Year: 2024
Type: article
Abstract: A rapidly proliferating corpus of clinical research harnessing the power of machine learning has substantial implications for healthcare feature engineering. As a broad umbrella encompassing data preprocessing, quality control, transformation, and generation, feature engineering addresses a major bottleneck in the production of predictive models. Automated feature engineering systems are increasingly deployed at scale to meet the challenges of generating the vast quantity of predictive features necessary for successful, generalizable, and clinically useful machine learning systems. Such clinical feature engineering systems produce features that are applied in a predictive setting after the fact and not explicitly linked to clinical care, but nevertheless involve substantial risk. A principled examination of a clinical feature engineering system can be framed in terms of six core components: data governance; compliance with privacy and regulatory constraints; appropriate validation and prospective evaluation; consideration of data biases; the use of effective data-quality and preprocessing pipelines; and sound candidate feature generation, scoring, and selection strategies. Health systems typically possess an assemblage of rich and diverse, yet underutilized, information with the potential to contribute meaningfully to clinical prediction problems. Electronic health record (EHR) data, comprising clinical notes, laboratory values, medication orders, and procedure codes; over a decade’s worth of length and width dataset and point-of-care laboratory test results; continuous biosensor measurements; DNA sequencing data; biomarkers derived from imaging; and drug compounds targeting genotypes provide raw material for hundreds of prediction problems in diverse specialties. However, machine learning in healthcare exhibits a stunning lack of reproducibility: many predictive models fail to retain their accuracy in different cohorts, and those that do are seldom incorporated into routine clinical care. A significant bottleneck underlying this failure lies with the feature engineering step...
Source: Journal of Neonatal Surgery
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7219389935
