Intrahepatic cholestasis of pregnancy: machine-learning algorithm to predict elevated bile acid based on clinical and laboratory data

• Published in: Archives of gynecology and obstetrics 2021-09, Vol.304 (3), p.641-647
• Main Authors: Asali, Aula, Ravid, Dorit, Shalev, Hila, David, Liron, Yogev, Eran, Yogev, Sabina Sapunar, Schonman, Ron, Biron-Shental, Tal, Miller, Netanella
• Format: Article
• Language: English
• Subjects: Acids; Bile; Endocrinology; Gallbladder diseases; Gynecology; Human Genetics; Laboratories; Maternal-Fetal Medicine; Medicine; Medicine & Public Health; Obstetrics/Perinatology/Midwifery; Pregnancy
• ISSN: 0932-0067; 1432-0711
• DOI: 10.1007/s00404-021-05994-z

Purpose Applying machine-learning models to clinical and laboratory features of women with intrahepatic cholestasis of pregnancy (ICP) and creating algorithm to identify these patients without bile acid measurements. Methods This retrospective study included 336 pregnant women with a chief complaint of pruritis without rash during the second/third trimesters. Data extracted included: demographics, obstetric, clinical and laboratory features. The primary outcome was an elevated bile acid measurement  ≥ 10 µmol/L, regardless of liver enzyme levels. We used different machine-learning models and statistical regression to predict elevated bile acid levels. Results Among 336 women who complained about pruritis, 167 had bile acids  ≥ 10 µmol/L and 169 had normal levels. Women with elevated bile acids were older than those with normal levels ( p  = 0.001), higher parity ( p  = 0.001), and higher glutamic oxaloacetic transaminase ( GOT) ( p  = 0.001) and glutamic-pyruvic transaminase (GPT) levels ( p  = 0.001). Using machine-learning models, the XGB Classifier model was the most accurate (area under the curve (AUC), 0.9) followed by the K-neighbors model (AUC, 0.86); and then the Support Vector Classification (SVC) model (AUC, 0.82). The model with the lowest predicative ability was the logistic regression (AUC, 0.72). The maximum sensitivity of the XGB model was 86% and specificity 75%. The best predictive parameters of the XGB model were elevated GOT (Importance 0.17), elevated GPT (Importance 0.16), family history of bile disease (0.16) and previous pregnancy with ICP (0.13). Conclusion Machine-learning models using clinical data may predict ICP more accurately than logistic regression does. Using detection algorithms derived from these techniques may improve identification of ICP, especially when bile acid testing is not available.