Artificial intelligence (AI)–based systems have been shown to reliably recognize cardiovascular disease risk1 and diagnose conditions such as diabetic retinopathy2,3 and melanoma4 from medical images. These advances in image-based medical diagnosis have been widely publicized in the media and similar tools have been approved by the US Food and Drug Administration (FDA). In April of 2018, the FDA approved the first AI device to provide screening decision for a disease (ie, diabetic retinopathy) without assisted interpretation by a clinician.5 Kanagasingam et al6 evaluated a similar approach—a convolutional neural network algorithm, a deep learning method—for identifying diabetic retinopathy from medical images in a primary care setting in Midland, Western Australia. Their system correctly classified the 2 severe cases captured in the data (193 patients with diabetes), and misclassified 15 (false-positives) individuals as having diabetic retinopathy. The number of patients needing to be reviewed by an ophthalmologist was less than 10%. These findings demonstrate the potential for these systems to support efficient and improved care, while also highlighting the need for rigorous evaluation in clinical settings.
Most deep learning algorithms require large data sets for training, usually consisting of thousands or millions of images. Medical data sets of this magnitude are typically expensive to produce and annotate. Individuals developing AI diagnostic tools might therefore rely on whatever data are available to produce initial results. However, certain deficiencies might not be evident until an AI diagnostic tool is evaluated in a clinical setting because data sets used for training have been carefully curated to remove imperfect data samples. For example, a system trained only on high-quality images might provide incorrect diagnosis when classifying low-quality images or images affected by sheen or other defects present in real-world clinical settings, as observed by Kanagasingam and colleagues.6 Also, evaluation of AI diagnostic tools in clinical settings will enable researchers and clinicians to ascertain its potential effect on patient outcomes and health care decisions. Problems identified can be corrected prior to deployment. Findings from these evaluations should also be published in peer reviewed literature to monitor progress and allow for comparison of different systems. There is currently a dearth of published studies on the evaluation of AI diagnostic tools used in clinical settings.
Of course, evaluating an AI diagnostic tool in a clinical setting does not guarantee generalizability of findings. The article by Kanagasingam et al6 is based on a single algorithm evaluated in a single health care location. The authors acknowledge this limitation. Moving from good initial performance to a device that can be used across varied clinical settings might not be feasible in some cases. For example, AI diagnostic tools hold significant potential for improving health care in low-resource settings and regions where adequate medical infrastructure is lacking. However, observations made in a clinical context in a health care setting in a developed country might not be reproducible in a low-resource setting. This highlights the fact that different geographic regions and clinical settings might require tailored tools. Furthermore, training an AI diagnostic tool on a single data set or clinical setting might lead to outcomes that are dependent on a particular type of device used in capturing images or overrepresentation of a particular symptom or demographic group.7
Although multiple studies have demonstrated that AI can perform on par with clinical experts in disease diagnosis, most of these tools have not been evaluated in controlled clinical studies to assess their effect on health care decisions and patient outcomes. While AI tools have the potential to improve disease diagnosis and care, premature deployment can lead to increased strain on the health care system, undue stress to patients, and possibly death owing to misdiagnosis.
Published: September 28, 2018. doi:10.1001/jamanetworkopen.2018.2658
Open Access: This is an open access article distributed under the terms of the CC-BY License. © 2018 Nsoesie EO. JAMA Network Open.
Corresponding Author: Elaine O. Nsoesie, PhD, Institute for Health Metrics and Evaluation, University of Washington, 2301 Fifth Ave, Ste 600, Seattle, WA 98121 (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.
Nsoesie EO. Evaluating Artificial Intelligence Applications in Clinical Settings. JAMA Netw Open. 2018;1(5):e182658. doi:10.1001/jamanetworkopen.2018.2658
Customize your JAMA Network experience by selecting one or more topics from the list below.
Create a personal account or sign in to: