Prospectus for 'The Practical Implementation of Artificial Intelligence Technologies in Medicine': A Nature Medicine Comprehensive Review

Editor’s Note: This article is reprinted from Academic Horizon, with authorization granted to VCBeat.

Artificial intelligence (AI) is experiencing explosive growth, impacting numerous industries and ushering in a new revolution in healthcare. “AI + Healthcare” has emerged as a hotly debated field, drawing significant attention from academia, industry, and regulatory agencies alike.

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

▲Professor He Jianxing (left) and Professor Zhang Kang (right) (Image source: the research institutes to which the two scholars belong)

“AI + Healthcare” refers to the application of artificial intelligence—leveraging technologies such as machine learning, representation learning, deep learning, and natural language processing—to extract information from data via computer algorithms. With the aim of assisting clinical decision-making, it enables a range of functions, including auxiliary diagnosis, treatment selection, risk prediction, disease triage, reduction of medical errors, and improvement of operational efficiency.

In the healthcare sector, AI applications with significant impact will span four key areas: diagnosis, treatment, population health management, and monitoring and regulation.

▲ Four Major Directions for the Potential Applications of “AI + Healthcare” (Image source: Adapted from a figure in Nature Medicine)

Researchers have predicted several ways in which AI-based technologies can be implemented in clinical practice.

First, as a triage and screening tool, it can theoretically alleviate pressure on the healthcare system by allocating resources to patients in greatest need of medical assistance. For example, through deep learning, AI tools can examine retinal images to identify patients with blinding eye diseases and ensure timely referral to ophthalmologists. Additionally, Babylon Health, a UK-based company, has developed a mobile application featuring a chatbot that interacts directly with users; this is essentially an AI-powered triage tool designed to determine whether patients require further evaluation by a physician.

AI technology can also serve as a substitute for manual labor in tasks that are theoretically uncomplicated but time-sensitive and labor-intensive, thereby allowing healthcare professionals to focus on more complex responsibilities. Examples include the automated analysis of radiological images for bone age assessment; the automated interpretation of optical coherence tomography (OCT) scans to diagnose treatable retinal diseases; and the automated analysis of cardiovascular images to quantify vascular stenosis and other metrics, among others.

Perhaps the most effective way to demonstrate the value of AI is by having it assist professional physicians. By integrating clinicians with AI to generate a synergistic effect where 1+1>2, we can support real-time clinical decision-making and advance precision medicine.

Although AI technologies in healthcare continue to achieve breakthroughs, there remains a considerable gap between these technological advances and their practical implementation in clinical settings. To truly achieve industrialization, it is necessary to acquire large-scale datasets, integrate AI into actual clinical workflows, and align with regulatory frameworks. Researchers believe that the following key issues need to be addressed.

Data serves as the foundational cornerstone for both the initial training of AI models and the validation and refinement of algorithms. Currently, international initiatives such as 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 comprising both imaging and non-imaging data. Nevertheless, researchers argue that a greater degree of data sharing is essential to facilitate broader adoption of AI technologies in healthcare.

Transparency spans multiple levels. For instance, in supervised learning, predictive accuracy largely depends on the accuracy of the annotations fed into the algorithm. A large volume (ranging from tens of thousands to hundreds of thousands) of high-quality, well-annotated data is a fundamental prerequisite for algorithmic accuracy and constitutes a scarce resource. Furthermore, the transparency of input data labels plays a critical role in assessing whether the training process of supervised learning algorithms is accurate.

Transparency also affects the interpretability of models, enabling humans to understand or explain the logic behind specific predictions or decisions. AI technologies applied in healthcare need to open the "black box" and provide sufficient transparency to assess the rationality of diagnoses, treatment recommendations, or predictive outcomes.

Another critical reason for transparency is that AI technologies may harbor algorithmic biases, which can exacerbate discrimination based on race, gender, or other characteristics. Transparency in training data and model interpretability enable the identification of potential biases. Ideally, algorithmic approaches can be employed to address these biases; if systems are designed to compensate for known biases, machine learning may even help address genetic and biological health disparities among different populations.

Accountability is a critical issue concerning patient safety. When AI technologies cause harm to patients, who should be held liable? Undoubtedly, AI will transform the traditional doctor-patient relationship. Governments of multiple countries and regulatory bodies of the World Health Organization (WHO) are making efforts to strike a delicate balance between safeguarding patient safety and fostering technological innovation.

Given the complexity and large scale of healthcare data, to effectively leverage data collected through various means, AI technologies should prioritize data standardization during the initial development phase, converting data into a universal format interpretable across different tools and methodologies.

A typical clinical workflow comprises multiple components, imposing significant demands on interoperability. Taking AI-assisted radiology as an example, algorithms for examination procedures, study prioritization, feature analysis and extraction, and automated report generation may be products provided by different vendors. A set of workflow interoperability standards must be established to integrate these algorithms and enable their deployment across diverse devices. Failure to optimize interoperability at an early stage will severely constrain the practical effectiveness of AI technologies.

The Digital Imaging and Communications in Medicine (DICOM) standard and Picture Archiving and Communication Systems (PACS) have revolutionized medical imaging by providing a consistent platform for data management. Similar standards should be applied to AI technology, with unified nomenclature developed to facilitate data storage and retrieval.

For example, the Fast Healthcare Interoperability Resources (FHIR) framework, designed to facilitate clinical translation, is a rapidly evolving global standard built on a series of modular components known as “resources.” These resources can be easily integrated into operational systems, enabling seamless data sharing across electronic health records, mobile applications, and cloud-based communication platforms, which is critical for the future implementation of AI technologies in healthcare.

Researchers specifically emphasize that, given the complexity of clinical decision-making and the potential consequences of misuse, the implementation of AI technologies in the medical field requires the active participation of all stakeholders, fostering communication and collaboration among physicians, healthcare providers, data scientists, computer scientists, and engineers.

In July 2017, the U.S. Food and Drug Administration (FDA) launched the Digital Health Innovation Action Plan, introducing new regulatory measures for medical software. On this basis, several AI technologies have already received FDA approval. For example, IDx-DR, the first FDA-approved medical device utilizing AI, is an “autonomous” diagnostic system that uses AI algorithms to automatically detect mild diabetic retinopathy (DR) in patients. It provides recommendations on whether referral to an ophthalmologist is necessary based on screening results, making it suitable for primary care settings. The market authorization process for this AI product followed the FDA’s “De Novo classification” pathway for low-to-moderate-risk devices and was granted Breakthrough Device designation.

Furthermore, the FDA launched the Software Pre-Certification Program, which focuses on reviewing software technology developers rather than individual products, thereby improving access to technology and concentrating resources on high-risk products.

▲Conceptual diagram of the FDA’s pre-certification program for Software as a Medical Device (Image source: Adapted from Nature Medicine)

The European Union officially implemented the General Data Protection Regulation (GDPR) in May 2018, which stipulates that citizens have the right to obtain explanations for algorithmic decisions. This means that when implementing AI technologies in the healthcare sector, informed consent must be obtained for any collection of personal data; furthermore, patients who provide their data should have the right to view how their data is used and to request its deletion. Researchers anticipate that the enforcement of GDPR will foster public trust and patient engagement, thereby facilitating the long-term implementation of AI technologies.

China is also a major player on the global AI stage. AI technology has become one of the key opportunities to achieve equitable access to healthcare resources. In July 2017, the State Council issued the Development Plan for New-Generation Artificial Intelligence, and in April 2018, the General Office of the State Council issued the Guiding Opinions on Promoting the Development of “Internet Plus Healthcare,” explicitly encouraging the vigorous development and application of technologies such as artificial intelligence in the healthcare sector.

In actual clinical practice, there are already implemented cases of AI technology in diagnostic tools for diseases such as lung cancer, esophageal cancer, and diabetic retinopathy, as well as in diagnostic assistance for pathological examinations. A successful case based on AI technology is the assisted diagnosis and screening system introduced at the First People's Hospital of Kashgar, Xinjiang, and its health network points. This system uses retinal photographs to screen and diagnose blinding eye diseases such as diabetic retinopathy, glaucoma, and age-related macular degeneration. Preliminary results have demonstrated the high accuracy of AI-based diagnoses.

Researchers anticipate that clinical specialties such as radiology, pathology, ophthalmology, and dermatology will be the earliest to achieve AI technology translation. These imaging-centric fields are well-suited for training AI systems to perform automated analysis or diagnostic predictions. In contrast, in specialties that require the integration of diverse data types (such as internal medicine) or where surgical procedures are essential (such as surgical specialties), the integration of AI into practical applications may take longer. Nevertheless, across the entire medical field, AI-related applied research is advancing rapidly.

Researchers also caution that while AI technologies hold promise for enhancing productivity, they are not infallible, much like the humans who create them. It is therefore essential for researchers, developers, and policymakers to evaluate and implement AI technologies with a critical eye, keeping their limitations in mind.

References:

[1] Jianxin He, et al., (2019), The practical implementation of artificial intelligence technologies in medicine, Nature Medicine, DOI: https://www.nature.com/articles/s41591-018-0307-0