Development of early prediction model for pregnancy-associated hypertension with graph-based semi-supervised learning

Seung Mi Lee, Yonghyun Nam, E. Choi, Young Mi Jung, Vivek Sriram, Jacob S. Leiby, Ja Nam Koo, Ig Hwan Oh, Byoung Jae Kim, Sun Min Kim, Sang Youn Kim, Gyoung Min Kim, Sae Kyung Joo, Sue Shin, Errol R. Norwitz, Chan‐Wook Park, Jong Kwan Jun, Won Kim, Dokyoon Kim, Joong Shin ParkScientific 2022 · DOI

TL;DR

This paper introduces a novel prediction model for pregnancy-associated hypertension (PAH) using graph-based semi-supervised learning (SSL) on routine clinical variables collected in early pregnancy. The model achieves superior predictive performance compared to current clinical guidelines and even a known biomarker (PlGF), offering a more accurate and accessible tool for early risk identification.

What Problem Does This Paper Address?

Pregnancy-associated hypertension (PAH) is a serious complication affecting 1-8% of pregnancies, leading to increased maternal and neonatal morbidity and mortality. While clinical guidelines exist to identify high-risk individuals for early intervention (like aspirin prophylaxis), these methods often suffer from low predictive accuracy. Existing advanced methods, such as using biomarkers like placental growth factor (PlGF) or ultrasound, can be costly and not universally accessible. Furthermore, real-world datasets often contain "unlabeled" data (patients lost to follow-up), which traditional supervised machine learning methods struggle to utilize effectively. This paper aims to bridge this gap by developing a highly accurate early prediction model for PAH using readily available clinical data and a machine learning approach that can leverage incomplete datasets.

Methods

The core of this research lies in the application of graph-based semi-supervised learning (SSL). Unlike traditional supervised learning, SSL can learn from both labeled (outcomes known) and unlabeled (outcomes unknown) data. The authors constructed a "patient network" where each node represents a pregnant woman, and edges represent the similarity between patients based on their clinical variables. This network structure allows the model to propagate label information across patients, effectively utilizing the information from those with unknown outcomes.

Before building the SSL model, the researchers employed four different feature selection methods (logistic regression, SVM, random forest with Gini, and random forest with entropy) to identify the most important clinical variables. They then aggregated the rankings from these methods to select the top 11 variables. The prediction model was developed using these selected variables, and its performance was compared against models using variables from clinical guidelines and all available routine variables.

Datasets

The study utilized a secondary analysis of a prospective cohort of 1742 healthy pregnant women enrolled before 14 weeks of gestation. After exclusions, the final dataset comprised 1404 women for the graph-based SSL analysis, including 1347 women with complete follow-up (labeled data) and 57 women with incomplete follow-up (unlabeled data). The data was split into a training set (971 patients) and a test set (376 patients) based on enrollment year.

Key Insights and Ideas

  • Leveraging Unlabeled Data is Crucial: The study highlights that incorporating unlabeled data (from patients lost to follow-up) significantly improves the predictive performance of the SSL model, demonstrating the power of semi-supervised approaches in real-world clinical scenarios.
  • Routine Variables Can Be Highly Predictive: The research successfully identified a set of 11 routine clinical variables that are more predictive of PAH than traditional risk factors recommended by clinical guidelines. This suggests that a wealth of predictive information is already being collected and can be better utilized.
  • Graph-Based SSL Outperforms Traditional Methods: The graph-based SSL approach, particularly when using the selected top variables, demonstrated superior performance in terms of AUC, sensitivity, and specificity compared to logistic regression, SVM, and random forest models.
  • Biomarkers Aren't Always Superior: The proposed model achieved a higher AUC than placental growth factor (PlGF), a known biomarker for early PAH prediction, suggesting that a well-constructed model using routine variables can be equally, if not more, effective and certainly more accessible.

Key Results

  • The graph-based SSL model using the top 11 selected variables achieved an Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.811 in the test set.
  • This model demonstrated higher sensitivity (72.7%) compared to clinical guidelines (45.5%) while maintaining similar specificity (80.0% vs. 80.5%).
  • The proposed model's AUROC (0.80) was significantly higher than that of placental growth factor (PlGF) (0.71) for the total study population.
  • The top 11 variables identified were primarily related to blood pressure in early and late first trimester, maternal BMI, maternal age, and history of preeclampsia.

Limitations and Future Work

  • External Validation Needed: The study acknowledges that while the model performed well, further validation with larger, diverse cohorts (across different ethnicities and races) is necessary to confirm its generalizability.
  • Cost-Effectiveness Analysis: Future research should compare the cost-effectiveness of this model against predictive models that incorporate additional markers like ultrasound indices.
  • Algorithmic Enhancements: The authors suggest exploring more advanced graph-based machine learning models, such as graph convolutional neural networks (GCNs), which could potentially further improve performance by better capturing patient features.
  • Handling Excluded Patients: The study excluded patients with abortion or previable preterm birth. Future work could explore methods to incorporate these cases, potentially through missing value imputation, to create a more comprehensive patient network.
  • Gestational Age at Diagnosis: Incorporating gestational age at diagnosis of PAH as an outcome variable could further enhance the model's clinical utility.

Summary updated

Abstract

Clinical guidelines recommend several risk factors to identify women in early pregnancy at high risk of developing pregnancy-associated hypertension. However, these variables result in low predictive accuracy. Here, we developed a prediction model for pregnancy-associated hypertension using graph-based semi-supervised learning. This is a secondary analysis of a prospective study of healthy pregnant women. To develop the prediction model, we compared the prediction performances across five machine learning methods (semi-supervised learning with both labeled and unlabeled data, semi-supervised learning with labeled data only, logistic regression, support vector machine, and random forest) using three different variable sets: [a] variables from clinical guidelines, [b] selected important variables from the feature selection, and [c] all routine variables. Additionally, the proposed prediction model was compared with placental growth factor, a predictive biomarker for pregnancy-associated hypertension. The study population consisted of 1404 women, including 1347 women with complete follow-up (labeled data) and 57 women with incomplete follow-up (unlabeled data). Among the 1347 with complete follow-up, 2.4% (33/1347) developed pregnancy-associated HTN. Graph-based semi-supervised learning using top 11 variables achieved the best average prediction performance (mean area under the curve (AUC) of 0.89 in training set and 0.81 in test set), with higher sensitivity (72.7% vs 45.5% in test set) and similar specificity (80.0% vs 80.5% in test set) compared to risk factors from clinical guidelines. In addition, our proposed model with graph-based SSL had a higher performance than that of placental growth factor for total study population (AUC, 0.71 vs. 0.80, p < 0.001). In conclusion, we could accurately predict the development pregnancy-associated hypertension in early pregnancy through the use of routine clinical variables with the help of graph-based SSL.
BibTeX
@inproceedings{lee2022development,
  author = {Seung Mi Lee and Yonghyun Nam and E. Choi and Young Mi Jung and Vivek Sriram and Jacob S. Leiby and Ja Nam Koo and Ig Hwan Oh and Byoung Jae Kim and Sun Min Kim and Sang Youn Kim and Gyoung Min Kim and Sae Kyung Joo and Sue Shin and Errol R. Norwitz and Chan‐Wook Park and Jong Kwan Jun and Won Kim and Dokyoon Kim and Joong Shin Park},
  title = {{Development of early prediction model for pregnancy-associated hypertension with graph-based semi-supervised learning}},
  booktitle = {Scientific},
  year = {2022},
  doi = {10.1038/s41598-022-15391-4},
}
Figure