Predictive value of ultrasonic artificial intelligence in placental characteristics of early pregnancy for gestational diabetes mellitus

Front Endocrinol (Lausanne). 2024 Mar 13:15:1344666. doi: 10.3389/fendo.2024.1344666. eCollection 2024.

Abstract

Background: To explore the predictive value of placental features in early pregnancy for gestational diabetes mellitus (GDM) using deep and radiomics-based machine learning (ML) applied to ultrasound imaging (USI), and to develop a nomogram in conjunction with clinical features.

Methods: This retrospective multicenter study included 415 pregnant women at 11-13 weeks of gestation from two institutions: the discovery group from center 1 (n=305, control group n=166, GDM group n=139), and the independent validation cohort (n=110, control group n=57, GDM group n=53) from center 2. The 2D USI underwent pre-processed involving normalization and resampling. Subsequently, the study performed screening of radiomics features with Person correlation and mutual information methods. An RBF-SVM model based on radiomics features was constructed using the five-fold cross-validation method. Resnet-50 as the backbone network was employed to learn the region of interest and constructed a deep convolutional neural network (DLCNN) from scratch learning. Clinical variables were screened using one-way logistic regression, with P<0.05 being the threshold for statistical significance, and included in the construction of the clinical model. Nomogram was built based on ML model, DLCNN and clinical models. The performance of nomogram was assessed by calibration curves, area under the receiver operating characteristic curve (AUC) and decision curve analysis (DCA).

Results: The AUCs for the ML model in the discovery cohort and independent validation cohort were 0.91 (0.88-0.94) and 0.86 (0.79-0.93), respectively. And 0.65 (0.59-0.71), 0.69 (0.59-0.79) for the DLCNN, 0.66 (0.59-0.72), 0.66 (0.55-0.76) for the clinical model, respectively. The nomogram exhibited the highest performance with AUCs of 0.93 (0.90-0.95) and 0.88 (0.81-0.94) The receiver operating characteristic curve (ROC) proved the superiority of the nomogram of clinical utility, and calibration curve showed the goodness of fit of the model. The DCA curve indicated that the nomogram outperformed other models in terms of net patient benefit.

Conclusions: The study emphasized the intrinsic relationship between early pregnancy placental USI and the development of GDM. The use of nomogram holds potential for clinical applications in predicting the development of GDM.

Keywords: deep learning; early pregnancy; gestational diabetes mellitus; machine learning; placenta; ultrasound imaging.

Publication types

  • Multicenter Study

MeSH terms

  • Artificial Intelligence*
  • Diabetes, Gestational* / diagnosis
  • Female
  • Humans
  • Neural Networks, Computer
  • Placenta / diagnostic imaging
  • Pregnancy
  • Ultrasonics

Grants and funding

The author(s) declare financial support was received for the research, authorship, and/or publication of this article. This work was supported by the Guangzhou Basic Research Program Jointly Funded by Universities and Municipalities (grant numbers 202201020526).