A Risk Prediction Model for Hospitalization in Patients with Hyperemesis Gravidarum Based on Traditional Logistic Regression Algorithms.

OBJECTIVE: To construct and test a hospitalization risk prediction model and nomogram in patients with hyperemesis gravidarum (HG), which can objectively estimate hospitalization risk at initial clinical presentation and assist early-stage risk-stratification for symptomatic hyperemesis gravidarum patients. METHODS: Retrospective analysis of clinical data of 685 pregnant women assessed at Qinghai Red Cross Hospital during the period between January 2018 and April 2025 was conducted, with 433 of them hospitalized (63.2%). Univariate logistic regression analysis (P < 0.10) was utilized to screen potential predictors. Variables that passed the screening criterion were then analyzed using multivariate logistic regression and LASSO regression to identify independent predictors and create a nomogram. The performance of the models was evaluated based on clinical utility, discrimination, and calibration in the form of decision curve analysis (DCA), area under the receiver operating characteristic curve (AUC), and calibration curves, respectively. RESULTS: Univariate logistic regression found nine variables that were significantly related to hospitalization risk, such as a history of gynecological surgery, urinary ketone levels, body mass index (BMI), total bilirubin (TBIL), thyroid-stimulating hormone (TSH), serum potassium, serum sodium, fasting blood glucose, and lipid levels (all P < 0.05). The following LASSO and multivariate logistic regression analyses revealed four independent predictors. The independent variables were BMI, serum potassium (K), and serum sodium (Na) that were linked to a lower risk of hospitalization (OR < 1, all P < 0.05), and high TBIL was linked to a higher risk (OR > 1, P < 0.05). The prediction model was found to have a high discriminative power with an AUC of 0.903 and good consistency between the observed and predicted results. The average absolute errors in the training and test sets were 0.028 and 0.034, respectively. DCA also showed good clinical applicability, with net benefits being positive at threshold probabilities of 6 to 95 in the training set and 9 to 98 in the test set. CONCLUSION: BMI, TBIL, serum potassium, and serum sodium are accurate predictors of hospitalization risk in patients with HG in a nomogram. This nomogram may serve as an objective adjunctive tool for risk-stratification and supports clinical decision-making, but it cannot replace comprehensive clinical assessment.