Several studies and academic articles recently published offer updates regarding the use of artificial intelligence (AI) in medicine. Topics span a range of applications and specialties, including endocrinology, neurology, radiology, cardiology, pain care, and more.
In the field of radiology, a paper in JAMA described a computer that developed algorithms which enabled it to accurately detect fractures on radiographs.1 The authors proposed that the less complex tasks of radiologists, such as the detection of lung nodules on computer tomography, could ultimately be delegated to AI.
Concerning neurology, an article in Movement Disorders explored the potential of machine-learning algorithms to transform data from wearable, sensor-based systems that quantify signs of Parkinson’s disease into meaningful information that could significantly improve research and clinical practice.2
Within cardiology, researchers at the University of Ioannina in Greece and the University of Lyon in France tested algorithms designed to classify patients with heart failure as medication adherent and global adherent (pertaining to medication, nutrition, and physical activity). Their results showed detection accuracy rates of 82% and 91% for each respective measure.3 Additionally, new research reported in JAMA found high specificity and sensitivity of algorithms designed to detect referable diabetic retinopathy.4
In response to that paper, Isaac Kohane, MD, PhD, chair of the Department of Biomedical Informatics and professor of Biomedical Informatics and Pediatrics at Harvard Medical School in Boston, and Andrew L. Beam, PhD, a postdoctoral fellow who works with Dr Kohane, published an editorial on the topic of transforming AI in to clinical care.5
Cardiology Advisor recently interviewed Dr Kohane to better understand the role of artificial intelligence in medicine (AIM).
Cardiology Advisor: Since readers may be unfamiliar with the use of AIM, could you describe what it means and a few of the top ways it is currently being used?
Dr Kohane: In many ways, the definition AIM has been slippery because some of the tasks that we used to think of as requiring uniquely human expertise have been effectively implemented as computer programs that are used routinely and therefore no longer seen as requiring “intelligence.” For example, in the past, if you would have told the average physician that computers could do a good job of characterizing arrhythmias in an ECG or interpreting gene expression patterns to predict breast cancer recurrence risk, they might have exclaimed, “It’s an expert!” But now we just think of those tasks as mechanical.
So, to paraphrase William Schwartz from a 1970 article, I would say that AIM is the use of computing as an intellectual tool to assist in clinical assessment and decision making.6 [Editor’s note: William B. Schwartz, MD, was one of the first pioneers in the field of AIM.] I already gave a few examples, but in our recent editorial we highlighted the success of machine learning methods and deep neural networks in particular at performing at the expert level in image classification—the classification of retinopathy in that specific article, but clearly similar work is being followed for a number of clinical imaging modalities.5
Cardiology Advisor: What are your thoughts on why, as mentioned in your editorial, the promise of AI to transform medicine has not yet fully materialized?
Dr Kohane: There are many causes, but at the root it is a failure on the part of the medical system and the medical educational system to recognize that at its core, medicine is an information and knowledge processing discipline. You take in data from the patient, assess it based on what you know from other patients and compiled knowledge sources, and then transmute that assessment into a set of therapeutic actions that will be continually reviewed and revised based on the patient’s response.
Because this view is not at the core of medical practice, the implementation and use of information technologies lags far behind what we experience on the web as informed consumers, for example on Amazon and Netflix, and the educational program does not see automation as a near-term, and certainly not present, companion to the clinical decision-making process.
Cardiology Advisor: What are some of the major benefits that AI could offer the field of medicine?
Dr Kohane: First of all, let me stipulate that the smooth integration into a natural clinical workflow is essential. We should not wish to reproduce the effect of health information technology such as electronic health records which takes doctors away from the already limited time they have with patients and distracts them from their core mission. But supposing that part is done right, then the benefits include, in no particular order:
- Clinical decision-making would be informed at every step by what is known broadly in medicine, from textbooks to up-to-date population analyses.
- Patients could get their entire assessment and plan automatically translated into patient language, to be revisited at will and often.
- The plans across multiple care providers could be automatically scanned for incompatibility and danger before the clinicians are even aware of the possibility.
- The entirety of the data volunteered by the patient and multiple questionnaires would be integrated into decision-making on a day to day basis–or even more frequently–rather than waiting for a call from a doctor or nurse or a clinic visit.
- Repurposing existing drugs for new indications could be inferred from the timelines of millions of patients undergoing various therapies for thousands of diseases.
Cardiology Advisor: What are the current implications of these developments for our clinician audience?
Dr Kohane: Just like taxis lost their guild advantage from medallions to the Uber disruption, we should expect AIM to disrupt those activities that are truly rote in medicine and yet for which we bill as if they are acts of higher-level “slow-thinking” cognition. From tracking growth to scanning X-rays to reviewing pathology slides, to estimating gestational age from an ultrasound to interpreting whole genome sequences for variants with clinical impact, AIM is going to change the value proposition of clinicians.
It therefore behooves us to start thinking about which parts of our practice are the most valuable to patients and payors. I would start with the value of our common sense and human empathy, but we all know some colleagues who are not superlative in that respect either. I do think there are many important activities that computers will have a hard time reproducing, but if we and our professional societies do not start thinking hard about what those are and investing in our human capital in those areas, then Dr Robot is going to come a lot faster than need be.
- Jha S, Topol EJ. Adapting to artificial intelligence: radiologists and pathologists as information specialists. JAMA.2016; 316(22):2353-2354. doi:10.1001/jama.2016.17438.
- Kubota KJ, Chen JA, Little MA. Machine learning for large-scale wearable sensor data in Parkinson’s disease: Concepts, promises, pitfalls, and futures. Mov Disord. 2016;31(9):1314-1326. doi:10.1002/mds.26693.
- Karanasiou GS, Tripoliti EE, Papadopoulos TG, et al. Predicting adherence of patients with HF through machine learning techniques. Healthc Technol Lett. 2016;3(3):165-170. doi:10.1049/htl.2016.0041.
- Gulshan V, Peng L, Coram M, et al. Development and validation of a deep learning algorithm for detection of diabetic retinopathy in retinal fundus photographs. JAMA. 2016;316(22):2402-2410. doi:10.1001/jama.2016.17216.
- Beam AL, Kohane IS. Translating artificial intelligence into clinical care. JAMA. 2016;316(22):2368-2369. doi:10.1001/jama.2016.17217.
- Schwartz WB. Medicine and the computer—the promise and problems of change. N EngI J Med. 1970;283:1257-1264. doi:10.1056/NEJM1970012032832305.
This article originally appeared on Endocrinology Advisor