Sharpening the resolution on data matters: a brief roadmap for understanding deep learning for medical data

Abstract

Statistics has historically been the workhorse for analyzing clinical trials, building prediction models, and analyzing most types of medical data. However, recently large amounts of available healthcare data and increasing computing power have enabled techniques from the field of machine learning to play an ever larger role in clinical applications. The fields of machine learning and statistics in fact have many things in common, but they can appear at first glance to be quite different given the way each community describes its goals and techniques [ [1] Beam A.L. Kohane I.S. Big data and machine learning in health care. JAMA J Am Med Assoc. 2018; 319: 1317-1318 Crossref PubMed Scopus (506) Google Scholar ]. This false sense of difference is further heightened because much of the recent work in machine learning has been described under the umbrella term of “artificial intelligence” (AI). It should be noted that AI refers to a goal (ie, computers that behave “intelligently”) and does not in itself describe a method to achieve that goal. Much of the recent and rapid progress towards the goal of medical AI has been enabled by advancements in the subfield of machine learning known as deep learning [ 2 LeCun Y Bengio Y Hinton G Deep learning. Nature. 2015; 521: 436-444 Crossref PubMed Scopus (32521) Google Scholar , 3 Hinton G Deep learning—a technology with the potential to transform health care. JAMA. 2018; https://doi.org/10.1001/jama.2018.11100 Crossref PubMed Scopus (238) Google Scholar , 4 Schmidhuber J Deep learning in neural networks: an overview. Neural networks. 2015; 61: 85-117 Crossref PubMed Scopus (8573) Google Scholar ]. However, beyond the AI hyperbole, the successes afforded by the use of deep learning to date have been more modest and narrow in high-risk settings, such as medicine, but are nonetheless likely to expand in coming years [ [5] Topol E.J. High-performance medicine: the convergence of human and artificial intelligence. Nat Med. 2019; 25: 44-56 Crossref PubMed Scopus (1169) Google Scholar , [6] Ghassemi, M., T. Naumann, P. Schulam, and A. L. Beam. 2018. “Opportunities in machine learning for healthcare.” arXiv Preprint arXiv. Available at: https://arxiv.org/abs/1806.00388. Google Scholar ], though many challenges remain [ [7] Beam A.L. Manrai A.K. Ghassemi M Challenges to the reproducibility of machine learning models in health care. JAMA J Am Med Assoc. 2020; https://doi.org/10.1001/jama.2019.20866 Crossref PubMed Scopus (66) Google Scholar ]. The relevant literature stretches multiple decades and is rapidly growing, and we provide here a high-level overview of deep learning, with an emphasis on medical applications and natural language data. We offer a key insight for applications of deep learning in medicine, a dichotomy that presents both challenges and opportunities for real-world applications: Deep learning models can amplify biases and other issues in the underlying data, but those same models can be leveraged to uncover data issues and patterns that might not be readily discoverable via more parsimonious models.