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.