最近一周医学人工智能方向的高质量论文推荐

Background The use of artificial intelligence (AI) in medicine will generate numerous application possibilities to improve patient care, provide real-time data analytics, and enable continuous patient monitoring. Clinicians and health informaticians should become familiar with machine learning and deep learning. Additionally, they should have a strong background in data analytics and data visualization to use, evaluate, and develop AI applications in clinical practice. Objective The main objective of this study was to evaluate the current state of AI training and the use of AI tools to enhance the learning experience. Methods A comprehensive systematic review was conducted to analyze the use of AI in medical and health informatics education, and to evaluate existing AI training practices. PRISMA-P (Preferred Reporting Items for Systematic Reviews and Meta-Analysis Protocols) guidelines were followed. The studies that focused on the use of AI tools to enhance medical education and the studies that investigated teaching AI as a new competency were categorized separately to evaluate recent developments. Results This systematic review revealed that recent publications recommend the integration of AI training into medical and health informatics curricula. Conclusions To the best of our knowledge, this is the first systematic review exploring the current state of AI education in both medicine and health informatics. Since AI curricula have not been standardized and competencies have not been determined, a framework for specialized AI training in medical and health informatics education is proposed.

Artificial intelligence (AI) and deep learning are entering the mainstream of clinical medicine. For example, in December 2016, Gulshan et al1 reported development and validation of a deep learning algorithm for detection of diabetic retinopathy in retinal fundus photographs. An accompanying editorial by Wong and Bressler2 pointed out limits of the study, the need for further validation of the algorithm in different populations, and unresolved challenges (eg, incorporating the algorithm into clinical work flows and convincing clinicians and patients to “trust a ‘black box’”). Sixteen months later, the Food and Drug Administration (FDA)3 permitted marketing of the first medical device to use AI to detect diabetic retinopathy. FDA reduced the risk of releasing the device by limiting the indication for use to screening adults who do not have visual symptoms for greater than mild retinopathy, to refer them to an eye care specialist. This issue of JAMA contains 2 Viewpoints on deep learning in health care. Hinton4 explains the technology underlying AI and deep learning, using clinical examples. AI is the general term for imitating human intelligence with computer systems. Early AI systems represented human reasoning with symbolic logic. As computer processing and storage became more powerful, researchers developed machine-learning techniques to imitate the way the human brain learns. The first machine learning continued to rely on human experts to label the data the system trained on (eg, the diagnosis) and to identify the significant features (eg, findings). Machine learning weighted the features from the data. With continued advances in computational power and with larger data sets, researchers began to develop deep learning techniques. The first deep learning algorithms were “supervised” in that human experts continued to label the training data, and the deep learning algorithms learned the features and weights directly from the data. The retinopathy screening algorithms are an example of supervised deep learning. Hinton4 describes continuing development of new deep learning techniques, including ones that are completely unsupervised. He also points out that it is not feasible to see the features learned by deep learning to explain how the system reaches a conclusion. Naylor5 identifies 7 factors driving adoption of AI and deep learning in health care: (1) the strengths of digital imaging over human interpretation; (2) the digitization of health-related records and data sharing; (3) the adaptability of deep learning to analysis of heterogeneous data sets; (4) the capacity of deep learning for hypothesis generation in research; (5) the promise of deep learning to streamline clinical workflows and empower patients; (6) the rapid-diffusion open-source and proprietary deep learning programs; and (7) of the adequacy of today’s basic deep learning technology to deliver improved performance as data sets get larger. Factors 3, 4, and 6 are specific to deep learning; the other factors apply to other AI techniques as well. Artificial intelligence is a family of technical techniques in the same way the radiologic imaging tool kit includes flat images, computed tomography scans, and functional imaging such as magnetic resonance imaging. Advances in computational technology, computer science, informatics, and statistics improve existing techniques and make new techniques possible. The addition of deep learning to the AI family of techniques represents an advance similar in magnitude to the addition of the computed tomography scanner to the radiology tool kit. Each AI technique has strengths and weaknesses. Symbolic logic is self-explaining but difficult to scale.6 For example, knowledge engineers extract the logic by interviewing or observing human experts. Statistical techniques such as supervised deep learning scale, but are subject to bias in the training data, and the reasoning cannot be explained. Since deep learning systems are trained on data from the past, they are not prepared to reason in the way humans do about conditions that have not been seen before. In the future, unsupervised deep learning may reduce this gap between human intelligence and AI. The potential applications of AI in health care present a range of computational difficulty. Narrow tasks, in which the context is predefined, are relatively easy. Imageprocessing tasks such as recognizing the border of an organ to suggest where to cut off a scan, or highlighting a suspicious area in an image for the radiologist or pathologist, are examples of narrow tasks. Image analysis and diagnostic prediction tasks such as the diabetic retinopathy example are broader and harder, but doable with today’s technology. Very broad data analysis and pattern prediction tasks such as analyzing heterogeneous data sets from diverse sources to suggest novel associations are feasible today because the purpose is limited to hypothesis generation. Thinking in the way humans do—reasoning, for example, from a few observations to suggest a novel scientific framework as Einstein did with the theory of relativity—is beyond technology on the horizon. Clinicians should view the output of AI programs or devices as statistical predictions. They should maintain an index of suspicion that the prediction may be wrong, just as they Viewpoint pages 1099 and 1101 Opinion

Artificial Intelligence in Medicine is looking for novelty in the methodological and/or theoretical content of submitted papers. Such kind of novelty has to be mainly acknowledged in the area of AI and Computer Science. Methodological papers deal with the proposal of some strategy and related methods to solve some scientific issues in specific domains. They must show, usually through an experimental evaluation, how the proposed methodology can be applied to medicine, medicallyoriented human biology, and health care, respectively. They have also to provide a comparison with other proposals, and explicitly discuss elements of novelty. Theoretical papers focus on more fundamental, general and formal topics of AI and must show the novel expected effects of the proposed solution in some medical or healthcare field.