OC14.01: *Constructing small‐for‐gestational‐age prediction models before 32 weeks of gestation with machine learning

Abstract

To develop machine learning prediction models with multimodal clinical data and compare the predictive performance for SGA with datasets before 32 weeks' gestation. This retrospective study included singleton pregnancies and the pregnancy data before 32 weeks' gestation were categorised into three datasets: first, second and early third trimesters (T1, T2, T3). The LightGBM framework was utilised to assess the variable importance by employing a five-fold cross-validation. Seven machine learning algorithms were used with an 8:2 ratio for training and testing. The model performance was evaluated using receiver operating characteristic curve (ROC) and sensitivity of test sets at the false positive rate of 10%. We included data of 4,394 women with 148 SGA. Among them, 1353, 2705 and 2964 women had T1, T2 and T3 ultrasound screening (SGA = 41, 82, 93). Seven datasets were generated after merging. The top twenty predictors were incorporated into machine learning. Areas under the ROC of ultrasonic and multimodal clinical prediction models were better than clinical prediction models. In T1, the ultrasonic prediction model performed well (AUC of 0.74, sensitivity of 0.50) with best predictor of the Crown–rump length (CRL). In T2 and T3, baseline characteristics and biochemical tests improved significantly performance based on ultrasound prediction models (AUC: 0.87 vs. 0.70, sensitivity: 0.56 vs. 0.25 in second trimester; AUC: 0.83 vs. 0.82, sensitivity: 0.63 vs. 0.42 in early third trimester). Combining ultrasound screening data of three stages, the model performed extremely excellent (AUC of 0.97, sensitivity of 0.88) with most predictive factor of ultrasound estimated birthweight in T3. Ultrasound screening in T1 is predictive for SGA, with CRL being closely relative. In T2 and T3, the multimodal prediction models have best sensitivity of SGA. Combining ultrasound screening data from three stages provides richer information, especially fetal growth trends, facilitating a better capture of complex relationships.

Ähnliche Arbeiten