Assessment, External Verification and Modification of Models in Diagnosing Primary Aldosteronism Subtypes

ABSTRACT

Diagnostic models are used to identify primary aldosteronism (PA) subtypes, including unilateral primary aldosteronism (UPA) and bilateral primary aldosteronism (BPA), only UPA could be treated through surgery. PA diagnostic models were screened from the NCBI, Embase, and Web of Science databases and assessed using the Prediction Model Risk of Bias Assessment Tool (PROBAST). External validation of these models was conducted using our Cohort. Models with high specificity were selected as the foundation for further modification. Based on the data from Cohort A and Cohort B, a new diagnostic model was developed by incorporating plasma aldosterone concentration (PAC) changes during postural stimulation testing (PST) as a variable. Eleven UPA diagnostic models were selected, with the concordance index (C index) varying from 0.745 to 0.938. Among these, three models conducted external validation, and one conducted internal validation. All models were identified as having a high risk of bias by using PROBAST. During external validations, the value of the C‐index ranged from 0.600 to 0.708. In the modified diagnostic model, the PAC change was set at ≤ ‐23.0% during PST. This adjustment increased specificity from 94.5% to 98.7% in cohort A, while specificity from cohort B increased to 96.3%. The performance of all included models needs improvement. Adding a PAC change of ≤ −23.0% during PST as an additional variable to the published UPA diagnostic models improved their specificity and may broaden the models’ applicability across different populations.

Abbreviations

ACE

angiotensin‐converting enzyme

ACTH

adrenocorticotropin

AVS

adrenal venous sampling

BPA

bilateral primary aldosteronism

CCT

captopril challenge test

CHARMS

critical appraisal and data extraction for systematic reviews of modelling studies

CT

adrenal computed tomography

LI

lateralization index

MR

magnetic resonance (MR)

PAC

plasma aldosterone concentration

PASO

primary aldosteronism surgery outcome

PROBAST

prediction model risk of bias assessment tool

PST

postural stimulation testing

RAAS

renin‐angiotensin‐aldosterone system

SI

selectivity index

SIT

saline infusion test

UPA

unilateral primary aldosteronism

1. Introduction

Primary aldosteronism (PA) is caused by excessive aldosterone secretion by the adrenal cortex (hyperplasia, adenoma, or cancer) and is characterized by hypernatremia, hypertension, and hypokaliema [1]. PA is a major cause of secondary hypertension [1, 2]. Compared to patients with essential hypertension, patients with primary aldosteronism have a higher risk of cardiovascular and cerebrovascular disease [3, 4, 5]. PA consists of two major subtypes: unilateral primary aldosteronism (UPA) and bilateral primary aldosteronism (BPA), with completely different treatments. Adrenalectomy is the primary treatment for UPA, and drug treatment is for BPA. Adrenal venous sampling (AVS) is the golden standard for identifying these two subtypes [1, 2, 6]; however, it is costly, invasive, and technically challenging [7, 8].

Aldosterone is a mineralocorticoids which is produced by the adrenal gland that retains sodium and expels potassium by the kidney, which is regulated by the renin‐angiotensin‐aldosterone system (RAAS). Adrenal computed tomography (CT) or magnetic resonance imaging (MR) is employed to diagnose PA but has a limited ability to identify PA subtypes. AVS is the golden standard for distinguishing PA subtypes; however, as an interventional procedure with a relatively high failure rate, it is invasive and expensive. With advancements in nuclear medicine, radionuclides such as [11C]metomidate for positron emission tomography (PET) and [68Ga]pentixafor for PET/CT have been utilized to diagnose various types of primary aldosteronism (PA). However, the high costs associated with these methods limit their widespread adoption.

The postural test, introduced in the 1970s, is a commonly used clinical examination for routine screening of PA. Studies have shown that patients with BPA exhibit an increase in aldosterone levels after standing upright for 2 h [9], in contrast, patients with UPA exhibit a decrease in aldosterone levels in response to postural changes [10]. Fuss et al. [11] found that PAC changes in postural tests could be used to diagnose UPA with a high specificity, which may be used as one of the variables for the Diagnosing model for PA subtypes.

The noninvasive clinical diagnostic model aims to alleviate the burden of the healthcare system and provide a reference to identify PA subtypes for clinician decision [7, 12, 13, 14, 15, 16, 17, 18, 19]. However, the efficiency of the PA subtype models varied. Therefore, we systematically reviewed and critically evaluated each of the published PA subtype diagnostic models. We then selected a diagnostic model to validate with the data from our retrospective study. The model was subsequently refined using data from our cohort.

2. Methods

2.1. Diagnosis of PA

Before clinical diagnosis, it is recommended that patients discontinue mineralocorticoid receptor antagonists and nonpotassium‐sparing diuretics for 4 weeks, β‐blockers, angiotensin‐converting enzyme inhibitors, and angiotensin II receptor blockers for 2 weeks. During this period, these antihypertensive medications should be replaced with non‐dihydropyridine calcium channel blockers and/or alpha‐blockers. Aldosterone and renin were measured at 8 am during bed rest during hospitalization. For patients with ARR>3.7 ng/dl/uIU/ml (according to the diagnostic criteria of UPA), a confirmatory test (captopril stimulation test [CCT) or saline infusion test [SIT]) was required. The diagnosis of primary aldosteronism was established when at least one of the following criteria was met: plasma aldosterone concentration (PAC) failed to suppress by at least 30% after the administration of 50 mg captopril for 2 h; or PAC exceeded 10 ng/dl following the infusion of 2 liters of isotonic saline while the patient was in the supine position [1]. The concentration of aldosterone was measured by KingMed, a third‐party laboratory. When the levels approached the upper detection limits, the samples were proportionally diluted and then reanalyzed.

Once PA was confirmed, a thin‐slice CT of the adrenal gland is required. CT findings of a UPA included a unilateral macroadenoma (>10 mm) [20, 21, 22], while the contralateral gland was smooth. The CT manifestations BPA included bilateral macroadenomas, normal‐appearing adrenals, minimal adrenal limb thickening, or microadenomas (≤10 mm).

An experienced radiologist used AVS to estimate the lateralization of aldosterone hypersecretion. The selectivity index (SI) was used to assess successful cannulation of the adrenal vein, which was calculated as the ratio of cortisol in the adrenal vein to cortisol in the peripheral vein (> 2 without ACTH stimulation, or > 3 with ACTH stimulation was considered positive). The ratio of cortisol‐corrected PAC on the dominant and non‐dominant side was used to calculate the lateralization index (LI); a ratio > 3 without ACTH stimulation or > 4 with ACTH stimulation was considered positive [17, 23].

UPA was diagnosed using the following methods. After surgery, complete biochemical remission was assessed according to the Primary Aldosteronism Surgical Outcome (PASO) criteria during a follow‐up period of 6 to 12 months after adrenalectomy [24]. Patients with manifested lateralization of aldosterone hypersecretion in the AVS without undergoing surgery were also diagnosed with UPA for further analysis.

2.2. Date Source

We utilized clinical data from patients diagnosed with primary aldosteronism (PA) at the First and Second Affiliated Hospitals of Xi'an Jiaotong University (2017–2025). The diagnosis of PA was based on the captopril challenge test (CCT) or saline infusion test (SIT), and subtype diagnosis (unilateral PA, UPA; bilateral PA, BPA) was determined by adrenal venous sampling (AVS) and/or post‐adrenalectomy outcomes. As a retrospective study, verbal informed consent was obtained from each participant during telephone follow‐up, in agreement with the ethics committee. Due to the limited number of cases from the Second Affiliated Hospital, which was insufficient for standalone validation, we combined the data from both hospitals. The combined cohort was then randomly divided into Cohort A (70%, n = 209) for model development and cutoff determination, and Cohort B (30%, n = 81) for internal split‐sample validation. Cohort A served as the development cohort and Cohort B as the validation cohort. The ethics committee of the Second Affiliated Hospital of Xi'an Jiaotong University approved the study protocol. All included patients were hospitalized with relatively complete clinical records; no imputation methods were applied, and complete case analysis was used throughout model validation and modification.

To minimize inter‐center variability, both hospitals adhered to unified protocols for PA diagnosis and subtyping, including consistent criteria for patient recruitment, medication washout, confirmatory tests (CCT or SIT), and AVS procedure. AVS was performed with or without ACTH stimulation following a standardized protocol, with successful cannulation defined as selectivity index ≥2 (without ACTH) or ≥3 (with ACTH). The lateralization index was calculated as the cortisol‐corrected aldosterone ratio, with a cutoff >3 (no ACTH) or >4 (ACTH) indicating unilateral hypersecretion. Adrenal CT images were interpreted independently by experienced radiologists at each center using identical definitions, including unilateral nodule ≥10 mm.

A total of 290 patients with confirmed PA were included. For patients without AVS, UPA was diagnosed by post‐operative remission (PASO criteria, n = 112) and BPA was diagnosed by clinical follow‐up outcomes (n = 58). The total cohort was randomly divided into Cohort A (n  =  209, 70%) for model development and cutoff determination, and Cohort B (n  =  81, 30%) for validation, stratified by UPA/BPA status. For sensitivity analysis, we further identified a subgroup of 120 patients who had a definitive subtype diagnosis by AVS alone (76 UPA, 44 BPA), independent of post‐operative outcomes.

2.3. The Review and Evaluation of UPA Diagnostic Models

We searched PubMed, Embase, and Web of Science for published UPA diagnostic models. The last search was conducted on January 1, 2024, using a merged method consisting of MeSH headings search strategy and the following terms “primary aldosteronism”, “diagnostic model”, “unilateral primary aldosteronism”. Two independent reviewers (JYY and GHN) conducted standardized data extraction based on the CHARMS (Critical Appraisal and Data Extraction for Systematic Reviews of modelling Studies) checklist [25] and PROBAST (model risk of bias assessment tool) [26].

2.4. Position Tests

Samples were collected for serum aldosterone and renin levels were measured for PA using postural stimulation testing. For the supine test, the sample was collected at 8:00 am when the patients had bed rest overnight. For the upright test, the sample was taken after the patient had been in an upright position or ambulance for four hours. We then calculated the ratio of PAC change in response to posture according to the following formula:

$$PACchange\left(\%\right)=\frac{uprightPAC-supinePAC}{supinePAC}$$

The postural stimulation test (PST) was performed under standardized conditions for all patients. Blood samples were collected in the supine position after at least 60 min of recumbency and again after 2 h of upright posture (standing or walking). All PAC measurements were conducted using the same electrochemiluminescence immunoassay platform (Roche Diagnostics, Germany) in a single certified laboratory (KingMed Diagnostics) throughout the study period. Intra‐assay and inter‐assay coefficients of variation for PAC were <5% and <8%, respectively. PAC change was calculated uniformly across all patients using the formula above. For patients in Cohort B, the same PST protocol and assay platform were applied to ensure comparability between cohorts.

2.5. Statistical Analysis

Continuous variables with a normal distribution are presented as mean ± standard deviation (SD) and categorical variables are presented as frequencies (percentages). T‐tests and chi‐squared tests were conducted to compare the differences between the UPA and BPA groups. The change in PAC during the position tests was used as a continuous variable and assessed using the concordance index (C‐index). The The cutoff value was determined by selecting the earliest point at which the change in PAC change demonstrated specificity greater than 90%. To assess the potential risk of overfitting in the process of cutoff determination, we performed internal validation of the PAC change cutoff (≤−23.0%) using 1000 bootstrap resampling iterations within cohort A. After bootstrap correction, the discriminative performance of the cutoff value was re‐evaluated, and the calibrated C‐index with its 95% confidence interval was reported. We modified the published UPA diagnostic model (Song [18]) by adding a variable for PAC change. Data from cohort A was used to screen out the cut‐off value of PAC change and to validate the modified model using the data from cohort B.

2.5.1. Model Calibration and Decision Curve Analysis

Model calibration was assessed using the Hosmer‐Lemeshow goodness‐of‐fit test with 10 groups based on the predicted risk, and visualized via calibration plots. Decision curve analysis (DCA) was performed across threshold probabilities from 0% to 50% to evaluate the clinical net benefit of the modified model, compared with treat‐all and treat‐none strategies.

3. Results

3.1. Baseline Data

Baseline data of the patients from the two Affiliated Hospitals of Xi'an Jiaotong University showed a point prevalence rate of 0.006%; 290 of these patients were diagnosed with UPA and BPA by AVS (120 cases, 41.4%) and/or postadrenalectomy results (170 cases, 58.6%). Cohort A  (n = 209) was composed of 132 UPA and 77 BPA, cohort B  (n = 81) was consisted of 56 UPA and 25 BPA. The diagnostic source for subtype confirmation was as follows: 120 patients (76 UPA, 44 BPA) were diagnosed by AVS alone, and 170 patients (112 UPA, 58 BPA) were diagnosed by either post‐operative outcomes (for UPA) or clinical follow‐up (for BPA) according to PASO criteria. Table 1 shows the baseline characteristics, and significant statistical differences were found in serum potassium, supine PAC, and PAC change between the UPA and BPA patients.

TABLE 1.

Baseline characteristics of patients in PA.

Variables	UPA (n = 188)	BPA (n = 102)	P‐vaule
Age (years)	47.8 ± 0.94	49.3 ± 1.14	0.40
Male/Female	72/116 *	62/40	0.02
SBP (mmHg)	167.4 ± 1.97 *	161.7 ± 2.72	0.03
BMI (kg/m2)	25.9 ± 0.6	24.4 ± 0.4	0.42
Organ damages [n (%)]	78(41.5)	50(49.0)	0.57
Serum potassium (mmol/l)	2.92 ± 0.55	3.25 ± 0.82	0.09
eGFR(ml/min)	99.4 ± 1.88	105.9 ± 4.15	0.67
Supine PRC (uIU/ml)	8.18 ± 2.85	7.06 ± 2.52	0.53
Upright PRC (uIU/ml)	8.49 ± 3.10	10.34 ± 4.30	0.56
Supine PAC (ng/dl)	37.30 ± 2.20 *	26.1 ± 2.65	0.03
Upright PAC (ng/dl)	35.1 ± 2.02	33.0 ± 3.57	0.07
PAC change (%)	0.68 ± 2.67 *	−26.8 ± 4.24	0.02
Adenoma ≥10 mm on CT	115(61.2) *	37(36.2)	0.01

Note: Data are expressed as percentages or medians (interquartile range), and the latter were all subject to normal distribution. P‐values represent the comparison between BPA and UPA. eGFR, estimated glomerular filtration rate; PAC, plasma aldosterone concentration; PRC, plasma renin concentration; UPA, Unilateral primary aldosteronism; BPA, Bilateral primary aldosteronism.

^* P < 0.05.

To evaluate potential heterogeneity between the two centers, we compared the baseline clinical and biochemical parameters of patients from the First Affiliated Hospital and the Second Affiliated Hospital. No statistically significant differences were observed in age, sex, systolic blood pressure, serum potassium, eGFR, supine or upright PAC, PAC change, or CT findings (all p > 0.05). These results indicate good homogeneity of patient characteristics across the two centers, supporting the validity of data pooling (detailed comparisons are provided in  Table S5).

3.2. Literature Search and Selection

Utilizing the described derivation strategy, we conducted a comprehensive search across three databases. From the available data at two participating centers, we identified 11 UPA diagnostic model studies for subsequent evaluation and validation (Figure 1).

FIGURE 1.

Flow diagram outlining the literature search and study selection for PA diagnostic model.

3.3. Diagnostic Models

The selection of 11 published UPA diagnostic models is shown in Figure 1. The characteristics of these diagnostic models are presented in Table S1 [27]. All of the studies were case‐control studies, five were multi‐center studies(include two international multi‐center cohorts.) [12, 14, 17, 18, 28] and six were single‐center studies [13, 15, 29, 30, 31, 32]. The diagnostic models were developed mainly in Asia (China, Japan, and Singapore), Europe (Italy, Spain, and Germany), Australia. For the selection of variables, logistic regression analysis was employed in six studies [12, 15, 17, 18, 29, 32], Lasso regression in one study [13], and four studies were not reported [14, 28, 30, 31]. Finally, two to five variables were included in these models, including age (n = 4), blood potassium level (n = 9), adrenal CT image (n = 9), and plasma aldosterone level (n = 4). In terms of evaluation, seven studies used the C‐index to evaluate performance, with the C‐index ranging from 0.938 to 0.745 [13, 15, 17, 18, 28, 29, 32]. Five studies employed as score‐scaled model [12, 18, 29, 32, 33], while one study used a nomogram to perform the model [13], and the remaining models only presented variables. In the development set, two models exhibited a specificity of 100% and a sensitivity of 52%–53% [18, 29], indicating that 52%–53% of UPA patients could bypass AVS before curative surgery. One study performed internal validation with a C‐index of 0.912 [13]. External validation was reported in two studies [12, 18], but only one study reported a sensitivity of 13% and a specificity of 100%. Additionally, one model was assessed using a calibration plot [13].

All included studies were evaluated by PROBAST, as shown in Table S2 [27]. All included studies were case‐control designs and were therefore rated as high risk of bias in the participant domain. For variables, only two studies [15, 28] were evaluated as low risk and nine were evaluated as unclear due to CT imaging [12, 13, 14, 17, 18, 29, 30, 31, 32], which was considered an opaque variable. In the analysis section, all studies were measured as having a high risk of bias because the number of events should increase by at least 10 for each included variable to avoid overfitting; only one study met this requirement [12]. The continuous variable was transformed into a categorical variable with specific descriptions (five studies) [12, 13, 18, 29, 31]. Nine models (81.8%) were not subjected to either internal or external validation. Moreover, the evaluation of discrimination and calibration was incomplete, and the evaluation of discrimination and calibration was incomplete, and inappropriate statistical methods were used in all 11 studies (n = 11, 100%). In terms of applicability, all included studies showed high risks for participants and outcomes, whereas the predictors were unclear (Table S3 [27]).

3.4. External Validation of UPA Diagnostic Models

We also evaluated the Endocrine Society Guideline criteria [1] as an additional reference; results are in Table 2. The baseline data of PA patients are shown in Table 1. Significant differences were observed in gender, SBP, serum potassium, supine and upright PAC, PAC change, CT findings among patients with UPA and BPA. What's more, we used our cohort to verify the performance of the published UPA diagnostic models, and the results are shown in Table 2. The C‐index values were ranged from 0.600 to 0.708 (Figures S1–7 [27]), with the models from The Endocrine Society Guideline [1] and Song [18] showing a specificity > 90%, the performance of all included models needs improvement. The primary objective of the modified model was to identify UPA patients eligible for surgery, with the goal of minimizing diagnostic errors and suboptimal treatment, rather than achieving the highest possible AUC. Based on this, we focused on the model with the highest specificity (Song [18]) to make further modifications to avoid unnecessary surgery in BPA patients.

TABLE 2.

Sensitivity and specificity of other diagnostic models in the validation cohort.

		Cohort(n = 290)
References	Variables	Sensitivity	Specificity	Positive prediction value	Negative prediction value	Accuracy	Positive likelihood ratio	C‐Index
The Endocrine Society Guideline [1]	serum potassium <3.5 mmol/l, PAC > 30 ng/dl, unilateral nodule > 10 mm, age <35 years	3.2%	90.1%	37.5%	33.6%	33.8%	0.32	0.687
Young [31]	Age <40 years, unilateral adrenal nodule > 10mm	17.0%	86.2%	69.5%	36.1%	41.3%	1.41	0.600
Küpers [29]	A typical Conn's adenoma, (unilateral adrenal nodule > 10 mm and encapsulated), serum potassium <3.5 mmol/L, eGFR ≥ 100 mL/min/1.73 m2	27.6%	85.3%	77.6%	39.0%	47.9%	1.88	0.606
Umakoshi [30]	PAC > 15.9 ng/dl, serum potassium <3.5 mmol/l, unilateral nodule > 10 mm, age <35 years	6.3%	84.3%	42.8%	32.8%	33.8%	0.41	0.678
Lee [14]	serum potassium <3.5 mmol/l, PAC > 30 ng/dl, Unilateral lesion > 7 mm,	25.0%	85.3%	75.8%	38.1%	46.2%	1.70	0.703
Rossi 2022 [17]	age > 45 years, serum potassium <3.6 mmol/l, unilateral nodule > 5mm	25.0%	86.3%	72.3%	38.4%	46.5%	1.82	0.656
Song [18]	PAC ≥ 20 ng/dl, PRA ≤ 5uIU/ml, serum potassium ≤ 3.5 mmol/l, unilateral nodule ≥ 10mm	34.0%	91.1%	87.7%	42.3%	54.1%	3.80	0.708

Note: The Endocrine Society Guideline criteria are not part of the 11 systematic models but were included as an additional reference.

Abbreviations: PAC, plasma aldosterone concentration; PRA, plasma renin activity.

3.5. Development of the Modified UPA Model

For the variables that showed significant differences between UPA and BPA in Table 1, we included PAC change to modify the Song model [18]. From the literature review, we found that Fuss et al. [11] employed PAC change during the postural stimulation test to diagnose UPA, using a cut‐off value of at least −28% or −30%. PA patients was used to conduct validation and make some improvements, as shown in Table 3. When the cut‐off value of PAC change was set to ≤ −28%, the specificity was 92.0%, with only 25 predictive values. When the cutoff value of PAC change was set to ≤ −30% [34], the specificity was 94.7% and the sensitivity was 16.2%, with 22 predictive values. However, such a small positive predictive value limits the generalization of the model. Considering the low sensitivity corresponding to specificity and the limited generalization of the model, we used cohort A to reset the cutoff value of PAC change. The r results indicated that when PAC change was ≤ −23.0%, the specificity reached 92.0% (exceeding the predefined threshold of 90%; Table S4) [27], with 34 positive values in cohort A. After internal validation using 1000 bootstrap resamples, the modified model with the PAC change cutoff (≤ −23.0%) achieved a bootstrap‐corrected C‐index of 0.724 (95% CI: 0.696–0.752) in cohort A, with an optimism estimate of 0.009, suggesting minimal overfitting and robust model stability. The Song [18] model, when validated in Cohort A and Cohort B, yielded specificities of 94.5% and 81.5%, respectively. The optimal diagnostic model should have as high a specificity as possible, meaning that BPA patients should reduce the possibility of being misdiagnosed as UPA to avoid unnecessary adrenalectomy using this model. Therefore, PAC change ≤ −23.0% was included in the modified model, which also contained PAC ≥ 20 ng/dl, PRA ≤ 5uIU/ml, serum potassium ≤ 3.5 mmol/l, unilateral nodule ≥ 10 mm. In cohort A, the specificity was 98.7%, and the specificity was 96.3% for cohort B by the modified model. The 92.0% specificity in Table S4 represents the PAC change alone, whereas the 98.7% specificity here is for the combined model (Song [18] criteria plus PAC change) Table 4.

TABLE 3.

The modification of the cut‐off value for PAC change.

PAC change (%)	Sensitivity	Specificity	Sample size of the positive predictive value
≤−30.0	16.2%	94.7%	22
≤−28.0	18.4%	92.0%	25
≤−23.0	25.0%	92.0%	34

Note: Specificity threshold was set at ≥90% for cutoff selection.

Abbreviation: PAC, plasma aldosterone concentration.

TABLE 4.

Performance of the Song [18] model, PAC change alone, and the modified UPA diagnostic model in the development and validation cohorts.

	Cohort A (n = 209)							Cohort B (n = 81)
Models	Sensitivity	Specificity	Positive prediction value	Negative prediction value	Accuracy	Positive likelihood ratio	C‐index	Sensitivity	Specificity	Positive prediction value	Negative prediction value	Accuracy	Positive likelihood ratio	C‐Index
Song 2022 [18]	32.3%	94.5%	91.6%	43.5%	54.5%	6.06	0.682	38.4%	81.5%	80.0%	40.7%	53.2%	2.07	0.715
PAC change (%)	25.0%	92.0%	85.0%	40.3%	48.8%	3.13	0.752	34.6%	88.9%	85.7%	41.3%	53.2%	3.11	0.809
Modified UPA diagnostic model	10.3%	98.7%	93.3%	37.7%	41.7%	7.72	0.733	7.7%	96.3%	80.0%	35.1%	37.9%	2.07	0.717

Note: Song [18] model: PAC ≥20 ng/dl, PRA ≤5 uIU/ml, serum potassium ≤3.5 mmol/l, unilateral nodule ≥10 mm on CT. Modified UPA diagnostic model: Song [18] criteria plus PAC change ≤ −23.0% during postural stimulation test. Modified UPA diagnostic model: Song [18] criteria plus PAC change ≤ −23.0%.

Abbreviations: Acc, accuracy; NPV, negative predictive value; PAC, plasma aldosterone concentration; PLR, positive likelihood ratio; PPV, positive predictive value; PRA, plasma renin activity; Sen, sensitivity; Spe, specificity.

3.6. Sensitivity Analysis for Reference Standard Verification

To assess the potential impact of including patients diagnosed without AVS, we performed a sensitivity analysis restricted to the 120 patients whose UPA or BPA diagnosis was confirmed by AVS alone (76 UPA, 44 BPA). In this AVS‐only subgroup, the modified model achieved a bootstrap‐corrected C‐index of 0.695 (95% CI: 0.655–0.735) with an optimism estimate of 0.007, a specificity of 94.2%, and a sensitivity of 28.6%. These results were consistent with those from the full cohort analysis (corrected C‐index: 0.724; specificity: 98.7% in Cohort A), suggesting that the inclusion of post‐operatively diagnosed patients did not substantially affect our primary conclusions.

3.7. Model Calibration and Decision Curve Analysis

Calibration of the modified UPA model was evaluated in Cohort A (n = 209). The Hosmer‐Lemeshow test showed no significant lack of fit (χ ² = 6.53, P  = 0.59). The calibration plot demonstrated good agreement between predicted probabilities and observed UPA rates (Figure S8).

Decision curve analysis (Figure S8) showed that the modified model provided higher net clinical benefit than both treat‐all and treat‐none strategies for threshold probabilities ranging from 8% to 32%, supporting its clinical utility in identifying UPA patients suitable for adrenalectomy while minimizing unnecessary surgery in BPA patients.

4. Discussion

Our study involved a systematic search and review of existing diagnostic models for UPA, with 11 studies selected for detailed assessment and validation. The primary aim of these models was to minimize the reliance on AVS for surgical decision‐making and to support clinical decision‐making in settings where AVS is unavailable. Nevertheless, most of the models developed in these studies were not validated, leaving their clinical utility and performance uncertain.

Each of the incorporated studies exhibited a high risk of bias owing to limitations such as inadequate sample sizes, inaccuracies in variable translation, insufficient documentation or standardized management of missing data, suboptimal variable selection techniques, overfitting, and incomplete assessment of model performance. Owing to the overall high risk of bias for the included models, few low‐risk UPA diagnostic models are available. In addition, we validated these published UPA diagnostic models in our cohort. Inevitable differences existed among PA patients, impling their selection bias; selection biases arose from the follows reasons: Firstly, medications such as ARBs, ACE inhibitors, β‐blockers, and spironolactone, which may affect the RAAS, could compromise the accuracy of the results. In PA patients managed with these antihypertensive agents, clinicians replaced them with non‐dihydropyridine calcium channel blockers or α‐blockers. Patients were subsequently admitted for precise diagnostic testing following a washout period. Nevertheless, loss to follow‐up was observed during the washout phase for multiple reasons. Secondly, the determination of PA subtypes via AVS or adrenal surgery is significantly influenced by the clinical habits of physicians at different hospital. In an effort to prevent misdiagnosis and suboptimal treatment, urologists often prioritize surgical intervention for patients with larger unilateral adrenal adenomas, potentially leading to selection bias.

According to the data of our cohort, we included 11 studies, each of which developed diagnostic models for UPA. A comprehensive literature review revealed that high plasma aldosterone concentration (PAC), low plasma renin concentration (PRC), and hypokalemia were the most frequently utilized variables in these UPA diagnostic models. The implementation of these diagnostic models has the potential to allow 3.2%–34.0% of UPA patients to bypass AVS—an expensive, invasive procedure with a high failure rate—before undergoing adrenalectomy. The 11 UPA diagnostic models had different variables, but most of them were focused on the PAC level, which included PRA, age, eGFR level, and the size of unilateral adrenal nodules. The models from the Endocrine Society Guidelines [1] set a high cut‐off value of PAC level, and made unilateral adrenal nodule > 10 mm as a variable, but did not take into account the inhibitory effect of high PAC level on PRA in UPA patients, which may be the reason for the relatively low specificity in the validation cohort. The aims of the included age restriction [1, 30, 31] may distinguish it from essential hypertension, and the inclusion of eGFR≥100 mL/min/1.73 m² as a variable is to exclude hypertension and hypokalemia caused by chronic kidney diseases. After the external validation by our cohorts, we found that the study of Song [18] had the highest specificity of 91.1%, which included the variables of PAC≥20 ng/ml, PRA≤5uIU/ml, serum potassium≤3.5 mmol/l, and unilateral nodule>10 mm. However, when we conducted validation in our cohort, we found that 9 BPA patients were incorrectly diagnosed with UPA and underwent unnecessary adrenalectomy.

Fuss et al. found that PAC changes in postural tests could be utilized for the diagnosis of UPA. Postural examination is a common clinical procedure and is relatively easy to popularize. We aimed to improve the specificity of the model based on Song [18] by adding PAC change during postural tests. This supplementation aims to reduce the number of patients misdiagnosed with UPA and undergoing unnecessary adrenalectomy. The postural test, proposed in the 1970s, can differentiate between UPA and BPA. The theoretical basis is that patients with BPA are sensitive to slight increases in angiotensin II during these tests, leading to elevated PAC levels [34]. In contrast, some UPA patients show a tendency of decreased aldosterone levels [10, 34, 35, 36, 37].

Fuss et al. [11]. found that a PAC change <−28% in the postural tests was a variable that could screen out patients with UPA. However, the data from our cohort showed that a PAC change <−28% had a high specificity. Its predictive value was low, limiting its generalizability. When the threshold for PAC change was adjusted to ≤ −23.0%, specificity slightly reduced, but sensitivity and positive predictive values increased. Therefore, we added PAC change ≤ −23.0% during the Position tests into the Song model [18] to create a modified UPA model. Of note, the PAC change cut‐off of ≤ −23.0% alone showed a low sensitivity of 25.0% (Table S4); therefore, it was not intended as a standalone test but as an additional criterion within the combined model to enhance specificity.

Compared with the Song model [18], the specificity of the modified model increased from 94.5% to 98.7% in cohort A, indicating that the modified diagnostic model reduced the number of BPA patients misdiagnosed as UPA, thereby preventing unnecessary adrenalectomies. In cohort B, the specificity increased to 96.3% after modification. This result may be attributed to the the presence of large adrenal nodules in UPA patients at this center and the limited number of UPA cases diagnosed by AVS.

Several limitations should be acknowledged. First, as a retrospective study, the modified UPA model requires prospective validation across more centers. Second, over half of the PA patients (n = 170, 58.6%) were diagnosed by postoperative outcomes rather than AVS, which may introduce misclassification bias. To minimize this, we applied strict PASO‐based criteria for postoperative remission (blood pressure improvement, normalization of serum potassium and aldosterone, pathological confirmation). A sensitivity analysis restricted to AVS‐confirmed patients (n = 120) yielded consistent results (Section 3.6), supporting the robustness of our findings. Nevertheless, excluding patients without AVS or adrenalectomy, failed AVS, missing follow‐up or data may have introduced selection bias. Third, data were pooled from two affiliated hospitals. Despite unified protocols, subtle inter‐center differences in patient characteristics, AVS procedures, and CT interpretation may still exist. To mitigate this, we randomly split the pooled data into cohort A (70%) and cohort B (30%), each containing patients from both centers, thereby balancing the center effect. However, the lack of formal inter‐center reliability testing is a limitation. Future multi‐center studies with standardized protocols are needed to validate the model's generalizability. Fourth, as a retrospective study, minor variability in PST implementation (e.g., exact upright duration) and assay conditions cannot be fully excluded. Although all PAC changes were calculated uniformly using the same platform, batch‐to‐batch assay variability may have introduced measurement uncertainty. Prospective studies with strictly standardized protocols are warranted to further validate these findings. Furthermore, calibration analysis confirmed good agreement between predicted and observed UPA probabilities, and decision curve analysis demonstrated a favorable net benefit profile across clinically relevant threshold probabilities, further supporting the clinical applicability of the modified model.

In conclusion, the 11 UPA diagnostic models analyzed in this study exhibited a significant risk of bias, and the specificity declined after external validation. The modified UPA model, which incorporates PAC change as a variable, reduced the number of BPA patients misdiagnosed as UPA instead of BPA, thereby preventing unnecessary adrenalectomies.

Author Contributions

Yingyu Jing: project administration, data curation, writing – original draft. Ning Wang: methodology, formal analysis, resources, supervision. Bo Sun: data curation, review and editing. Yifan Zhang: original draft, data curation. Haonan Guo: software, formal analysis, Rongrong Hou: investigation. Jing Xu, Jingyue Wang, Xingyan Pan, Xinyi Niu, Lin Song, Lili Huang, Jinyan Zhang, and Fang Jia: data curation.

Supporting information

Supplementary Table 1: Overview of UPA diagnostic models.

Supplementary Table 2: Risk of bias assessment (using PROBAST) based on 4 domains across 11 studies that created models for UPA.

Supplementary Table 3: The detail of the items of Risk of bias assessment (using PROBAST) for the selected diagnostic models.

Supplementary Table 4: The modification of the cut‐off value for PAC change.

Supplementary Figure 1: The validation of the UPA model from Endocrine Society Guidelines (Funder et al., 2016).

Supplementary Figure 2: The validation of the UPA model from Young 2004 (Young et al., 2004).

Supplementary Figure 3: The validation of the UPA model from Küpers 2012 (Kuepers et al., 2012).

Supplementary Figure 4: The validation of the UPA model from Umakoshi 2018 (Umakoshi et al., 2018).

Supplementary Figure 5: The validation of the UPA model from Lee 2021 (Lee et al., 2021).

Supplementary Figure 6: The validation of the UPA model from Rossi 2022 (Rossi et al., 2022).

Supplementary Figure 7: The validation of the UPA model from Song 2022 (Song et al., 2022).

Data Availability Statement

The datasets are available from the corresponding author on reasonable request.