Patient-specific surrogate model to predict pelvic floor dynamics during vaginal delivery

• Type: Article in International Scientific Journal
• Date: 2024

Authors

Moura, R (Author) Other; Oliveira, DA (Author) Other; Parente, MPL (Author) Other; Kimmich, N (Author) Other; Hyncík, L (Author) Other; Hympánová, LH (Author) Other; Renato Natal Jorge (Author) FEUP

Journal

Title: Journal of the Mechanical Behavior of Biomedical Materials

Vol. 160

ISSN: 1751-6161

Publisher: Elsevier

Indexing

ISI Web of Knowledge - 0 Citations

Scopus - 0 Citations

Other information

Authenticus ID: P-017-056

DOI: 10.1016/j.jmbbm.2024.106736

Abstract (EN): Childbirth is a challenging event that can lead to long-term consequences such as prolapse or incontinence. While computational models are widely used to mimic vaginal delivery, their integration into clinical practice is hindered by time constraints. The primary goal of this study is to introduce an artificial intelligence pipeline that leverages patient-specific surrogate modeling to predict pelvic floor injuries during vaginal delivery. A finite element-based machine learning approach was implemented to generate a dataset with information from finite element simulations. Thousands of childbirth simulations were conducted, varying the dimensions of the pelvic floor muscles and the mechanical properties used for their characterization. Additionally, a mesh morphing algorithm was developed to obtain patient-specific models. Machine learning models, specifically tree-based algorithms such as Random Forest (RF) and Extreme Gradient Boosting, as well as Artificial Neural Networks, were trained to predict the nodal coordinates of nodes within the pelvic floor, aiming to predict the muscle stretch during a critical interval. The results indicate that the RF model performs best, with a mean absolute error (MAE) of 0.086 mm and a mean absolute percentage error of 0.38%. Overall, more than 80% of the nodes have an error smaller than 0.1 mm. The MAE for the calculated stretch is equal to 0.0011. The implemented pipeline allows loading the trained model and making predictions in less than 11 s. This work demonstrates the feasibility of implementing a machine learning framework in clinical practice to predict potential maternal injuries and assist in medical-decision making.

Language: English

Type (Professor's evaluation): Scientific

No. of pages: 14