Critical Issues in Neonatal Artificial Intelligence and Machine Learning Applications
Tuesday, January 31, 2023 
Clinical decision support systems developed using artificial intelligence and machine learning approaches possess transformative potential in improving neonatal care.
FREMONT, CA: Recent advances in artificial intelligence and machine learning (AI/ML) in healthcare have created computer-based clinical decision support systems (CDSS). CDSS may utilize vast quantities of patient data and clinical knowledge to improve diagnostics, alert systems, illness management, prescription, and drug control, among other applications. The trend in CDSS has shifted from using basic thresholds and ad hoc rules for diagnosis to data-driven training, which allows the model to uncover embeddings within very complex and non-linear data to generate unique insights and methods. Despite this, neonatal providers frequently dispute the clinical utility of these systems due to their lack of interpretability and explainability, and numerous statistical learning algorithms continue to evolve to increase therapeutic relevance. Due to out-of-sample difficulties, AI/ML algorithms optimized for one class of data obtained from local systems may need to be more extensively disseminated. In addition, the deployment and spread of these models have been limited due to the immense implementation problems posed by the real-world dynamics of patient care.
AI/ML-enhanced CDSS can address a variety of distinct clinical difficulties faced by neonatal providers. Current AI/ML technologies have become essential components of an evidence-driven and active neonatology care management pipeline from prognosis to clinical management to drug interactions.
As a primary illustration, AI/ML algorithms may prevent excessive or unnecessary blood collection. A leading cause of anemia in neonates in neonatal intensive care units (NICUs) is blood loss from laboratory testing. Due to low sample volumes, infectious diseases such as sepsis are frequently overlooked even when blood is collected correctly. AI/ML algorithms can be trained to detect irregularities from even the most minor changes in clinical indications or to do multivariate analysis utilizing their computing capability. Consequently, AI/ML can be trained to employ physiological time-series measurements that are frequently collected as an alternative to intrusive testing.
Infants in the NICU frequently require continuous, long-term electroencephalography monitoring to diagnose subclinical brain impairment. In addition, because the EEG of neonatal patients differs from that of adults, a neurophysiologist with expertise in neonatal neurologic care must be present or accessible in the NICU. AI/ML has an advantage in recognizing even the smallest change in patterns and discovering fresh features to distinguish between categorization labels. AI/ML can thereby improve the process of continuous EEG monitoring and reduce human labor, optimizing the allocation of medical resources.
Lastly, newborn health is intrinsically linked to maternal health. AI/ML techniques can utilize the real-time data collected from newborns and perinatal markers. By completely exploiting both the mother's and neonate's clinical characteristics, the interrelationship between maternal health issues, birth characteristics, and neonatal outcomes may be easily depicted in decision-making.
Despite the promise for AI/ML to improve neonatal care practices, AI/ML applications aimed at neonates were rather restricted until very recently. In addition, there are still many significant obstacles to using these ideas in clinical settings.
