What You Need to Know About Machine Learning
1. What is machine learning in medicine? How does it differ from predictive modeling and computer-based algorithms?
The goal of machine learning is to use machines to detect patterns or correlations in data. In medicine, the specific purpose of machine learning-using the right data sets, approaches, and systems-is to unveil insights that convert into actionable knowledge for better, more timely patient outcomes.
Although the mining of data is not new, what sets machine learning apart is its nondirectional nature; there are no specific guidelines or explicit instructions to sign the way. Instead, on the basis of the data sets they are fed, machines are, in a sense, tasked with delivering their own findings. It's a nonlinear, interactive process that can handle complex quantities of data that would boggle the human mind, within a timeframe that defies human competition.
Where machine learning is open-ended, both predictive modeling and computer-based algorithms are more directional, linear, and rule-based. They are designed to provide specific responses in given circumstances-for instance, to identify a possible drug interaction, interpret an ECG, or deliver a result within the Framingham Risk Score.
In a commentary in the New England Journal of Medicine, Drs Zaid Obermeyer and Ezekiel Emanuel, both at Harvard Medical School, likened machine learning in medicine to a doctor in residency who seeks to learn rules from data. In contrast, they compare computer-based algorithms to a medical student who applies general principles to new patients.
2. How successful has machine learning in medicine been so far?
Machine learning has provided advances in varied fields, helping to create systems that detect potential fraud in banking and furthering the study of astronomy, for instance. A review of the literature reveals the extent to which machine learning is being applied to medicine, with recently published studies in epilepsy, dementia, and post-myocardial infarction survival rates, to name only a few.
The most important success of machine learning in medicine so far relates to its potential in diagnostics and prognostics. At this point, it's entirely plausible that machines will one day-perhaps sooner, rather than later-become accurate and rapid tools to identify physical anomalies, for instance. In this light, machine learning clearly has the potential to reshape the practices of radiology, anatomical pathology, dermatology, and oncology.
With its rapid expansion into medical fields, machine learning has engendered tremendous excitement and a growing set of expectations. Yet the science is still in its early stages and faces sizeable obstacles. In another recent New England Journal of Medicine commentary, Drs Jonathan Chen and Steven Asch of the Stanford University Department of Medicine warned about the risk for "hyping" the potential of machine learning and predictions in medicine, calling for the need to recalibrate expectations "beyond the peak of inflated expectations."
3. What are the current obstacles to machine learning in medicine?
Perhaps the greatest obstacles to the success of machine learning in medicine are the quality and appropriateness of the data it uses. It's a foregone conclusion that incomplete, inappropriate, biased, inconsistent, or otherwise faulty data will provide findings that are at best meaningless and at worst misleading and potentially harmful.
Fundamental questions relating to data sets for machine learning in medicine include:
- What constitutes the "right" data sets for a given area of study?
- Where do you obtain those data sets?
- How do you prioritize multiple sets of data within a given study?
- What constitutes too little or too few data? What are the consequences?
- What are the costs relating to data acquisition?
- Is there a danger of overgenerating data? What is the relative cost of creating data that might never serve a purpose?
- How should possible bias in data, such as patient selection, confounding indications, and inconsistent outcome data, be accounted for?
- How do you incorporate key "nonmeasurable" information, including social and cultural factors?
- How do you guarantee the security and anonymity of data?
Many of these issues are, of course, fundamental to medical research in general.
Other outstanding questions relating to machine learning in medicine include:
- How do you validate findings? By comparison with "the best" practitioners? By using clinical trials? What are the mechanisms? What are the delays?
- Medicine is a moving target; how do we build for the future, knowing that it will not necessarily look like the past?
- What mechanisms do we put in place to adopt findings into clinical practice?
- What about pragmatic issues relating to reimbursement and legal liabilities?
- Who takes responsibility for errors?
- What is the "black box" problem? In machine learning, for the time being at least, we might know "what" but not "how."
- What is the issue of providing predictions as opposed to identifying causes and necessary actions?
Fundamentally, for now, machine learning in medicine does not help to explain how or why illness occurs.
4. What is the potential for machine learning in medicine?
Despite the hurdles, machine learning in medicine is already ubiquitous. Its ability to turn data into knowledge has transformational potential that has already begun to reshape the accuracy of prognostics and diagnostics while pointing the way to upcoming changes throughout medicine. With its proven predictive possibilities, machine learning has the potential to accelerate timeframes and assume an increasingly important role, beyond the basic detection of anomalies. And as technologies and medical conditions become increasingly complex, machine learning will probably prove to be an invaluable tool-an aid, rather than a threat-for clinicians.
"Dr Eric Topol, Medscape editor-in-chief and director of the Scripps Translational Science Institute, believes that 'machine learning will get to the root cause of illness, even at the individual level."
Can we truly envisage more complicated models of machine learning to address root causes of illness? Dr Eric Topol, Medscape editor-in-chief and director of the Scripps Translational Science Institute, believes that "machine learning will get to the root cause of illness, even at the individual level. Once we have all the granular data from a person-biological, physiologic, anatomical, environmental, behavioral/social-and it's processed, we'll get there. But it'll take a while."
5. Will machine learning transform patient care and the practice of medicine?
The age of machine learning will probably necessitate adaptation and a potential shift in how medicine is practiced. The specifics are yet to be determined, but if machines can provide better outcomes, why wouldn't we want to use them?
The practice of medicine combines, of course, multiple skill sets. Perhaps machine learning will subsume the more technical side of the profession, freeing doctors to focus elsewhere? With this said, at this point, it's hard to tell how the cookie will crumble, and, as Drs Jonathan Chen and Steven Asch point out, it's essential to move forward with realistic expectations.
What about patients? Patients appear to have the most to gain from advances in machine learning. The role of the physician is undeniably central to patient care. However, if sophisticated models of machine learning succeed in improving diagnoses, prognostics, determination of causes of illness, and treatment options, we may well be entering a new age in healthcare.