Medicine finds itself in an age of data. From electronic medical records to wearables, the capturing, storing, and synthesizing of data are becoming an important part in the changing landscape. Deep learning uses computer modeling, known as artificial neural networks, to identify complex relationships in large data sets, and to prospectively apply this knowledge to newly added data.1,2 The study by Lu et al3 presents a unique opportunity for channeling the power of deep learning.
In brief, the authors created a convolutional neural network to predict long-term mortality from chest radiograph findings. Testing of this model was done using screening radiographs from the Prostate, Lung, Colorectal, and Ovarian Cancer Screening Trial, with subsequent external testing from the screening radiograph arm of the National Lung Screening Trial. The authors were able to identify patients with an increased risk of mortality at 6 and 12 years using diagnostic findings and standard risk factors. This study expands the possibilities of deep learning from models that reliably predict short-term outcomes to those that could reliably predict outcomes in the distant future.
Study findings by Lu et al3 highlight one of the allures of deep learning: the prospect of identifying patients at risk for adverse outcomes and then trying to avoid that adverse outcome. But what use is this prediction if we do not yet know what to do with this information? What is a worthy preventive strategy? Lu et al3 point to this challenge, arguing that, for now, clinicians do not have much guidance in treating these patients beyond optimizing known risk factors for mortality. Their research speaks to the existing artificial intelligence chasm, defined as “the gulf between developing a scientifically sound algorithm and its use in any meaningful real-world applications.”2 This chasm is apparent in numerous other studies1,4 that have used deep learning for various clinical scenarios. The chasm’s existence also raises the questions, what are the boundaries of this chasm, and how does one successfully cross it? Prevention strategies have their own inherent costs and risks, and this must be accounted for when determining what is worth predicting and what should be done with that prediction.
Although studies have not yet addressed the problem of the artificial intelligence chasm, a growing body of research suggests that 1 solution lies in harnessing data in real time that may affect short-term outcomes. In a 2016 study, Taylor et al5 used deep learning to predict in-hospital mortality for patients with sepsis in the emergency department. This retrospective study5 of adults with emergency department visits for sepsis found that the machine learning model outperformed established clinical decision rules and previously validated analytic techniques for classification. This, along with other research,6 suggests that deep learning could be used to identify patients at increased risk of deterioration and prompt clinicians to consider more intensive monitoring and interventions. One can imagine more appropriate allocation of resources earlier during a hospitalization, resulting in improved outcomes, such as shorter length of stay, fewer unanticipated transfers to the intensive care unit, and lower short-term mortality.
Beyond guiding clinical decisions early in a disease process, deep learning also has the potential to help guide appropriate treatment at the end of life. Avati et al7 looked at how deep learning could potentially identify patients who may benefit from palliative care consultation. They used machine learning and electronic medical record data to identify patients at risk for death within the next 3 to 12 months. The palliative care team was then notified about patients who were screened, and the team performed medical record reviews to evaluate whether the clinical situations were appropriate for a palliative care consultation. Data gathered from deep learning models could help clinicians anticipate clinical needs in real time, respond accordingly, and engage patients and their families in informed decision making.
Briefly indulging in the artificial intelligence hype, one can imagine a world in which a patient admitted to the hospital for the management of sepsis undergoes an infectious diseases workup, including a chest radiograph. Although findings may be normal for acute disease, evidence of cardiomegaly and calcification of the aortic notch might trigger notification to the patient’s outpatient clinician to ensure optimization of cardiac risk factors after hospitalization. Meanwhile, inpatient clinicians are notified that the patient is at an increased risk for decompensation, and they institute interventions accordingly, including the more frequent reporting of vital signs or admission to a step-down unit. A prolonged hospitalization with numerous complications might trigger an automatic palliative care consultation to ensure that the patient and family are considering both short-term and long-term goals of care.
During each stage of a hospitalization (and, by extension, all levels of care), deep learning has the potential to augment the clinician’s judgment, which may lead to improved delivery of care. That hyperbole should be tempered by the reality that the technology we have thus far is not nearly as ambitious. The evidence is lurching toward the ability to better predict outcomes, but even the most optimistic artificial intelligence supporters must acknowledge that successful prevention of clinically significant adverse outcomes remains an even more distant possibility. We should not simply be intoxicated by the idea of what we can do but must be clear-headed about what is worth knowing and is worth doing.
Published: July 19, 2019. doi:10.1001/jamanetworkopen.2019.7447
Open Access: This is an open access article distributed under the terms of the CC-BY License. © 2019 Tsega S et al. JAMA Network Open.
Corresponding Author: Hyung J. Cho, MD, New York City Health and Hospitals, 125 Worth St, Room 507, New York, NY 10013 (email@example.com).
Conflict of Interest Disclosures: None reported.
Identify all potential conflicts of interest that might be relevant to your comment.
Conflicts of interest comprise financial interests, activities, and relationships within the past 3 years including but not limited to employment, affiliation, grants or funding, consultancies, honoraria or payment, speaker's bureaus, stock ownership or options, expert testimony, royalties, donation of medical equipment, or patents planned, pending, or issued.
Err on the side of full disclosure.
If you have no conflicts of interest, check "No potential conflicts of interest" in the box below. The information will be posted with your response.
Not all submitted comments are published. Please see our commenting policy for details.
Tsega S, Cho HJ. Prediction and Prevention Using Deep Learning. JAMA Netw Open. Published online July 19, 20192(7):e197447. doi:10.1001/jamanetworkopen.2019.7447
Customize your JAMA Network experience by selecting one or more topics from the list below.
Create a personal account or sign in to: