Machine learning-based prediction of impaired arginine vasopressin secretion during hypertonic saline test

Abstract

Hypertonic saline infusion and water deprivation tests are commonly used to stimulate arginine vasopressin (AVP) for the diagnosis of AVP deficiency (AVP-D); however, these procedures impose a considerable burden on patients. We aimed to develop machine learning models to predict impaired AVP secretion and thereby reduce the need for AVP stimulation testing. This retrospective cohort study included 64 patients who underwent the hypertonic saline test (HST) at Nagoya University Hospital, Japan, between 2018 and 2024. Impaired AVP secretion was defined as a predicted plasma AVP level <1.0 pg/mL at a serum sodium level of 149 mEq/L during the HST. Feature selection was performed using univariate screening and systematic selection procedures. Logistic regression and support vector machine models were developed using baseline clinical and laboratory parameters obtained before the HST and validated using nested cross-validation. The primary outcome was the area under the receiver operating characteristic curve (AUC); secondary outcomes included the sensitivity, specificity, and positive and negative predictive values. Four variables (urinary osmolality, serum sodium, plasma AVP, and blood urea nitrogen) were selected. The logistic regression model achieved an AUC of 0.862 for predicting impaired AVP secretion. Among 35 patients who underwent the HST for suspected AVP-D, at the optimized threshold, the positive and negative predictive values were 100% (14/14 patients) and 33.3% (7/21 patients), respectively. The developed machine learning model can identify a subset of patients with impaired AVP secretion without requiring the HST, potentially reducing the number of patients who need AVP stimulation testing for AVP-D diagnosis.

Graphical Abstract

Introduction

Arginine vasopressin (AVP), an antidiuretic hormone, is synthesized in the hypothalamus and released from the posterior pituitary to maintain water balance in response to changes in plasma osmolality, with the serum sodium concentration serving as the primary osmotic stimulus [1]. AVP deficiency (AVP-D), formerly termed central diabetes insipidus, results in impaired urinary concentration capacity, leading to hypotonic polyuria and compensatory polydipsia. Accurate diagnosis is critical because AVP-D must be distinguished from other causes of polyuria-polydipsia syndrome, including AVP resistance (nephrogenic diabetes insipidus) and primary polydipsia, as the management strategies differ fundamentally.

Both the water deprivation test and hypertonic saline test (HST) are established diagnostic methods for evaluating polyuria-polydipsia syndrome [2]. The former has limited diagnostic accuracy (70–78%) [3, 4] and carries a substantial risk of dehydration due to prolonged fluid restriction. The HST assesses AVP secretion responses by inducing hyperosmolality via intravenous hypertonic saline, which normally stimulates AVP release but elicits an impaired response in patients with AVP-D. Copeptin, a stable peptide co-secreted with AVP [5, 6], has significantly improved diagnostic accuracy in evaluating polyuria-polydipsia syndrome [4], and copeptin-based stimulation tests using hypertonic saline or arginine are increasingly being adopted worldwide as practical alternatives for direct AVP measurement [3, 7, 8]. In Japan, AVP-D is diagnosed by the plasma AVP level obtained during the HST, according to a diagnostic cutoff of <1.0 pg/mL predicted plasma AVP at a serum sodium level of 149 mEq/L (99% sensitivity and 95% specificity) [9]. Although the HST provides excellent diagnostic accuracy, it imposes a considerable burden on patients, requires close medical supervision, and is often impractical for routine use, particularly in vulnerable subjects with severe heart failure and renal dysfunction.

Given these limitations, there is a need for less invasive tests to diagnose AVP-D. Machine learning, the process of integrating multiple clinical and biochemical parameters into predictive models, has been shown to be useful for diagnosis in diverse areas of clinical medicine. A previous study developed a model to distinguish AVP-D from primary polydipsia based on five clinical variables: urinary osmolality (U-Osm), serum sodium level, plasma glucose level, history of transsphenoidal surgery, and anterior pituitary deficiency [10]. However, that study employed AVP-D diagnosis by experts as the reference standard; no studies have developed machine learning models using the HST-obtained AVP value as the reference standard.

We therefore aimed to develop and validate a model to predict impaired AVP secretion, defined as a predicted plasma AVP level <1.0 pg/mL at a serum sodium level of 149 mEq/L during the HST [9, 11]. The model is intended as a decision-support tool for endocrinologists and other physicians evaluating patients with suspected AVP-D, to help identify individuals who may not require the HST, thereby reducing unnecessary invasive testing and patient burden.

Materials and Methods

Study design and participants

We retrospectively analyzed consecutive patients who underwent the HST at Nagoya University Hospital between January 2018 and December 2024. The sample size was determined by the availability of all eligible patients during this period, with no formal a priori calculation as this was an exploratory study. The exclusion criteria were as follows: a suspected residual effect of desmopressin, early termination resulting in an incomplete HST protocol, failure to achieve hypernatremia during the HST (peak serum sodium level <146 mEq/L), and missing U-Osm data.

Impaired AVP secretion was defined as a predicted plasma AVP level <1.0 pg/mL at a serum sodium level of 149 mEq/L during the HST, according to established Japanese criteria [11]. Patients who did not meet this criterion were classified as controls. The cutoff (i.e., a predicted plasma AVP level <1.0 pg/mL at a serum sodium level of 149 mEq/L) has been reported to identify AVP-D during the HST with 99% sensitivity and 95% specificity [9]. To determine the predicted plasma AVP level for each patient, we performed a simple linear regression analysis using serial plasma AVP and serum sodium measurements obtained during the HST. The resulting regression equation was then used to estimate the plasma AVP level at a serum sodium level of 149 mEq/L. We determined the average coefficient of determination (R²) and standard error (SE) of the estimated AVP value at a serum sodium level of 149 mEq/L for these individual regression analyses.

This retrospective study involving human participants was in accordance with the ethical standards of the institutional research committee and with the 1964 Helsinki Declaration and its later amendments. The study was approved by the Institutional Review Board of Nagoya University (approval number: 2025-0107). Informed consent was waived due to the retrospective nature of the study; however, patients were given the opportunity to opt out via the hospital’s website.

Hypertonic saline test (HST)

The HST was started at 9:00 am. Before the test, participants remained in the supine position for 30 min. A 5% saline solution was administered intravenously at a rate of 0.05 mL/kg/min for 120 min. To measure serum sodium and plasma AVP levels, blood samples were collected before infusion (0 min) and at 30, 60, 90, and 120 min after the start of infusion. The target level of hypernatremia during the HST was a serum sodium level ≥150 mEq/L; however, a peak serum sodium level ≥146 mEq/L was used as an inclusion criterion (patients with levels <146 mEq/L were excluded). The test was terminated early if participants could not continue because of adverse events (e.g., headache, nausea, vomiting, or infusion-site pain) or if the attending physician decided the test should be discontinued for safety reasons.

Data collection

The following data were extracted from electronic medical records: demographics (age, sex), clinical history (transsphenoidal surgery, anterior pituitary deficiency), and laboratory values (renal function, complete blood count, serum electrolytes, liver enzymes, glucose, plasma AVP, and U-Osm). To ensure that the model inputs reflected the information available before testing, baseline measurements obtained immediately before HST initiation and after an overnight fast were used as the laboratory variables for the machine learning models. If a variable was not measured immediately before HST initiation, the most recent value available prior to the HST was used. Measurements obtained during the HST were used for outcome definition only and were not included as model inputs. Anterior pituitary deficiency was defined as documented impairment of one or more anterior pituitary hormones, diagnosed on the basis of basal hormone levels and/or stimulation tests performed prior to the HST. Plasma AVP was measured by radioimmunoassay (RIA; YAMASA Corporation, Choshi, Japan), with a detection limit of 0.40 pg/mL; values below the detection limit were imputed as 0.27 pg/mL (two-thirds of the detection limit).

Feature selection

Univariate screening was applied to 21 candidate variables, representing routinely available clinical and laboratory parameters extracted from the dataset. The Benjamini–Hochberg procedure was used to control the false discovery rate (FDR) at 0.05. Variables with FDR-adjusted p values <0.05 were further selected using Least Absolute Shrinkage and Selection Operator (LASSO) logistic regression with 5-fold cross-validation. Variables selected in ≥80% of folds were included in the final model. Data preprocessing included median imputation for missing values, binary encoding for categorical variables, and z-score standardization for continuous variables (i.e., subtracting the training-fold mean and dividing by the training-fold standard deviation). Standardization parameters were computed using each training fold and then applied to the corresponding test fold to prevent information leakage.

Model development and validation

Machine learning models were developed using the selected variables. Logistic regression (LR) and linear support vector machine (SVM) models were compared for their ability to predict impaired AVP secretion. Given the sample size and class distribution, no class-imbalance correction (e.g., class weighting or resampling) was applied. Model performance was assessed using nested cross-validation: outer 5-fold stratified cross-validation for performance estimation and inner 3-fold stratified cross-validation for hyperparameter tuning, ensuring complete separation of training and test data. Stratification was used to ensure that both positive and negative cases were evenly distributed between the training and test sets. The regularization parameter C for both LR and SVM models was optimized by systematically testing three values: 0.1, 1.0, and 10.0. SVM output scores were converted into probabilities, enabling receiver operating characteristic (ROC) analysis. In the sensitivity analyses, models excluding certain predictors (plasma AVP or U-Osm) were developed and evaluated under the same nested cross-validation framework.

Outcome measures

The primary outcome was the area under the ROC curve (AUC). Secondary outcome measures were the sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) at both the default (0.5) and optimized thresholds.

Feature importance analysis

SHapley Additive exPlanations (SHAP) was used to quantify the contributions of each variable to the predictions. SHAP LinearExplainer was applied to the best-performing model. Variable importance was ranked by the mean absolute SHAP values, and summary plots were generated to illustrate the relationships between variable values and predictive impact.

Clinical utility evaluation

Using the outer-fold predictions from the nested cross-validation in the overall cohort, we extracted predicted probabilities for the subgroup of patients who underwent the HST because of suspected AVP-D and evaluated model performance within this subgroup. As these probabilities were generated from held-out test folds in the nested cross-validation, no additional model training or refitting was performed for the subgroup analysis. Model performance was evaluated at two thresholds: the default threshold of 0.5 and an optimized threshold chosen to target a PPV ≥95%. The optimized threshold was defined as the lowest threshold at which the PPV reached ≥95% within this subgroup. Sensitivity, specificity, PPV, and NPV were calculated for each threshold.

Statistical analysis

Continuous variables are expressed as medians (interquartile range) and were compared using the Mann–Whitney U test. Categorical variables are expressed as numbers (percentage) and were compared using Fisher’s exact test. Two-sided p values <0.05 were considered statistically significant. The 95% confidence intervals (CIs) for the PPV and NPV were calculated using the Clopper–Pearson exact method. Analyses were performed in Python 3.12 using scikit-learn, SciPy, and related libraries.

Results

Patient characteristics

Of the 82 patients screened, 18 were excluded according to the exclusion criteria, leaving 64 patients for analysis (31 males and 33 females). All 64 patients underwent the HST: 35 for evaluation of suspected AVP-D and 29 for other clinical indications, such as assessment of AVP secretory capacity in patients with known hypothalamic–pituitary disorders (Fig. 1). In the suspected AVP-D subgroup (n = 35), the median daily urine volume was 5,330 (3,750–8,060) mL/day and the median spot U-Osm was 148 (96–250) mOsm/kg. Of the 64 patients analyzed, based on the HST results, 34 (53.1%) were classified as controls and 30 (46.9%) as having impaired AVP secretion. The control group included patients with heterogeneous underlying conditions (e.g., primary polydipsia and hypothalamic–pituitary disorders) and was defined operationally by failure to meet the criterion for impaired AVP secretion, rather than by a single reference diagnosis. Baseline demographic and clinical characteristics were comparable between the two groups (Table 1). Compared with the controls, patients with impaired AVP secretion had a significantly higher serum sodium level (143.0 vs. 140.5 mEq/L, p < 0.001), lower plasma AVP level (0.40 vs. 0.70 pg/mL, p < 0.001), lower U-Osm (227 vs. 539 mOsm/kg, p < 0.001), and lower blood urea nitrogen (BUN) level (10.1 vs. 14.7 mg/dL, p = 0.001) (Table 1). Among the patients classified as having impaired AVP secretion (n = 30), daily urine volume data were available for 28 patients, in whom the daily urine volume was 5,326 (3,288–8,954) mL/day. Among these 28 patients, 12 had mild to moderate polyuria (<5,000 mL/day), suggesting partial AVP-D.

Fig. 1. Schematic overview of the patient enrollment, diagnostic testing, feature selection, model training, and model evaluation for predicting impaired arginine vasopressin secretion. Of the 82 patients screened, 18 were excluded for the following reasons: suspected residual effect of desmopressin (n = 7), early termination resulting in an incomplete HST protocol (n = 4), peak serum sodium level <146 mEq/L during the HST (n = 5), and missing urinary osmolality data (n = 2). HST, hypertonic saline test; AVP, arginine vasopressin; LASSO, least absolute shrinkage and selection operator; ML, machine learning; LR, logistic regression; SVM, support vector machine; AVP-D, arginine vasopressin deficiency.

Table 1. Baseline characteristics of the patients analyzed (n = 64).

Variables	Controls (n = 34)	Impaired AVP secretion (n = 30)	p value
Sex, female	15 (44.1)	18 (60.0)	0.222
Age, years	55.5 (49.3–63.8)	53.5 (44.0–62.0)	0.798
History of TSS	8 (23.5)	7 (23.3)	1
Anterior pituitary deficiency	7 (20.6)	11 (36.7)	0.175
Blood values
Na, mEq/L	140.5 (140.0–142.0)	143.0 (142.0–144.0)	<0.001
K, mEq/L	4.0 (3.8–4.1)	4.0 (3.8–4.5)	0.326
Cl, mEq/L	106.0 (104.0–108.0)	106.0 (103.3–107.0)	0.496
Ca, mg/dL	9.2 (8.9–9.4)	9.2 (8.9–9.5)	0.945
BUN, mg/dL	14.7 (11.9–17.7)	10.1 (7.8–13.0)	0.001
Cre, mg/dL	0.72 (0.64–0.84)	0.71 (0.62–0.85)	0.652
eGFR, mL/min/1.73 m2	75.1 (63.0–88.4)	76.1 (64.6–92.7)	0.830
UA, mg/dL	4.8 (4.2–5.9)	5.6 (4.4–6.3)	0.172
Alb, g/dL	4.0 (3.7–4.1)	4.1 (3.6–4.3)	0.772
AST, IU/L	18.0 (15.3–24.8)	21.5 (18.3–33.8)	0.049
ALT, IU/L	18.0 (12.0–29.3)	22.5 (18.3–36.8)	0.114
WBC, /μL	5,500 (4,600–7,200)	7,000 (5,800–8,700)	0.019
Hb, g/dL	13.2 (12.3–13.8)	13.0 (11.4–14.5)	0.944
Plt, ×104/μL	23.7 (21.0–26.0)	25.1 (23.4–32.2)	0.040
Glucose, mg/dL	96 (90–114)	95 (86–105)	0.200
AVP, pg/mL	0.70 (0.50–0.90)	0.40 (0.27–0.60)	<0.001
Urine values
Osmolality, mOsm/kg	539 (427–615)	227 (117–387)	<0.001

Continuous and categorical variables are expressed as medians (interquartile range) and numbers (percentage), respectively.

TSS, transsphenoidal surgery; Na, sodium; K, potassium; Cl, chlorine; Ca, calcium; BUN, blood urea nitrogen; Cre, creatinine; eGFR, estimated glomerular filtration rate; UA, uric acid; Alb, albumin; AST, aspartate aminotransferase; ALT, alanine aminotransferase; WBC, white blood cell count; Hb, hemoglobin; Plt, platelet count; AVP, arginine vasopressin.

To assess the robustness of the outcome definition, we calculated the coefficient of determination (R²) and the SE for the estimated AVP level at a serum sodium level of 149 mEq/L for the individual AVP–sodium regression analyses (Supplementary Table 1). The median R² was 0.689 in the overall cohort, 0.807 in controls, and 0.414 in patients with impaired AVP secretion, and the respective median SE values were 0.145, 0.318, and 0.074 pg/mL (Supplementary Table 1).

Feature selection

Univariate analysis was performed to identify the variables significantly associated with impaired AVP secretion. Sex was included as a candidate variable in the initial screening but was not retained after feature selection. Of the 21 variables evaluated, the following 4 remained significant after Benjamini–Hochberg correction: U-Osm (FDR-adjusted p < 0.001), plasma AVP (FDR-adjusted p = 0.002), serum sodium (FDR-adjusted p = 0.003), and BUN (FDR-adjusted p = 0.006) levels. In the LASSO logistic regression, all four variables were selected in 100% of the cross-validation folds. These four variables were available for all 64 patients. Serum sodium, plasma AVP, and BUN levels were obtained immediately before HST initiation after an overnight fast in all patients. U-Osm was also obtained immediately before HST initiation in most patients; when unavailable, the most recent value available prior to the HST was used (6/64 patients, 9.4%).

Model performance evaluation

Model performance using the four selected variables is summarized in Table 2. The LR model yielded an AUC of 0.862, with a sensitivity of 83.3% and specificity of 82.4%. The linear SVM model yielded an AUC of 0.833, with a sensitivity of 76.7% and specificity of 82.4%. Both models showed good discriminative performance, but that of the LR model was superior. The ROC curves are shown in Fig. 2.

Table 2. Performance metrics of the machine learning models for predicting impaired AVP secretion.

Model	AUC	Sensitivity	Specificity	PPV	NPV
LR	0.862	83.3%	82.4%	80.6%	84.8%
Linear SVM	0.833	76.7%	82.4%	79.3%	80.0%

AVP, arginine vasopressin; AUC, area under the ROC curve; PPV, positive predictive value; NPV, negative predictive value; LR, logistic regression; SVM, support vector machine.

Fig. 2. Receiver operating characteristic (ROC) curves used to evaluate the performance of machine learning models in predicting impaired AVP secretion. AUC, area under the ROC curve; LR, logistic regression; SVM, support vector machine.

Because the plasma AVP level was used both as a predictor and as a component of the serial measurements used to define the outcome, we developed an LR model in which plasma AVP was excluded from the predictor set (retaining serum sodium, BUN, and U-Osm) under the same nested cross-validation framework. The discriminative performance of this three-variable model was maintained, with an AUC of 0.824, sensitivity of 76.7%, and specificity of 79.4% (Supplementary Table 2). To assess the incremental value of U-Osm, we evaluated another LR model in which U-Osm was excluded (retaining serum sodium, plasma AVP, and BUN). This model achieved an AUC of 0.820, with a sensitivity of 80.0% and specificity of 79.4% (Supplementary Table 3).

Feature importance

SHAP was used to quantify the contributions of each variable to the model predictions. The U-Osm had the highest variable importance (mean absolute SHAP value: 0.672), followed by the plasma AVP level (0.629), the serum sodium level (0.622), and BUN level (0.248) (Fig. 3a). Low U-Osm, low plasma AVP, high serum sodium, and low BUN levels were associated with predictions of positive cases (Fig. 3b).

Fig. 3. Feature importance for predicting impaired AVP secretion using Shapley Additive exPlanations (SHAP). (a) Ranking of variables by mean absolute SHAP values. (b) SHAP summary plot showing the distribution of SHAP values for each variable and the relationship between these values and their impact on the prediction. U-Osm, urinary osmolality; AVP, arginine vasopressin; Na, sodium; BUN, blood urea nitrogen; SHAP, Shapley Additive exPlanations.

Clinical utility in target populations

The performance of the LR model was evaluated in a subgroup of 35 patients who underwent the HST for suspected AVP-D, including 28 patients with impaired AVP secretion and 7 without impaired AVP secretion. At the default threshold (0.50), the sensitivity was 89.3% (25/28 patients), specificity was 71.4% (5/7 patients), PPV was 92.6% (25/27 patients; 95% CI, 75.7–99.1), and NPV was 62.5% (5/8 patients; 95% CI, 24.5–91.5). At the optimized threshold (0.67), the sensitivity was 50.0% (14/28 patients), specificity was 100% (7/7 patients), PPV was 100% (14/14 patients; 95% CI, 76.8–100), and NPV was 33.3% (7/21 patients; 95% CI, 14.6–57.0) (Table 3 and Fig. 4).

Table 3. Performance metrics of the LR model for predicting impaired AVP secretion in the 35 patients who underwent the HST for suspected AVP-D (28 with impaired AVP secretion and 7 without impaired AVP secretion).

Threshold	Sensitivity	Specificity	PPV	NPV
0.50	89.3%	71.4%	92.6% (75.7–99.1)	62.5% (24.5–91.5)
0.67	50.0%	100%	100% (76.8–100)	33.3% (14.6–57.0)

PPV and NPV are presented as % (95% CI). LR, logistic regression; AVP, arginine vasopressin; HST, hypertonic saline test; AVP-D, arginine vasopressin deficiency; PPV, positive predictive value; NPV, negative predictive value.

Fig. 4. Predicted probability distributions for impaired AVP secretion under two thresholds. (a) Default threshold of 0.5. (b) Optimized threshold of 0.67. Analysis was based on 35 patients who underwent the HST for suspected AVP-D (28 with impaired AVP secretion and 7 without impaired AVP secretion) using cross-validation predictions. AVP-D, arginine vasopressin deficiency; PPV, positive predictive value; NPV, negative predictive value.

Because the intended workflow prescreens patients based on U-Osm ≤300 mOsm/kg, we additionally evaluated an LR model excluding U-Osm (retaining serum sodium, plasma AVP, and BUN). At the default threshold (0.50), the sensitivity of this three-variable model was 82.1% (23/28 patients), specificity was 57.1% (4/7 patients), PPV was 88.5% (23/26 patients), and NPV was 44.4% (4/9 patients) (Supplementary Table 3).

Discussion

In this study, we developed and validated machine learning models to predict impaired AVP secretion during the HST using readily available clinical laboratory parameters. In the overall cohort (n = 64), the LR model achieved an AUC of 0.862. In the subgroup of patients who underwent the HST for suspected AVP-D, the LR model yielded a PPV of 100% at the optimized threshold.

Our LR model has two key features. First, it directly predicts the AVP secretion capacity during the HST. Based on Japanese diagnostic criteria [11], this model predicts whether the plasma AVP level will be <1.0 pg/mL at a serum sodium level of 149 mEq/L, allowing clinicians to infer HST results without performing the test itself. Second, the model requires only four laboratory values: serum sodium, plasma AVP, BUN, and U-Osm levels. Because the model relies solely on laboratory data, it requires no additional clinical information and enhances ease of implementation. In the validation among 35 patients who underwent the HST for suspected AVP-D, the PPV was 100% at the optimized threshold (0.67). Thus, this model can identify at least some patients with impaired AVP secretion without requiring administration of the HST.

A potential concern is that the plasma AVP level contributes to both the prediction (as a model input) and outcome definition (through serial AVP measurements used to estimate the AVP level at a serum sodium level of 149 mEq/L). However, dependence on any single AVP measurement is limited because the outcome is derived from the AVP–sodium relationship across five time points during the HST (0, 30, 60, 90, and 120 min). Furthermore, the model retained discrimination after excluding the plasma AVP level from the predictor set, with an AUC of 0.824, indicating that the remaining laboratory variables contribute meaningfully to the prediction. In addition, the fit of the individual AVP–sodium regressions differed between the groups (median R²: 0.414 in impaired AVP secretion group vs. 0.807 in controls), which likely reflects a blunted AVP response to increasing serum sodium levels under impaired AVP secretion rather than poor data quality. Importantly, the estimation uncertainty of the predicted AVP value at a serum sodium level of 149 mEq/L was small in the impaired AVP secretion group (median SE: 0.074 pg/mL) relative to the diagnostic cutoff of 1.0 pg/mL, supporting the robustness of the outcome definition based on the predicted AVP level at this sodium level.

The proposed diagnostic process for AVP-D using our prediction model is as follows. First, patients are evaluated using the standard diagnostic criteria for AVP-D, excluding HST results; these criteria include a daily urine volume ≥3,000 mL and U-Osm ≤300 mOsm/kg [11]. The prediction model is then applied to patients who meet these criteria. Using the optimized threshold (0.67), a positive prediction supports a diagnosis of AVP-D without necessarily requiring the HST, whereas a negative prediction indicates the need to undergo the HST for a definitive diagnosis. Among the 35 patients who underwent the HST for suspected AVP-D, 14 (40%) were classified as positive according to this threshold. These findings suggest that the HST may be avoided in approximately 40% of such patients, thereby reducing the need for invasive testing and mitigating patient burden (Fig. 5 and Graphical Abstract).

Fig. 5. Proposed diagnostic decision algorithm using the prediction model. Patients with a clinical suspicion of AVP-D (polyuria and polydipsia; urine volume ≥3,000 mL/day or ≥40 mL/kg/day and U-Osm ≤300 mOsm/kg) were evaluated using the LR model (inputs: plasma AVP, serum sodium, U-Osm, and BUN levels). A predicted probability ≥0.67 indicates a rule-in result (PPV, 100% [14/14 patients]; 95% CI, 76.8–100 in patients with suspected AVP-D), in which the HST may be deferred. A predicted probability <0.67 indicates the need to perform the HST for definitive diagnosis. AVP-D, arginine vasopressin deficiency; U-Osm, urinary osmolality; AVP, arginine vasopressin; Na, sodium; BUN, blood urea nitrogen; HST, hypertonic saline test; PPV, positive predictive value; CI, confidence interval.

Graphical Abstract.

Because the intended workflow prescreens patients based on U-Osm ≤300 mOsm/kg, the incremental value of including U-Osm as a predictor may appear limited. However, in sensitivity analyses excluding U-Osm from the model, performance in the suspected AVP-D subgroup was lower at the default threshold (Supplementary Table 3), suggesting that U-Osm provides complementary information even within a prescreened population.

SHAP was used to evaluate the contribution of each laboratory variable to the model predictions. Based on mean absolute SHAP values, U-Osm, plasma AVP, and serum sodium were top contributors, whereas BUN showed the lowest contribution. Notably, however, BUN was selected in 100% of the cross-validation folds during feature selection, suggesting that it provides information complementary to that provided by the other predictors. Serum sodium and U-Osm levels were also included in the model by Nahum et al. [10], underscoring their roles as markers of the physiological response to impaired AVP secretion. BUN, which was not incorporated in the model by Nahum et al., emerged as a meaningful predictor in our analysis. This is consistent with the known pathophysiology of AVP-D, which results in decreased urea reabsorption and reduced BUN levels [12, 13]. Collectively, these results indicate that our model captures physiologically coherent patterns consistent with the known pathophysiology of AVP-D.

Contemporary AVP-D diagnostics increasingly rely on copeptin-based stimulation tests (hypertonic saline-stimulated copeptin and arginine-stimulated copeptin), particularly outside Japan, because copeptin is more stable and more practical than direct AVP measurement in many settings. The diagnostic performance for differentiating AVP-D from other causes of polyuria-polydipsia syndrome has been reported for these approaches: 93.2%/100% sensitivity/specificity for hypertonic saline-stimulated copeptin [3] and 93%/92% sensitivity/specificity for arginine-stimulated copeptin [8]. Copeptin is strongly correlated with plasma AVP (Spearman’s r ≈ 0.8) [14], suggesting that copeptin could potentially be used in lieu of plasma AVP in prediction models. In addition, U-Osm and urinary specific gravity were strongly correlated in our cohort (r = 0.95), suggesting that urinary specific gravity can serve as an alternative marker in facilities without osmometry.

In Japan, AVP-D diagnosis has traditionally been supported by direct plasma AVP measurement using a highly sensitive RIA during the HST, whereas copeptin testing is not routinely available in daily clinical practice. Accordingly, our prediction model is intended to be a complementary decision-support tool in this setting: it aims to identify a subset of patients with suspected AVP-D in whom the HST may be deferred (rule-in at the optimized threshold), while reserving the HST for those with negative predictions or diagnostic uncertainty (Fig. 5). Future studies should evaluate model performance in cohorts managed under copeptin-based diagnostic pathways and assess whether copeptin can substitute for plasma AVP within the model.

This study has several limitations. First, as a single-center retrospective study, the generalizability of the results at other institutions is uncertain. External validation using multi-center datasets in more diverse patient populations will be essential to establish the broader applicability of the model. In addition, generalizability may be affected by population differences (including ethnicity) and by the referral bias inherent to a tertiary academic center. To mitigate the risk of overfitting due to the modest sample size (n = 64), we employed nested cross-validation, a robust technique that separates hyperparameter tuning from performance evaluation [15-17]. Second, we selected LR and linear SVM because these algorithms perform reliably in small datasets; more complex ensemble or deep learning methods were not used because of their higher risk of overfitting under such conditions [18]. Third, the diagnostic cutoff used to define impaired AVP secretion in this study (predicted plasma AVP level <1.0 pg/mL at a serum sodium level of 149 mEq/L) was established using the only commercially available AVP RIA in Japan (YAMASA Corporation). The study that originally reported the diagnostic accuracy of this cutoff during the HST employed the same AVP assay [9]. Accordingly, the direct applicability of this cutoff to other AVP assays has not been established. Future studies are needed to determine whether the cutoff of <1.0 pg/mL is applicable to different AVP assays and how corresponding cutoffs should be determined for other assay platforms or copeptin-based diagnostic approaches. Finally, because only a small number of patients tested negative for impaired AVP secretion in key subgroups (e.g., only 7 of 35 patients who underwent HST for suspected AVP-D), it was challenging to reliably determine how often the model correctly identifies negative cases, making accurate estimation of the NPV difficult. Therefore, the model is better suited as a rule-in rather than rule-out tool.

Regarding clinical implementation, the model requires all four predictors (serum sodium, plasma AVP, BUN, and U-Osm) to generate a prediction; therefore, it cannot be applied when any of these values are unavailable. Because the model uses routine laboratory parameters, implementation would not require specialized equipment beyond standard testing. For clinical application, the levels of each predictor should ideally be obtained from a standardized early-morning fasting sample when feasible because these laboratory measures can vary with hydration status, time of day, posture, and medication use.

In conclusion, we developed a machine learning model that predicts impaired AVP secretion using four routinely available laboratory parameters. The model demonstrated good discriminative performance, achieving a PPV of 100% at the optimized threshold. Notably, application of this model suggests that the HST can be avoided in approximately 40% of patients who meet the diagnostic criteria. Patients identified as positive for impaired AVP secretion by the model may be diagnosed without undergoing invasive testing. This diagnostic approach has the potential to reduce patient burden and contribute to more efficient and patient-centered evaluation of AVP-D.

Disclosure

Acknowledgements

Not applicable.

Author contributions

S.N. contributed to data acquisition, statistical analysis, manuscript drafting, and figure preparation. M.K., R.A., Y.K., T.M., T.H., T.K., M.S., T.O., S.I., H.S., and R.B. contributed to data collection and clinical interpretation. D.H. and H.A. contributed to study conception and design, and manuscript revision. All authors critically reviewed and approved the final manuscript and agree to be accountable for all aspects of the work.

Data Availability

The datasets generated and/or analyzed during the current study are not publicly available due to restrictions to preserve patient confidentiality but are available from the corresponding author on reasonable request.

Funding

The work submitted did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

Conflict of interest

S.I. and H.S. are members of Endocrine Journal’s Editorial Board.

Patent information

A patent application related to the prediction system described in this study has been filed in Japan (Japanese Patent Application No. 2025-153920, filed September 17, 2025; applicant: Nagoya University).

Ethics approval and consent to participate

This retrospective study involving human participants was in accordance with the ethical standards of the institutional research committee and with the 1964 Helsinki Declaration and its later amendments. The study was approved by the Institutional Review Board of Nagoya University (approval number: 2025-0107). Due to the retrospective nature of the study, the requirement for informed consent was waived; however, patients were given the opportunity to opt out via the hospital’s website.

Consent for publication

Not applicable.

Supplementary Material

Supplementary Tables