Editorial: Machine Learning in Clinical Decision-Making
Amanda C. Filiberto, Ira Lewis Leeds, Tyler John Loftus · Frontiers in Digital Health · 2021
Machine Learning in Clinical Decision-MakingThe "Machine Learning in Clinical Decision-Making" Research Topic assimilates evidence and perspectives from researchers and thought leaders that are pursuing the safe, effective development, and clinical application of machine learning systems to augment clinical decision-making across a wide array of specialties including cardiology, neurology, audiology, intensive care, and oncology.This editorial summarizes key points from the Research Topic.Electronic health record (EHR) systems have become widespread amongst health care systems globally.The resulting EHR databases have generated large, heterogeneous datasets that offer new opportunities to design and implement smarter health care systems by minimizing manual data entry, introducing objectivity where hypothetical-deductive reasoning fails, and providing accurate predictions and classifications that tailor care to individual patients' needs.Despite the promising role for artificial intelligence (AI) techniques and technologies to improve patient care, several substantial barriers to clinical adoption remain.Many ethical dilemmas surround the use of AI in healthcare.Machine learning algorithms trained to optimize a certain endpoint may make medically sound recommendations without reflecting a patient's ultimate goals of care, which often differ from textbook medical outcomes.Additionally, algorithms trained on biased data sets may produce biased outputs, which may be detrimental if the target population demographics and other characteristics are misaligned.There is also a lack of transparency regarding how many AI algorithms arrive at predictions, which has important implications for care delivery.Currently, despite impressive efficacy in retrospective and observational studies, there is limited level I evidence supporting the use of health care AI for decision support, suggesting a need for more high-level evidence, especially randomized trials.Electrocardiography (ECG) is an efficient, easily accessible, and commonly performed method for screening and diagnosing cardiovascular disease, a major contributor to potentially preventable mortality and morbidity.Among cardiovascular diseases, heart failure is particularly difficult to recognize due to heterogeneity of underlying pathology and clinical manifestations. Grun et al. demonstrate the ability of AI to accurately predict heart failure from standard 12-lead ECGs, highlighting the potential for AI to promote early diagnosis and treatment using routine clinical data.Electroencephalography (EEG) is used in the diagnosis, monitoring, and prognostication of many neurological ailments including seizure, coma, sleep disorders, brain injury, and behavioral abnormalities.Similar to cardiovascular disease, these neurologic diseases are heterogenous and often present with diagnostic uncertainty.Saba-Sadiya et al. propose a flexible, unsupervised model that applies to novel EEG data for a variety of clinical decision tasks, including coma prognostication and neurodegenerative illness detection.This work represents an important