Becker's Clinical Quality & Infection Control

45 DATA ANALYTICS & AI Viewpoint: AI is still years away from transforming oncology By Andrea Park D espite regular headlines and study findings claiming that artificial intelligence is already revolutionizing healthcare, in actuality, there is still much to be done before the technology has a real impact on routine clinical practice. In an editorial published Oct. 1, the editors of Nature Reviews Clinical Oncology described the challenges that must still be addressed before AI use becomes widespread in healthcare, and specifically in oncology. One such challenge is a lack of standards in the development of AI tools, which will be solved primarily through "collaboration among regulatory bodies, technology developers and clinical staff," per the journal's editors. Another challenge is posed by the cost of advanced technology. ough many pro- ponents of AI claim the new solutions will only be pricey at first, before becoming more affordable over time, their implementation will still require both outright funding and potentially costly training and upskilling efforts. As a result, "widespread access to AI-based healthcare might not happen in the near future," the editors wrote. One final challenge comes from the prevailing idea that automation will provide clinicians with much more time for face-to-face patient interactions. More likely, however, AI will be used to ameliorate staff shortages, keeping cli- nicians' schedules much the same. Addition- ally, patients may feel that the use of AI in the clinical setting is more of a disruption than a facilitator of productive interactions. In conclusion, the journal's editors wrote, "the practical implications of using AI in rou- tine oncology practice are not yet completely understood." Its adoption will require not only solutions to the aforementioned chal- lenges and further research describing AI's benefits, but also "multidisciplinary expertise and, more importantly, the input of patients and their families and the cooperation of regulatory bodies." n VUMC researchers use AI to detect heart disease phenotypes By Jackie Drees N ashville, Tenn.-based Vanderbilt University Medical Center research- ers applied artificial intelligence to de-identified patient EHR data to identify sub-phenotypes of cardiovascular disease. Researchers published the study in the Journal of Biomedical Informatics. For the research, the VUMC team gathered 12,380 de-identified patient re- cords of individuals who had been diagnosed with CVD. Records extended at least 10 years prior to the patient's CVD diagnosis. After applying an automated scan of the data, researchers discov- ered 1,068 distinct patient phenotypes in the dataset. The team then used machine learning in conjunction with a technique called tensor decomposition to identify the long-term emergence of 14 distinct CVD patient subtypes. The researchers compared the subsequent heart attack rates among the six most prevalent sub-phenotypes. Results of the study showed that heart attack risk was noticeably different among the six most prevalent subtypes. Through an association analysis with estimated CVD risk for each subtype, researchers found that some phenotypic topics such as Vitamin D deficiency and depression could not be explained by conventional risk factors. Researchers concluded that because the six most prevalent sub-pheno- types presented noticeably different risks of subsequent heart attack, the topics may identify clinically relevant sub-phenotypes of CVD. n Mayo Clinic study finds AI can predict post-PCI hospitalization, death By Andrea Park A machine learning algorithm from Tel Aviv-based Medial EarlySign effectively predicted patients' risk of complications and readmission after undergoing percutaneous coronary intervention, according to study published in JACC: Cardiovascular Interventions. In the study, Mayo Clinic researchers applied the algorithm to a retrospec- tive analysis of data from the Rochester, Minn.-based medical center's PCI registry. The information included EHR, demographic and social data from nearly 12,000 Mayo Clinic patients, who had collectively undergone more than 14,000 PCIs. Compared to standard regression methods, the algorithm was proven to be a better predictor of mortality 180 days post-PCI and of 30-day rehos- pitalization for congestive heart failure. Additionally, the algorithm suc- cessfully identified patient subgroups at an elevated risk of other post-PCI complications and readmission. Medial EarlySign has developed several other machine learning-powered solutions. Most recently, the company partnered with Danville, Pa.-based Geisinger to develop and deploy a suite of new solutions to assess pa- tients' risk of contracting various high-burden diseases. n

• Cover