AI is poised to revolutionize use of medical data, but challenges remain.
There has been an explosion in investment and application of artificial intelligence (AI) in health care. Leveraging a combination of large volumes of digital medical data available after the advent of electronic medical records, increasing sophistication in medical testing, and computing power, AI is projected to revolutionize our use of medical data. Health care AI investment is booming, totaling $2.14 billion over 323 deals from 2012-2017 and has increased each year, including by 31% in 2016, according to a 2017 CB Insights research report.
When it comes to health care applications, AI is a topic that remains clouded in mystery. The Oxford Dictionary defines AI as “the theory and development of computer systems able to perform tasks that normally require human intelligence.” In health care, AI is not specifically defined but is generally thought of as using machine learning algorithms to complete tasks such as drug discovery, virtually assisting patients, and automating complex medical tasks such as analysis of diagnostic tests.
AI is already being applied to urologic practice in three general areas: radiologic and pathologic diagnosis, patient monitoring and telemedicine, and automating repetitive tasks. It is being tested in two other areas: data analytics and precision medicine, and surgical training and quality improvement (table 1). This article examines current and potential future uses of AI as well as its benefits and shortcomings.
Current applications in urology
Radiologic and pathologic diagnosesare some of the more publicized and prolific applications of AI. There has even been a term coined for it in radiology-radiomics-which uses data algorithms to analyze large volumes of quantitative features from medical images. Lu et al in a recent study designed a machine learning algorithm to identify potentially metastatic pelvic lymph nodes and compared the AI results to those of radiologists. Their machine learning model was comparable to radiologist reads with an area under the curve of 0.912, indicating a high level of concordance while doing so in 20 seconds per case versus the radiologist average of 600 seconds (Cancer Res 2018; 78:5135-43).
Among more urology-specific examples (table 2), Sun et al used a machine learning algorithm to predict intra-prostatic tumor location from multiparametric prostate MRI with 70%-87% accuracy (Australas Phys Eng Sci Med 2017; 40:39-49).
Donovan et al developed and validated an AI-based test that automated Gleason grade scoring and combined this with biomarkers to accurately predict post-op prostate cancer progression. They were able to predict clinical failure of treatment with a C-Index of 0.82, which is better than conventional nomograms (Prostate Cancer Prostatic Dis Aug. 7, 2018 [Epub ahead of print]). AI has the potential to increase speed of radiologic and pathologic reads, reduce costs, and uncover subtle disease characteristics that might otherwise go unnoticed by humans in the diagnosis of urologic disease.
Next:Patient monitoring and telemedicinePatient monitoring and telemedicine. Wearable technology has been growing annually, with estimates that approximately 18% of Americans use such devices at least once a month, according to research by eMarketer. The large volume of recorded data that inherently comes with this technology is beyond human ability to analyze and sift through unaided. Currently, we are able to monitor blood glucose, electrolytes, activity levels, arrhythmias, and oxygen saturation, all of which can cloud to electronic health systems and providers.
While each data point may not be overly onerous to analyze individually, as these sensors multiply, they outgrow our ability to integrate them all without AI assistance. AI is being applied to augment this task and may make the promise of telemedicine, often touted but not fully realized, more practical by adding a layer of security/monitoring and pulling in greater objective information.
Increased AI-guided telemedicine has also taken the form of “remote-presence” mobile robots like the Vita by InTouch Health (figure). Telemedicine robots provide video remote access to patients. The existing shortage of U.S. urologists will only deepen in the coming years, and greater use of telemedicine may blunt this shortage by allowing us to reach patients in clinic or on the wards in areas without urologists, see consults from afar, and even round on post-op patients.
Automated repetitive tasks. Another growing use of AI is in automation and completion of complex but repetitive tasks normally performed by humans. Robots are already completing complex tasks in hospitals in several capacities. TUG is an autonomous robot commercially available from Aethon and has been in use in hospitals around the world including at UCSF Benioff Children’s Hospital and some VA medical centers. TUGs can use elevators, open doors, and use a camera and laser-based system to guide themselves without a set track to intelligently deliver supplies, drugs, and food, and remove waste around the hospital.
AI supply chain computers are also coming into use that can automatically order supplies when low and change order amounts based on usage patterns so surgical materials are there when you need them.
Xenex has designed robots to sterilize ORs and other areas of the hospital using high-intensity UV light, causing cellular damage to bacteria, viruses, and spores already in use at Mayo Clinic and Ochsner Health System.
The machines described above can help urologists increase nursing and OR staff efficiency, decrease OR turnover times, and help ensure that needed supplies are in stock and where they need to be.
Next:Future applicationsFuture applications
Data analytics and precision medicine. Precision or individualized medicine is a hot topic in medicine today. IBM has recently partnered Watson, its AI system, with Memorial Sloan Kettering Cancer Center as well as Quest Diagnostics to study ways to identify and diagnose cancers earlier, conduct faster genomic analyses, and even guide cancer treatment. Personalized medicine requires complex integration of a vast array of data points that is inherently labor intensive. AI has the potential to bring “super nomograms” to the fore of medicine by merging diverse data such as imaging features, genomics, molecular markers, and clinical data into one integrated analysis.
Patel et al compared a Watson-based model to a traditional molecular tumor board model to identify oncologic genomic events. In 32% of the 1,018 study patients, the Watson-based model identified genomic events that were not identified via the standard molecular tumor board model, the majority of which were actionable by qualifying patients for previously unconsidered biomarker-based clinical trials. The study authors concluded that AI could rapidly decrease the time needed for genetic sequencing-guided cancer care as well as aid physicians in identifying clinical trial eligibility (Oncologist 2018; 23:179-85).
Surgical training and quality improvement. AI is also being used to improve innate human performance. Hung et al used AI to track and analyze instrument motion tracking and camera manipulation on Cartesian coordinates and other performance metrics on a da Vinci Surgical System during radical retropubic prostatectomy. The AI model was then able to use this data to predict clinical outcomes, including prolonged length of stay, with 87% accuracy. Additionally, they found some of the metrics measured by their AI model also helped predict patients’ ultimate length of Foley catheter duration (JAMA Surg 2018; 153:770-1).
In the future, this type of AI system could give instant feedback to surgeons regarding case performance and objective insights into surgical technique.
Next:Challenges and benefitsChallenges and benefits
While AI technology presents many opportunities, its application has been rocky to date, particularly in the areas of data mining and clinical decision support. Despite some successes, IBM’s Watson division has also been beset by shortcomings and failures. MD Anderson ended an oncology clinical decision support system powered in part by Watson in 2017 after investing $62 million from 2013 to 2017. Numerous reported problems were found in an audit, including inaccurate treatment advice and problems integrating and extracting data from different electronic health records, according to a March 8, 2017 article in Becker’s Hospital Review.
After other setbacks, the Watson health care division underwent a large round of layoffs in June of 2018. In spite of the challenges, it is worth noting that MD Anderson has not completely abandoned AI efforts and is considering new partners. Other hospital systems such as Cleveland Clinic and Memorial Sloan Kettering have continued to invest in Watson.
AI expansion in health care raises other concerns, including the gradual replacement of human jobs and issues with data security and HIPAA compliance. AI also presents new trust and liability issues; ie, who is responsible when things go wrong? In other endeavors, such as automated driving vehicles, which have led to fatalities, developers and legal experts are already being forced to confront this scenario.
The potential benefits of AI likely mean that it is here to stay, however. With increasing pressure to reduce health care costs nationally, the promise of cost and efficiency savings with AI-assisted health care will likely continue to drive investment. There is increasing peer-reviewed data demonstrating AI’s ability to factor in hundreds or thousands of data points quickly, well beyond human abilities. Analysis of large volumes of difficult to integrate data streams is likely exactly what is needed for individualized medicine to move forward.
Lastly, there is a hope that AI may improve providers’ satisfaction and quality of life by automating some tasks, allowing physicians to spend more time with patients.
The future of AI in health care is bright and we should expect to see continued investment in the technology, leading to more peer-reviewed studies and eventually more commercially available applications. AI could make urologists’ lives easier by improving diagnosis, individualizing medicine, expanding telemedicine, automating tasks, augmenting surgical training, and extending our reach in an era of increasing work force shortages.
Daniel Au, MD
J. Brantley Thrasher, MD
Dr. Au is a urology resident and Dr. Thrasher is professor of urology, University of Kansas School of Medicine, Kansas City.
Section Editor James M. Hotaling, MD, MS
Section Editor Steven A. Kaplan, MD
is assistant professor of surgery (urology) at the Center for Reconstructive Urology and Men's Health, University of Utah, Salt Lake City, and
is professor of urology, Icahn School of Medicine at Mount Sinai, New York.