Work

When Predictive Models Collide

Year: 2020

Type: article

Abstract: No trend is more visible in medicine than integrating machine learning and artificial intelligence (AI) into clinical care.Terms like "random forests" and "convolutional neural networks" that were previously the domain of computer science PhDs are now part of the vocabulary of both C-suite administrators and frontline clinicians.As the high-performance storage and computing power that underlie these models become a commodity, both the supply and demand for predictive models based on machine learning will only increase.Advances in genomics 1 and patient-generated health data will enrich predictive models, but will also increase model complexity even as they increase precision.How can clinicians and modelers prevent the messy accidents that could happen when these models collide? GENERALIZING PREDICTIVE MODELS TO COMPLEX PATIENTS][8] Yet, although models may perform well for a single disease or clinical situation, the circumstances of the laboratory where they are developed are rarely replicated in real-world clinical settings.Advanced cognitive computing solutions have been shown to have variable performance in the wild 9-12 and may require additional patient characteristics (especially genetic data) to perform optimally in certain diseases.

Source: JCO Clinical Cancer Informatics

DOI: 10.1200/cci.20.00024

Authors: Peter D. Stetson, Michael N. Cantor, Mithat Gonen

Institution: Memorial Sloan Kettering Cancer Center

Language: English

Topic: Artificial Intelligence in Healthcare and Education

Subfield: Health Informatics

Field: Medicine

Domain: Health Sciences

SDG: Good health and well-being

Open Access status: green

Funder: National Cancer Institute