Possibilities of applying machine learning methods to improve the quality of prenatal diagnosis of congenital malformations: scoping review

Machine learning algorithms are used in many areas of medicine. Prenatal screening (PS) is no exception. Implementing machine learning techniques to evaluate PS results can help overcome the problems inherent in human analysis: reduce subjectivity and inter-expert variability when reading medical images, reduce examination time, and stratify pregnant women into risk groups with greater reliability. The scoping review was conducted to evaluate the diagnostic performance of machine learning technologies in PS. Twenty-seven relevant papers were identified by through PubMed, Cochrane and eLibrary databases. All included papers demonstrated the potential of machine learning methods to detect, classify, or predict of the risk of congenital anomalies. Interpreting medical images, machine learning allows to reduce the diagnostic time, improve its quality, ensure screening performance in remote areas or in conditions of staff shortage and to maintain sufficient sensitivity and specificity, regardless of the doctor's qualifications. Algorithms based on metabolomic analysis have advantages in accuracy and efficiency in predicting chromosomal anomalies. Clinical decision support systems based on factors of anamnesis and results of prenatal diagnostics can improve the prediction of congenital anomalies in the first trimester of pregnancy, both in terms of screening accuracy and in reducing the cost of the screening program. However, current evidence is mainly derived from the implementation of machine learning systems with low autonomy, and the authors of most of the studies included in the analysis describe a number of limitations that must be taken into account when implementing such solutions.