AI in healthcare: clinical uses, evidence requirements and governance
Artificial intelligence is used in healthcare for tasks such as image analysis, clinical decision support, and drug discovery. Regulators like the FDA require evidence of safety and effectiveness for AI-enabled medical devices. The WHO provides governance and ethics guidance to ensure responsible use.
Clinical applications in healthcare
Artificial intelligence in healthcare is designed to support clinicians and improve care delivery. According to IBM, AI is used to automate routine work, make better use of data, and support clinical decision-making rather than replace healthcare professionals. IBM's overview of AI in healthcare
In oncology, AI helps with screening, detection, and treatment planning. The National Cancer Institute reports that AI-based software has been authorized to help pathologists find cancer in prostate biopsy images, and imaging algorithms can improve breast cancer detection and predict long-term risk. AI also accelerates drug discovery and precision treatment by predicting T-cell responses and enabling genetic subtyping of brain tumors during surgery. National Cancer Institute's AI and cancer page
Types of AI and dataset requirements
AI encompasses several approaches, each suited to different tasks. The National Institute of Biomedical Imaging and Bioengineering explains that machine learning identifies hidden patterns in large datasets, deep learning uses layered neural networks to learn from complex data, and generative AI creates new content such as text or images. NIBIB's artificial intelligence science topic
For medical imaging, datasets must be curated, de-identified, assessed for quality, and annotated by clinical experts to teach AI tools effectively. The same source notes that researchers must assure that AI tools are fair and perform equally across groups and populations.
Data repositories like the Medical Imaging and Data Resource Center provide large collections of X-rays, CT scans, and MRIs for algorithm development. Such high-quality data is essential for building AI that can assist with earlier diagnoses and more effective treatment plans.
Evidence and regulatory oversight
In the United States, AI-enabled medical devices must meet the Food and Drug Administration's premarket requirements, including a focused review of safety and effectiveness. The FDA maintains a list of authorized AI-enabled medical devices, which provides transparency for providers and patients. FDA's list of AI-enabled medical devices
Devices are included in the list when AI-related terms appear in their marketing authorization summaries; the list is not a comprehensive registry of every such device. The FDA also plans to identify devices that incorporate foundation models, including large language models, in future updates.
The authorization process involves evaluation of study appropriateness for the device's intended use and technological characteristics. This helps ensure that AI devices are both innovative and safe for clinical use.
Governance and ethics
The World Health Organization's guidance on ethics and governance of artificial intelligence for health identifies six consensus principles to ensure that AI works for the public benefit of all countries. The guidance also includes recommendations that hold stakeholders in the public and private sectors accountable and responsive to healthcare workers and affected communities. WHO's ethics and governance guidance
According to the same document, AI technologies must put ethics and human rights at the heart of their design, deployment, and use. This includes addressing challenges such as privacy, security, transparency, and regulatory compliance.
Healthcare organizations adopting AI should establish oversight structures to ensure responsible use, with attention to equitable performance across diverse patient groups.
Guidance for large multi-modal models
The WHO has also issued specific guidance on large multi-modal models (LMMs), a type of generative AI that accepts multiple data inputs and generates diverse outputs. This guidance addresses the predicted wide use of LMMs in healthcare, research, and drug development, while noting that it is not yet proven whether they can accomplish a wide range of tasks. WHO's guidance on large multi-modal models
LMMs are also known as general-purpose foundation models, but their reliability for specific medical applications remains under study. The WHO guidance emphasizes that these models should be evaluated carefully and that governance should evolve alongside technological capabilities.
The document forms part of the WHO's broader ethical framework, which covers not only LMMs but also other AI systems used in health. It advises on appropriate oversight, transparency, and accountability to protect patients and public health.
| AI type | Function | Healthcare example |
|---|---|---|
| Machine learning | Finds patterns in data and makes predictions | Clinical decision support systems |
| Deep learning | Uses layered neural networks on complex data | Analysis of X-rays and other medical images |
| Generative AI | Creates new text, images, or other content | Synthetic data generation or draft reports |
| Large language model | Understands and generates human-like text | Knowledge extraction from research publications |
Frequently asked questions
What are the main clinical uses of AI?
AI supports image interpretation, clinical decision support, cancer screening, drug discovery, and wearable monitoring.
How does the FDA regulate AI-enabled medical devices?
The FDA requires premarket review of safety and effectiveness for AI-enabled medical devices and publishes a list of authorized devices.
What ethical principles does WHO outline for AI in health?
The WHO guidance identifies six consensus principles, including autonomy, safety, transparency, accountability, equity, and sustainability.
Why is high-quality data important for AI in medicine?
AI algorithms require curated, de-identified, and labeled datasets created by clinical experts to perform reliably and fairly across populations.