Risk factors and machine learning prediction models for intrahepatic cholestasis of pregnancy.

BACKGROUND

Intrahepatic cholestasis of pregnancy (ICP) is a liver disorder that occurs in the second and third trimesters of pregnancy and is associated with a significant risk of fetal complications, including premature birth and fetal death. In clinical practice, the diagnosis of ICP is predominantly based on the presence of pruritus in pregnant women and elevated serum total bile acid. However, this approach may result in missed or delayed diagnoses. Therefore, it is essential to explore the risk factors associated with ICP and to accurately identify affected individuals to enable timely prophylactic interventions. The existing literature exhibits a paucity of studies employing artificial intelligence to predict ICP. Therefore, developing machine learning-based diagnostic and severity classification models for ICP holds significant importance.

METHODS

This study included ICP patients and some healthy pregnant women from Jiaxing Maternity and Child Health Care Hospital in China between July 2020 and October 2023. We collected clinical data during their pregnancies and selected the top 11 critical risk factors through univariable and lasso regression analysis. The dataset was randomly divided into training and testing cohorts. Thirteen machine learning techniques, including Random Forest, Support Vector Machine, and Artificial Neural Network, were employed. Based on their various classification performances on the training set, the top five models were selected for internal validation.

RESULTS

The dataset included 798 participants (300 normal, 312 mild, and 186 severe cases). Through univariable and lasso regression analysis, total bile acid, gamma-glutamyl transferase, multiple pregnancy, lymphocyte percentage, hematocrit, neutrophil percentage, prothrombin time, Aspartate aminotransferase, red blood cell count, lymphocyte count and platelet count were identified as risk factors of ICP. The AUCs of the selected top five models ranged from 0.9509 to 0.9614. The CatBoost model achieved the best performance, with an AUC of 0.9614 (95% confidence interval, 0.9377-0.9813), an accuracy of 0.9085, a precision of 0.8930, a recall of 0.9059, and a F1-score of 0.8981.

CONCLUSIONS

We identified risk factors for ICP and developed machine learning models based on these factors. These models demonstrated good performance and can be used to help predict whether pregnant women have ICP and the degree of ICP (mild or severe).