In-depth overview: the implementation status and future development of "artificial intelligence + medical"

Artificial intelligence (AI) is experiencing explosive growth, impacting many industries, and is bringing a whole new revolution to the healthcare industry. AI...

Feb 09,2021 | Lydia

Artificial intelligence (AI) is experiencing explosive growth, impacting many industries, and is bringing a whole new revolution to the healthcare industry. "AI+healthcare" has become a hot area and has generated great interest among academia, industry and regulators.

Today, Professor Jianxing He, President of the First Affiliated Hospital of Guangzhou Medical University, and Professor Kang Zhang, Director of the Institute of Human Genomic Medicine at the University of California, San Diego (UCSD), published an in-depth review in the latest issue of Nature Medicine, outlining and predicting the current status and future development of AI technology implementation in healthcare. We have compiled the highlights from this review for our readers.

The current state of AI in healthcare

"AI+medicine" refers to the use of artificial intelligence to obtain information from data through various technologies such as machine learning, characterization learning, deep learning and natural language processing, using computer algorithms to assist in clinical decision making, achieving a series of functions such as diagnosis, therapy selection, risk prediction, disease triage, medical incident reduction and efficiency improvement.

In healthcare, the applications where AI will have a significant impact will cover four major directions: diagnosis, treatment, population health management, surveillance and regulation.

Researchers predict several ways in which AI-based technologies could be applied in clinical implementation.

The first is as a triage and screening tool that could theoretically reduce the strain on the healthcare system and allocate resources to patients most in need of medical help. For example, through deep learning, AI tools can examine retinal images to determine which patients have blinding eye disease and refer them to an ophthalmologist in a timely manner. There is also a mobile app developed by Babylon, a British company, that allows chatbots to interact directly with users, essentially an AI-based triage tool for distinguishing whether a patient needs to see a doctor for further examination.

AI technology can also be used as a replacement for manpower on tasks that are theoretically uncomplicated but time-critical and labor-intensive, allowing healthcare workers to tackle more complex tasks. Examples include automated analysis of radiographic images to estimate bone age; automated analysis of optical coherence tomography (OCT) images to diagnose treatable retinal diseases; automated analysis of cardiovascular images to quantify blood vessel stenosis and other indicators, and so on.

Perhaps the best way to demonstrate the value of AI is to allow AI to assist professional physicians. Allowing clinicians to combine with AI creates a 1+1>2 synergy that supports real-time clinical decision making and fuels precision medicine.

Key Issues for Implementing AI Technologies in Clinical Practice

Although medical-related AI technologies continue to achieve breakthroughs, there is still a certain distance between "translating" the technologies into real clinical applications. True "industrialization" requires access to large volumes of data, the embedding of AI into actual clinical workflows, and a regulatory framework. Researchers believe that the following major issues need to be addressed.

Data sharing

Data is central to both the initial training of AI and the validation and improvement of algorithms. Currently, the likes of the Cardiac Atlas Project, the Visual Concept Extraction Challenge in Radiology (VISCERAL), the UK Biobank "and the Kaggle Data Science Bowl, provide large-scale datasets of imaging and non-imaging data. However, researchers believe that the extent of data sharing needs to be further increased for broader adoption of AI technologies in healthcare.

Accuracy and transparency of data and algorithms

Transparency involves multiple dimensions. In supervised learning, for example, prediction accuracy relies heavily on the accuracy of the annotations fed into the algorithm. A large amount (tens to hundreds of thousands of levels) of high-quality well-labeled data is a fundamental condition for algorithm accuracy and a scarce resource. In addition the transparency of the labeling of the input data plays a key role in assessing the accuracy of the training process of supervised learning algorithms.

Transparency also affects the interpretability of the model, that is, the logic that allows humans to understand or interpret the logic that results from a particular prediction or decision. AI technology applied to healthcare needs to be open to the "black box" and transparent enough to judge the reasonableness of a diagnosis, treatment recommendation, or prediction.

Another important reason for transparency is that AI technologies may have algorithmic biases that can magnify discrimination based on race, gender, or other characteristics. Transparency in training data and interpretability of models allows us to check for potential bias. Ideally, algorithms can be used to address algorithmic biases, and even genetic and biological differences in health between groups can be addressed through machine learning if they are designed to make up for known biases.

Patient safety

air purifier for pet odor

Accountability is an important issue related to patient safety. When AI technologies cause harm to our bodies, who should be held accountable for it? Undoubtedly, AI technology will change the traditional doctor-patient relationship. Efforts are being made by multiple governments and WHO regulatory bodies to try to strike a delicate balance between protecting patient safety and promoting technological innovation.

Data standardization

Given the complexity and large scale of healthcare data, for AI technologies to effectively use data collected in a variety of ways, data standardization should be done during the initial development phase to translate data into a common format that can be understood across different tools and methods.

A typical clinical workflow consists of multiple components that place demands on interoperability. In AI-assisted radiology, for example, the algorithms used for exam operations, study prioritization, feature analysis and extraction, and automated report generation may be products from different vendors, and a set of workflow interoperability standards needs to be created for integration between algorithms and to allow the algorithms to run on different devices. Without early optimization of interoperability, the effectiveness of practical applications of AI technology will be severely constrained.

Embedding into existing clinical workflows

The Digital Imaging and Communications in Medicine (DICOM) standard and the Medical Image Archiving and Communication System (PACS) provide a consistent platform for data management that has revolutionized medical imaging dramatically, and similar standards should be applied to AI technology to develop uniform naming for easy data storage and retrieval.

For example, the Fast Health Interoperable Resources (FHIR) framework, which is designed to enable clinical translation, is a rapidly evolving set of standards worldwide, built on a series of modular components called "resources". These resources can be easily assembled into working systems to facilitate data sharing between electronic medical records, mobile applications, cloud communications, etc., which are critical to the future implementation of AI technologies in healthcare.

Economic Considerations and Talent Staffing Issues

In particular, the researchers suggest that given the complexity of clinical decision-making and the potential consequences of misuse, the implementation of AI technologies in medicine requires the active participation of all stakeholders, creating communication and collaboration among physicians, healthcare providers, data scientists, computer scientists, and engineers.

Policy and Regulatory Environment for Assessing Safety and Efficacy

Building on the U.S. FDA's Digital Health Innovation Action Plan (DHIAP), which launched in July 2017 with new regulatory initiatives for medical software, a number of AI technologies have already been approved by the FDA. For example, the first FDA-approved medical device to use AI, the "autonomous" diagnostic system IDx-DR, uses AI algorithms to automatically detect the presence of mild diabetic retinopathy (DR) for patients and, based on the results of the screening, provide advice on whether to Based on the results of the screening, it provides recommendations for referral to an ophthalmologist for use in primary care. The AI product was launched through the FDA's "De Novo reclassification" pathway for low to moderate risk, and qualified as a Breakthrough Device.

In addition, the FDA launched a software pre-certification program that focuses on reviewing software technology developers rather than individual products, improving access to technology and focusing resources on high-risk products.

The EU officially implemented the General Data Protection and Regulation (GDPR) from May 2018, which states that citizens have the right to an explanation of algorithmic decisions. This means that informed consent must be obtained for any personal data collection when implementing AI technologies in healthcare; after collection, patients who provide the data should have the right to see what the data was collected for and to delete it. Researchers expect the introduction of the GDPR to promote public trust and patient engagement, thereby facilitating the implementation of AI technologies in the long run.

China is also a major player in the global AI arena, and AI technology is one of the key opportunities to achieve equity in healthcare resources. Encourage the vigorous development of AI and other technology applications in the field of healthcare.

In actual clinical practice, AI technologies have been implemented in diagnostic tools for diseases such as lung cancer, esophageal cancer, diabetic retinopathy, and diagnostic aids for pathological examinations. The assisted diagnosis and screening system introduced at the First People's Hospital of Kashgar and its health outlets in Xinjiang is a successful case based on AI technology, using retinal photographs to screen and diagnose blinding eye diseases such as diabetic retinopathy, glaucoma, and age-related macular degeneration, with preliminary results demonstrating the high accuracy of AI diagnosis.

Future developments

Researchers expect that radiology, pathology, ophthalmology and dermatology will be the first clinical areas to translate AI technology, and that these primarily image-based fields are suitable for training AI technology for automated analysis or diagnostic prediction. In contrast, AI technologies may take longer to integrate into practical applications in areas that require the integration of multiple types of data (e.g., internal medicine) or where surgical procedures are a necessary component (e.g., surgical specialties). Overall, however, research on AI-related applications is growing by leaps and bounds throughout the medical field.

Researchers also caution that while AI technologies promise to improve productivity, they are not as absolutely reliable as the humans who created them, and it is necessary for researchers, developers, and decision makers alike to evaluate and implement AI technologies with a critical eye, keeping in mind their limitations.

AI Artificial Intelligence Health Field