“Machine learning may help identify patients who are highly likely to have impaired arginine vasopressin secretion using clinical laboratory data obtained before the test.” 

A machine learning model using four routinely available laboratory measurements predicted impaired arginine vasopressin secretion with good accuracy and identified a subset of patients with suspected arginine vasopressin deficiency who may be able to avoid hypertonic saline testing, according to a retrospective study published in Endocrine Journal.  

“The hypertonic saline test (HST) is highly informative for diagnosing arginine vasopressin (AVP) deficiency, but it is also burdensome for patients and requires careful medical supervision,” corresponding author Daisuke Hagiwara, MD, of the Nagoya University Graduate School of Medicine in Japan, said in an interview with AACE Endocrine AI. “Our study suggests that machine learning may help identify patients who are highly likely to have impaired AVP secretion using clinical laboratory data obtained before the test. With further validation, this approach may help reduce the need for invasive AVP stimulation testing in selected patients.” 

Researchers analyzed data from 64 patients who underwent HST between 2018 and 2024. Impaired AVP secretion was defined as a predicted plasma AVP concentration below 1 pg/mL at a serum sodium concentration of 149 mEq/L during HST, based on established Japanese diagnostic criteria. 

Of the 64 patients, 30 were classified as having impaired AVP secretion and 34 as controls. All 64 underwent HST: 35 because of suspected AVP deficiency (AVP-D), and 29 for other clinical indications, such as assessment of AVP secretory capacity in hypothalamic-pituitary disorders. 

Researchers evaluated 21 baseline clinical and laboratory variables, and 4 were consistently retained across all cross-validation folds: urinary osmolality, serum sodium, plasma AVP, and blood urea nitrogen. 

A logistic regression model was the best-performing model, achieving an area under the receiver operating characteristic curve (AUROC) of 0.862 with 83% sensitivity, 82% specificity, and positive and negative predictive values of 81% and 85%, respectively. A linear support vector machine model performed slightly less well with an AUROC of 0.833, 77% sensitivity, 82% specificity, and positive and negative predictive values of 79% and 80%, respectively.  

Patients with impaired AVP secretion had higher baseline serum sodium concentrations than controls (143 vs 140.5 mEq/L), lower plasma AVP concentrations (0.40 vs 0.70 pg/mL), lower urinary osmolality (227 vs 539 mOsm/kg), and lower blood urea nitrogen concentrations (10.1 vs 14.7 mg/dL). 

SHapley Additive exPlanations (SHAP) analysis identified urinary osmolality as the strongest predictor, followed by plasma AVP, serum sodium, and blood urea nitrogen. Lower urinary osmolality, lower plasma AVP, higher serum sodium, and lower blood urea nitrogen were associated with positive predictions for impaired AVP secretion. 

The model was also evaluated in a subgroup of 35 patients with suspected AVP deficiency (AVP-D). At the default probability threshold of 0.50, sensitivity was 89%, specificity was 71%, positive predictive value was 93%, and negative predictive value was 63%. Using an optimized probability threshold of 0.67 to prioritize confidence in positive predictions, the model achieved 100% specificity and a 100% positive predictive value, although sensitivity declined to 50% and the negative predictive value to 33.3%. At this threshold, 14 patients (40%) had positive model predictions, “suggesting that a meaningful proportion of patients may eventually be able to avoid the HST,” said Dr. Hagiwara. 

Sensitivity analyses showed that model performance remained relatively stable even when plasma AVP was excluded from the predictor set. A three-variable model incorporating serum sodium, urinary osmolality, and blood urea nitrogen achieved an AUROC of 0.824. Excluding urinary osmolality instead yielded an AUROC of 0.820. 

The study had several limitations, including its single-center retrospective design, modest sample size, and lack of external validation. The model was also developed using an AVP radio-immunoassay available in Japan, and the diagnostic cutoff may not be directly applicable to other assay platforms. In addition, only seven patients in the suspected AVP-D subgroup tested negative for impaired AVP secretion, limiting confidence in estimates of negative predictive value.  

The researchers noted that at this stage, the model is viewed as a potential decision-support tool rather than a replacement for established diagnostic testing. “A positive prediction may help clinicians identify patients who are highly likely to have impaired AVP secretion and in whom the HST might be deferred,” said Dr. Hagiwara. “In contrast, patients with negative or uncertain predictions would still require standard diagnostic evaluation, including the HST when appropriate. In the future, this approach could make the diagnostic workup of suspected AVP deficiency more patient-centered by reducing unnecessary invasive testing while preserving diagnostic accuracy.” 

Researchers Shintaro Iwama, MD, and Hidetaka Suga, MD, disclosed they are members of the Endocrine Journal Editorial Board. The team shared that Nagoya University has filed a patent application related to the prediction system.