Four perspectives on the prospects of medical AI

Although artificial intelligence has not yet formed a large-scale and normalized application in the medical and health field and is still subject to some policy, legal and ethical restrictions, and constraints. It is undeniable that with the deepening of artificial intelligence research and the emergence of the shortage of medical resources, medical artificial intelligence applications will become more extensive and more in-depth. The industry will become more and more mature.

The application of AI in the medical and health field will become more and more extensive

With the rapid development of medical and health informatization, medical institutions, various medical and health service enterprises will generate medical and health data. These data include medical images, electronic medical records, and health files. Artificial intelligence technology can perform semantics on these big medical data. Analysis and data mining, and realize early warning or automatic diagnosis of some diseases. These applications are mainly reflected in nine subdivisions, including disease screening and prediction, hospital management, health management, medical imaging, electronic medical records/document analysis, virtual assistants, intelligent medical devices, new drug discovery, genetic research, and interpretation.

Medical AI products will be officially approved as medical devices

In the United States, the FDA (Food and Drug Administration) approved the world’s first artificial intelligence medical device IDx-DR in April 2018. On November 19, the domestically developed ECG artificial intelligence automatic analysis and diagnosis system “AI-ECGPlatform” independently developed by Lepu Medical was approved by the FDA, becoming the first domestic artificial intelligence ECG product approved by the US FDA. Up to now, the FDA has approved 12 pan-AI medical products for clinical applications.

Since AI medical devices require a large amount of high-quality, labeled medical data for model training and learning in the early stage, high-quality data is currently difficult to obtain due to various reasons. Some algorithm models are not accurate enough, leading to many AI products in practical applications, problems such as misdiagnosis, misdiagnosis, and missed diagnosis will occur. The famous IBM Watson artificial intelligence product is continually being questioned because of misdiagnosis and prescribing unsafe drugs. Top medical institutions such as the University of Texas MD Anderson Cancer Center chose to abandon the project after investing $67 million.

The State Food and Drug Administration (CFDA) issued a new version of the “Medical Device Classification Catalogue” in September 2017 to better regulate the rapidly developing intelligent auxiliary diagnostic products in China, it will be implemented on August 1, 2018. The new version of the catalog adds categories corresponding to artificial intelligence-assisted diagnosis, explicitly analyzing and processing medical and pathological images. According to the latest classification regulations, if the diagnostic software provides diagnostic recommendations through algorithms, it only has auxiliary diagnostic functions and does not directly give a diagnosis conclusion, it can declare a Class II medical device. If the lesion is automatically identified and provides clear diagnostic tips, it can be considered the third category of medical device. On November 19, 2018, the Medical Device Technology Evaluation Center of the National Medical Products Administration issued a notice to publicly solicit information from domestic and foreign companies that produce artificial intelligence medical device products. It indicates that it has begun to prepare for the country’s AI medical device approval. It is estimated that by 2020 at the latest, China will have the first batch of artificial intelligence-based medical device products and be used in medical diagnostic services.

If medical artificial intelligence companies want to take the path of hospital procurement, it is the only way to obtain certification from the National Food and Drug Administration. If they are going to certify three types of medical devices, many real clinical application data will provide a vast amount for its application. To this end, the current third-party evaluation agency for artificial intelligence products for medical devices has begun to work, including building a smart product evaluation database, establishing smart product evaluation standards, and precise evaluation procedures, to provide companies with a true and reliable evaluation environment.

The supervision of intelligent medical applications will become more standardized and strict

Since medical data involves the patient’s privacy, it is a very sensitive issue in terms of ethics and law. Compared with AI products in other fields, medical intelligent products and services, especially those related to diagnosis and treatment, are bound to have a particular impact on the decision-making of human doctors and experts. Once there is a misjudgment, the life and health of patients Will face serious threats.

In the United States, medical devices are divided into three categories according to their risk levels. Category III is high-risk medical devices, specifically intended to support or maintain human life or prevent human health damage. It may cause the potential failure of medical devices with a reasonable risk of disease or injury. According to US federal law, Class III equipment requires pre-market approval (PMA) in addition to general supervision. But at the same time, the FDA also encourages enterprise medical AI products to update and iterate, for which it has a new accelerated approval channel.

Domestic regulatory agencies are more stringent in reviewing the use of artificial intelligence technology to provide diagnostic functions. The “Medical Device Classification Catalog” defines medical AI products: if the diagnostic software provides diagnostic recommendations through its algorithms, only the auxiliary diagnostic function does not directly offer the diagnosis conclusion is reached, it shall be declared and certified by the Class II medical device; if the lesion is automatically identified and a clear diagnosis prompt is provided, the clinical trial certification management must be conducted by the Class III medical device. Besides, the former National Health and Family Planning Commission issued the “Management Specifications for Artificial Intelligence Assisted Diagnosis Technology (2017 Edition)” and “Management System Specifications for Artificial Intelligence Assisted Therapy Technology (2017 Edition)” in February 2017. The application of AI-assisted diagnosis and treatment puts forward very operational requirements.

In the future, the world still needs to promote the formulation and improvement of related medical field systems. Each country and region will accelerate the formulation of a series of institutional standards for artificial intelligence products according to their own laws, regulations, and ethical environments, including product development, production, evaluation, and pricing. As a regulatory authority, the current focus is to formulate scientific, reasonable, and clear product classification and grading standards.

Research on AI theory and technology will be more in-depth

The rapid development of artificial intelligence technology has achieved specific results in many research projects. However, the medical and health industry is different from other industries, and its requirements for the credibility and interpretability of experimental results are still very high. For example, data analysis based on deep learning on medical images, although this type of technology can achieve high accuracy, but the model itself is a “black box technology”, resulting in a lack of judgment basis for its results. It is often difficult for human doctors and patients to believe the experimental results. The reliability of the final product is difficult to put into actual use. Also, many research contents aim to classify and detect a single disease or a small number of diseases. The artificial intelligence analysis of multi-disease tasks requires further improvement of the algorithm. Its generalization ability is improved on the basis of ensuring the accuracy of model experiments. As far as hardware equipment in the medical field is concerned, compared with developed countries, developing countries lack core R&D technologies, their innovation capabilities are feeble, and there are more significant difficulties in deploying artificial intelligence.

Artificial intelligence is not only the frontier of computer science, but also a new direction in mathematics, software engineering, neuroscience and other disciplines. Its future development in the medical field requires the joint efforts of computer software and hardware experts, medical experts, and statisticians. Collaboration across disciplines and fields. On the one hand, use more mature artificial intelligence theories to improve the thinking and logical ability of each system module, so that the expert system can accurately and quickly provide diagnosis and treatment plans in the most complex environment; on the other hand, continue to strengthen the artificial intelligence technology practice enables it to have more vital learning, self-organization, generalization and training capabilities, and accelerate the transformation from “weak artificial intelligence” to “strong artificial intelligence”.