William Hersh, MD, Professor and Chair, OHSU
Blog: Informatics Professor
Twitter: @williamhersh
Scarcely a week goes by without another study published of a deep learning algorithm that achieves accuracy comparable to or sometimes better than human experts. The vast majority of these studies focus on some application to diagnostic imaging or waveform interpretation. There is no question that the practice of medicine will be highly impacted by these systems, especially for practitioners in fields that directly use them, such as radiologists, pathologists, dermatologists, and so forth.
What about deep learning applied to other areas of medicine? One group of studies has applied deep learning to retrospective electronic health record (EHR) data. A number of studies have shown impressive abilities to use EHR data to predict or diagnose:
- Several dozen diseases [1]
- Length of stay, mortality, readmission, and diagnosis at two large medical centers [2]
- Prognosis in palliative care [3]
- 30-day readmission in heart failure [4]
- Patient mortality from coronary artery disease more accurately than traditional cardiovascular risk models [5]
- Early risk of chronic kidney disease in patients with diabetes [6]
- Many pediatric diagnoses at a major referral center [7]
- Clinical outcomes in rheumatoid arthritis [8]
One can see obvious use cases for these types of systems, such as being able to intervene when patients might fare poorly when hospitalized or are at risk for readmission or more serious outcomes after they are discharged. But making a faster or more accurate diagnosis from an imaging or waveform study is a different matter than trying to determine the best use of an algorithm that tells a clinical team that a patient may be heading toward a bad outcome. When do we apply the results of the system? Once we intervene, does this change the nature of future recommendations? These are fascinating research questions, but also big unknowns in terms of how to apply such data in the clinical setting.
I wrote back in 2014, and more recently in 2017, that these systems must go beyond their ability to predict and actually be used in the context of prescriptive systems that results in better outcomes for the patient and/or the healthcare system. I hope to see studies going forward where these sorts of systems become part of the patient care team, and lead to demonstration of Friedman’s Fundamental Theorem of Informatics [9], which is that humans aided by machines do better than humans or machines alone. A nice roadmap for such studies was recently published that notes the need for studies that include meaningful endpoints, appropriate benchmarks, transportable to other settings and systems, and including legal and ethical monitoring [10].
This sentiment is echoed in the excellent new book by Dr. Eric Topol, Deep Medicine [11]. Dr. Topol raises the notion of deep learning benefitting patient care in even additional ways. One could solve a problem that has vexed healthcare with the widespread adoption of EHRs, which is the introduction of a third entity into the patient-physician encounter, namely the computer. Clinicians now unfortunately spend too much time “feeding the beast,” especially when that beast distracts from the patient and has additional compliance and billing burdens, contributing heavily to our epidemic of burnout in clinicians [12].
Perhaps one area where deep learning might help reduce the clinician burden is in data entry. An intriguing new study was recently published by one of the same authors who contributed some of the EHR suites above, automatically charting symptoms from patient-physician conversations [13]. One also wonders whether the level of a patient visit for billing purposes, currently determined by the presence of various elements document in the medical record, could be replaced by other data easy to collect in the modern medical office, such as time spent with the patient, time with other aspects of care, and other requirements of care. While we may never be able to achieve a “computer-free” patient examination room, we will hopefully find ways to reduce its impact and burden.
Dr. Topol suggests a second major area of benefit for deep learning, which is monitoring patients in much more depth than our current approach to providing episodic healthcare. Of course, this must be done in ways that provide actionable information presented in ways that do not further overburden clinicians. But it is not beyond the pale to envision his view of algorithms the coalescence of -omics, personal sensors, physiological measurements, healthcare, and public health data coming together to give diagnostic, therapeutic, and prognostic advice. There are still many issues around cost, ethics, practicality, and so forth, so the benefits will need to outweigh the risks.
Thus, while we will likely see machine assistance in diagnosis from images and wave forms in the near future, the impact of what deep learning may do with patient data in the EHR and with the patient will likely take longer. I see exciting opportunities for research and development focus on how to prospectively determine how such systems fit into the workflow of patient interaction within and outside the healthcare system. A nice overview.
For all References, see original post.
This article post first appeared on The Informatics Professor. Dr. Hersh is a frequent contributing expert to HITECH Answers.