Part of PLOS Medicine‘s 15th Anniversary celebration, Academic Editor Steven Shapiro discusses the contributions of the Machine Learning Special Issue to the validation of precision medicine and its potential use in clinical research and health care.
15 years ago the Public Library of Science (PLOS) took a bold, innovative, and altruistic stand, launching the first major open access medical journal – PLOS Medicine. The journal continues to take on big and tough issues as exemplified by the November 2018 special issue “Machine Learning in Health and Biomedicine.” As computational power increases exponentially, the capacity to (more affordably) handle, store, and analyze “big data” using machine learning (ML) will revolutionize science and medicine.
The power of ML is to find patterns among variables in large data sets rather than being programmed with rules. Models become more complex when they move from supervised (input and outputs have labels) to unsupervised (no labels), and when they move from linear regression with decision trees to neural networks (> 3 neural networks is termed deep learning). As the complexity increases so does one’s ability to “interpret” the data. In addition, ML can very precisely come up with an algorithm that describes the data. Sometimes too precisely, and when the model “overfits” the data, it may not be applicable to new datasets.
Yet, this methodology is at odds with hypothesis-driven, mechanistic science and well controlled clinical trials: the hallmarks of medical publishing. Vast amounts of real-world data served up through complex “black box” machine learning algorithms will be an oncoming tsunami. In the special issue, PLOS Medicine editors along with guest editors Suchi Saria, Atul Butte and Aziz Sheikh got ahead of it discussing the opportunities, challenges and laid the groundwork for scientifically robust use of ML. This was followed with original manuscripts in a variety of medical disciplines.
Criteria used for manuscripts published in this issue were that models derived from ML must be fit for the stated clinical purpose, and researchers must report on their efforts to validate the models with external datasets. The original articles displayed a broad array of uses that ML will have in medicine including improved diagnosis, predicting disease course (including complications and mortality), and informing population and public health. Hence, there is a mix of population health that attempts to reduce variation, and precision medicine that aims to add back variation at an individual level to determine one’s disease susceptibility, trajectory, and best treatment for each patient.
Each article uses a limited range of data input to address specific problems including raw image data and text/language recognition from the electronic medical record. Full mining of these domains and aggregating and harmonizing additional patient information including social determinants of health, behaviors, diet, and physiologic data from wearable devices will give an even fuller view of each patient’s health status. Precision medicine has largely grown up using genetics and other “omic” information, which were not a focal point of this issue. As we continue to better understand the clinical manifestation of genetic variants and other “omic” features in common diseases, these will be important inputs to add as we create integrated holistic models resulting in a deeper understanding of clinical prediction. And, as causal inference models are added, in silico disease pathways amenable to discovery and intervention will arise.
Achieving these goals also requires another type of integration, one where computer scientists and clinicians/healthcare professionals (as well as social scientists and others) work hand in hand during all phases of the development, validation, and implementation of these algorithms. The complexity of the processes requires a multidisciplinary approach as science is becoming too complicated for individuals to be a jack of all trades. The final integrator will be modern-day clinical trialists who must rigorously, thoroughly, and unrelentingly perform prospective validation and study of how these models perform in the real world, the true test of advancing healthcare.
The state of healthcare is at a crossroads. Despite remarkable scientific advances that are starting to translate into improved patient outcomes, the business model is failing with exorbitant costs (particularly in the U.S.) that are neither sustainable nor clearly related to quality. To solve these problems we must move reimbursement of health services from volume (fee for service) to value (high-quality affordable care). The best way to achieve this, and the future of medicine, is to deliver individualized care through precision medicine powered by ML.
Steven D. Shapiro, MD is Executive Vice President, Chief Medical and Scientific Officer (CMSO) and President, Health Services Division at UPMC, which operates 40 hospitals and employs more than 4,800 physicians. Dr. Shapiro received his medical degree in 1983 from the University of Chicago and completed an internal medicine residency, chief residency and fellowship in respiratory and critical care at the Washington University School of Medicine. In 2001 he was named the Parker B. Francis Professor of Medicine at Harvard Medical School and appointed the Chief of the Division of Pulmonary and Critical Care at Brigham and Women’s Hospital. In 2006 Dr. Shapiro served as the Jack D. Myers Professor of Medicine at the University of Pittsburgh School of Medicine/Chair, Department of Medicine at UPMC, prior to being named UPMC’s CMSO in 2010.
Dr. Shapiro is a physician-scientist who remains active clinically and at the lab bench. His research is focused on novel molecular pathways of inflammation, tissue destruction, and host defense in chronic obstructive pulmonary disease (COPD), infectious diseases, vascular disease, and lung cancer. His efforts have led to several new potential therapies that are in clinical trials. He also serves on PLOS Medicine’s Editorial Board.
Image Credit: StockSnap, Pixabay