A multimodal deep learning model for predicting severe hemorrhage in placenta previa

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Obstetrics and Gynecology

Background:

• Placenta previa can cause severe, life-threatening hemorrhage.
• Accurate prediction of hemorrhage is crucial for risk stratification and intervention planning.

Purpose of the Study:

• To develop and evaluate a multimodal deep learning model for predicting severe intraoperative hemorrhage in placenta previa.
• To compare the model's performance against traditional machine learning methods and human expert performance.

Main Methods:

• A multimodal deep learning model was constructed using MRI T2-weighted images of the placenta and patient tabular data (demographics, preoperative blood tests).
• The model was trained to predict intraoperative blood loss exceeding 2000 ml.
• Performance was evaluated against machine learning models using only tabular data or MRI images, and against two human expert obstetricians.

Main Results:

• The multimodal deep learning model achieved the highest accuracy (0.68) and AUC (0.74) in predicting severe hemorrhage.
• Machine learning models using only tabular data or MRI images showed lower accuracies (0.61 and 0.53, respectively).
• Human experts had a median accuracy of 0.61.

Conclusions:

• Multimodal deep learning effectively integrates diverse data types (imaging and tabular) for hemorrhage prediction.
• The developed model demonstrates potential as an assistive tool for human experts in predicting intraoperative hemorrhage associated with placenta previa.