution to RF, followed by Categories A, D, and C. Combining Categories A and B improved predictive performance. These findings indicate that machine learning models, particularly RF, are effective for PA subtype prediction, with challenge test\u0026ndash;related features in Category B making a major contribution.\u003c/p\u003e","manuscriptTitle":"Machine Learning Approach for Subtype Prediction in Primary Aldosteronism: A Comprehensive Analysis of Models and Features","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-12-15 16:44:23","doi":"10.21203/rs.3.rs-8268064/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2026-01-20T09:57:12+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-01-19T07:17:52+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"214032036262304542855819896159844157929","date":"2026-01-19T06:51:24+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-12-31T13:07:01+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"19332021806322564322553044728538686724","date":"2025-12-12T06:14:46+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"244729778754168289219068736303505043746","date":"2025-12-11T12:30:37+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"329887202601113891941335172770866691749","date":"2025-12-10T15:55:21+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-12-09T23:26:31+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-12-09T23:25:13+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2025-12-08T13:08:02+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-12-06T07:43:50+00:00","index":"","fulltext":""},{"type":"submitted","content":"Scientific Reports","date":"2025-12-06T07:37:27+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"8c63a5be-7820-41ab-afe7-aeea04d4478b","owner":[],"postedDate":"December 15th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[{"id":59630441,"name":"Health sciences/Biomarkers"},{"id":59630442,"name":"Biological sciences/Computational biology and bioinformatics"},{"id":59630443,"name":"Health sciences/Diseases"},{"id":59630444,"name":"Health sciences/Endocrinology"},{"id":59630445,"name":"Health sciences/Medical research"},{"id":59630446,"name":"Health sciences/Risk factors"}],"tags":[],"updatedAt":"2026-02-23T16:14:25+00:00","versionOfRecord":{"articleIdentity":"rs-8268064","link":"https://doi.org/10.1038/s41598-026-41005-4","journal":{"identity":"scientific-reports","isVorOnly":false,"title":"Scientific Reports"},"publishedOn":"2026-02-20 15:57:09","publishedOnDateReadable":"February 20th, 2026"},"versionCreatedAt":"2025-12-15 16:44:23","video":"","vorDoi":"10.1038/s41598-026-41005-4","vorDoiUrl":"https://doi.org/10.1038/s41598-026-41005-4","workflowStages":[]},"version":"v1","identity":"rs-8268064","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8268064","identity":"rs-8268064","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}