Technological progress over the last decade has allowed for cheaper and faster data storage, more powerful computation, and greater automation via artificial intelligence. While in the corporate world these breakthroughs have allowed companies to unlock information about consumer behaviors, in medicine it has exacerbated physician burnout and moved the physician-patient interaction from the exam room to the computer screen.
Any physician who has had to review a ream of medical records after a hospital admission understands the frustration that electronic medical records (EMRs) bring to patient care. For example, I once reviewed 500 pages of hospital notes only to find that the entire admission could have been summarized in less than 1 page. In this way, technological progress has somehow failed to improve the quality of care and end-user experience in healthcare, in part largely because the benefactors of new technology aren’t the physicians who use the systems, but the hospitals and insurance companies. These organizations stand to benefit financially and thus emphasize billing over all other aspects of the EMR.
These issues are further exacerbated by growing time constraints that force physicians to simultaneously see more patients and increase documentation, both in shorter amounts of time. The result: less time to meet those documentations requirements and cut corners.
Physicians have turned to more automated features of the EMR to help them hit the quality care benchmarks set by the Centers for Medicare & Medicaid Services (CMS) and insurers while cutting down on documentation of their medical reasoning in favor of bulleted assessments and action plans. The result: more junk notes to sift through. Overflowing inboxes, abundance of useless alerts, poor peer-to-peer communication, and decreased time for physician and patient interaction add to the overwhelming discontent among physicians and create the framework for physician burnout.1
Machine learning and artificial intelligence may offer an improvement in both healthcare operations and patient care. Over the past decade, advancements in natural language processing and machine learning have yielded the ability to automate International Classification of Diseases (ICD) coding and estimate patient risks by analyzing the clinical documentation in the EMR.2 This could lead to decreased billing burdens for physicians, faster risk prediction models built into EMRs, and more time for patient interaction. Artificial intelligence, speech recognition software, and natural language processing may even lead to more automated charting — allowing physicians to turn back towards their patients and away from the computer screen, resulting in more positive interactions.1
However, automated models aren’t always perfect. For example, an automated risk assessment tool used by court systems in the United States has been found to mistakenly flag black prisoners as twice as likely to offend compared with white prisoners.1 Likewise, a risk prediction model for mortality in patients with pneumonia wrongly labeled patients with asthma at lower risk because the predictive model could not account for the fact that physicians were treating patients with asthma more aggressively by starting antibiotics earlier.1
These types of errors can lead to devastating consequences; physicians should view conclusions drawn by artificial intelligence as statistical predictions that may be incorrect. The accuracy and reliability of software-driven solutions in healthcare are heavily influenced by the fidelity of the data used in the learning models and by validating the model with external datasets. Since this data often comes from our EMRs, the same issues that challenge clinicians may also now pose problems for machine learning models. For example, issues such as copying and pasting of clinical notes may result in the perseverance of clinical findings and diagnoses that are non-existent or no longer relevant3 to the patient. This noise in the EMR datasets can result in abnormal, unexpected, and inaccurate output by artificial intelligence programs.
Luckily, there are scientists and clinicians working on an app for that. Utilizing advances in cluster and cloud computing, researchers are creating complex algorithms that can sort through and extract useful data from massive datasets; in other words, Big Data.4 These improvements in turn lead to improved speech recognition algorithms, improved natural language processing, and refined artificial intelligence algorithms that can produce “annotated” data sets that can be used for better decision making.
For the clinician this might mean a cleaner, more concise EMR. Imagine a future in which cameras and microphones, coupled with advanced artificial intelligence algorithms and deep neural networks, produce straight-to-the-point notes to get physicians up to speed in seconds, all while documenting and completing billing requirements as physicians examine and speak to their patients without any direct interaction with the computer.
For now, it would be satisfying if artificial intelligence could sort through that 500-page chart in milliseconds and produce a concise yet informative summary of what happened during the patient’s hospital course. We’re not there yet, but with the right goals in focus we will get there. Innovation and technology offer endless opportunities, but as we implement these advances into daily practice we need to remain cognizant of their limitations and how those flaws might have an impact on our decisions.
- Verghese A, Shah NH, Harrington RA. What this computer needs is a physician. JAMA. 2018;319(1):19-20.
- Liebovitz DM, Fahrenbach J. Counterpoint: Is ICD-10 diagnosis coding important in the era of big data? No. CHEST. 2018;153(5):1095-1098.
- Weiner MG. Point: Is ICD-10 diagnosis coding important in the era of big data? Yes. CHEST. 2018;153(5):1093-1095.
- Zhu L, Zheng WJ. Informatics, data science, and artificial intelligence. JAMA. 2018;320(11):1103-1104.