O-219 Reducing uncertainty in early pregnancy: Using clinical features and radiomics to develop a machine learning model to predict outcome

Abstract Study question Are there radiomic ultrasound features from the pregnancies where viability is unknown (PUV), which in combination with clinical features, may predict subsequent loss? Summary answer A machine learning method using radiomics alongside clinical features, can predict early pregnancy outcome. Further external validation is required. What is known already Miscarriage is the most common pregnancy complication causing significant psychological and physical morbidity to women and their partners Successful implantation, decidualisation and placentation in the first trimester are critical to pregnancy success. Currently specific ultrasound measurements are used as thresholds to definitively diagnose miscarriage. There are other ultrasound markers and clinical features which indicate likely miscarriage. Radiomic analysis quantifies high-dimensional tissue features that cannot be observed by direct human eye analysis. We propose applying machine learning methods to a combination of ultrasound radiomic and clinical features may predict miscarriage. Study design, size, duration A retrospective, multi-site study, included 500 cases of early pregnancy PUV collected from January 2021 to January 2023. Longitudinal ultrasound images were extracted and segmented, alongside clinical data, from Queen Charlotte’s and Chelsea Hospital (QCCH), London (n = 400, split 8:2 for training and validation) and St Mary’s Hospital (SMH), London (test data set n = 100). Participants/materials, setting, methods PUV cases were identified and included where longitudinal ultrasound images of the gestational sac of the pregnancy were available. These images were extracted and segmented to include the gestation sac and endometrial-myometrial border. Clinical data including age, ethnicity, gestational age, and symptoms were collected. A prediction model was developed comparing several machine learning methods. The area under the ROC curve (AUC), F1-score, and recall were used to assess model performance. Main results and the role of chance The QCCH and SMH data sets were well matched and consisted of 53.3% and 53.0% miscarriage cases by end of first trimester, respectively. The best performing PUV outcome (radiomics and clinical features combined) classification model (logistic regression model in combination with elastic net feature selection technique) for predicting miscarriage from early pregnancy PUV cases (PUVPS model); achieved a recall of 0.81 and an AUC of 0.91 (F1-score 0.84), 0.83 (F1-score 0.76) and 0.82 (F1-score 0.76) in the QCCH training, QCCH validation and SMH test set respectively. The clinical features analysed alone produced a model with test AUC score of 0.74, and the radiomics features analysed separately produced a model with test AUC score of 0.72. The data demonstrated that the combination of radiomics and clinical features did produce the best performing model to predict PUV outcome.This machine learning method is the first to use radiomics to predict early pregnancy outcome, and has been validated and externally tested. This study sets the stage for future trials to prospectively evaluate the performance of the model, which can then be used to help patients navigate the uncertainty of a PUV early pregnancy diagnosis. Limitations, reasons for caution The main limitation of this study is its sample size, which can make a ML model prone to overfitting. This is demonstrated in the observed drop in classification performance and calibration of the PUVPS model between the QCCH validation and SMH test set. Wider implications of the findings The next steps, to establish a large multi-centre cohort will look to overcome the challenges associated with relatively small datasets, to validate across diverse populations to ensure reliability, thus improve the overall calibration of this model and potential clinical translation. Trial registration number Yes