Predictive efficacy of machine-learning algorithms on intrahepatic cholestasis of pregnancy based on clinical and
Jianhu He1,2, Xiaojun Zhu3, Xuan Yang1
1Information Center, Women's Hospital, Zhejiang University School of Medicine, Hangzhou, China.
Insights
Machine learning models can predict intrahepatic cholestasis of pregnancy (ICP) using clinical data. Recurrent Neural Network (RNN) and Deep Neural Network (DNN) show high accuracy for early ICP detection and improved outcomes.
Area of Science:
- Medical Informatics
- Obstetrics and Gynecology
- Machine Learning in Healthcare
Background:
- Intrahepatic cholestasis of pregnancy (ICP) is a pregnancy-specific liver disease requiring timely diagnosis for better perinatal outcomes.
- Predictive models for ICP can aid in early detection and management.
Purpose of the Study:
- To develop and evaluate machine learning models for predicting ICP using clinical and laboratory indicators.
- To compare the efficacy of different models in identifying ICP.
Main Methods:
- Retrospective analysis of 1092 pregnant women's data (537 ICP, 555 control).
- Two schemes were used: Scheme 1 with first-trimester data and Scheme 2 with data from at least two trimesters.
- Models developed included Support Vector Machine (SVM), Deep Neural Network (DNN), Xgboost (Scheme 1), and Recurrent Neural Network (RNN), Long Short-Term Memory Network (LSTM), Gated Recurrent Unit (GRU) (Scheme 2).
Main Results:
- Scheme 1: DNN achieved 79.40% accuracy, SVM 67.90%, and Xgboost 73.40%.
- Scheme 2: RNN achieved 90.50% accuracy, LSTM 86.60%, and GRU 89.40%.
- RNN and DNN demonstrated superior predictive performance.
Conclusions:
- Deep Neural Network (DNN) and Recurrent Neural Network (RNN) are highly suitable for predicting ICP.
- These models offer a convenient and available tool for medical staff to optimize ICP management.
- Improved prediction can lead to better therapeutic strategies and enhanced clinical prognosis for ICP.
Objectives:
Intrahepatic cholestasis of pregnancy (ICP), a condition exclusive to pregnancy, necessitates prompt identification and intervention to improve the perinatal outcomes. This study aims to develop suitable machine-learning models for predicting ICP based on clinical and laboratory indicators.
Methods:
This study retrospectively analyzed data from 1092 pregnant women, with 537 diagnosed with ICP and 555 healthy cases as a control. Two study schemes were devised. For scheme 1, 62 indicators from the first period of gestation were utilized to establish predictive models. For scheme 2, 62 indicators from at least two periods of gestation were utilized to establish predictive models. Under each scheme, three different machine-learning models were developed based on the Arya Privacy Computing Platform, encompassing Support Vector Machine (SVM), Deep Neural Network (DNN), and Xgboost for Scheme 1, and Recurrent Neural Network (RNN), Long Short-Term Memory Network (LSTM), and Gated Recurrent Unit (GRU) for Scheme 2. The predictive efficacy of each model on ICP was evaluated and compared.
Results:
Under Scheme 1, the cohort comprised 1092 pregnant women (537 with ICP, 555 healthy). The SVM model exhibited a sensitivity, specificity, and accuracy of 85.5%, 47.50%, and 67.90%, respectively, while DNN showed 65.70%, 92.70%, and 79.40%, respectively, and Xgboost achieved 65.60%, 81.90%, and 73.40%, respectively. In Scheme 2, 899 pregnant women were analyzed (466 with ICP, 433 healthy). RNN demonstrated a sensitivity, specificity, and accuracy of 97.60%, 82.10%, and 90.50%, respectively; LSTM presented 90.70%, 81.70%, and 86.60%, respectively; and GRU achieved 89.90%, 83.80%, and 89.40%, respectively.
Conclusion:
DNN and RNN are the two most suitable models to predict ICP in a convenient and available way. It provides flexible choice for medical staff and helps them optimize the therapeutic strategies to meet different clinical setting and improve the clinical prognosis of ICP.
