# Development and validation of a multi-modality system combining radiomics and deep learning for predicting mid-pregnancy complications and enabling timely pregnancy care

## TL;DR

A new AI system combining ultrasound data with advanced machine learning improves early prediction of pregnancy complications like high blood pressure and diabetes.

## Contribution

A novel multi-modality AI model integrating radiomics and deep learning features from early ultrasound scans for predicting pregnancy complications.

## Key findings

- The combined model achieved the highest AUC of 0.987 in the training cohort and 0.963 in the test cohort.
- The model outperformed single-modality approaches in predicting hypertensive disorders and gestational diabetes.
- The system shows potential for non-invasive, early prediction to guide personalized prenatal care.

## Abstract

To improve the early prediction of hypertensive disorders of pregnancy (HDP) and gestational diabetes mellitus (GDM), we developed and validated an artificial intelligence (AI) model. This initiative was driven by the insufficient accuracy of current clinical tools. Our study aimed to determine whether integrating radiomics and deep learning features from first-trimester ultrasound scans could enhance predictive performance.

A total of 213 pregnant women who underwent ultrasound at 8 weeks of gestation were enrolled. Clinical data, radiomics features, and deep learning features were collected. Imaging features were selected using LASSO regression. Four predictive models were developed: a clinical model, a radiomics model, a deep learning model, and a fusion model combining all feature types. Model performance was evaluated on an independent test set using metrics including AUC, sensitivity, specificity, calibration, and decision curve analysis.

In the training cohort, all models demonstrated excellent discriminatory ability, with the combined model achieving the highest AUC of 0.987 (95% CI: 0.9733–0.9999), followed by the DLR model (AUC = 0.985). The clinical model (AUC = 0.941) and radiomics model (AUC = 0.939) also performed well. In the test cohort, the combined model maintained superior performance with an AUC of 0.963 (95% CI: 0.9152–1.0000), significantly outperforming all single-modality models. Overall, the combined model exhibited optimal and stable predictive performance across both training and test datasets.

This enables accurate early prediction of HDP and GDM. This non-invasive tool supports tailored prenatal care, with potential to improve outcomes. Further validation in diverse groups is needed.