Artificial Intelligence Resources

Tools & Resources

Artificial intelligence (AI) has become a widespread and prominent presence in nearly every field of human endeavor with a large amount of attention on health care. At the AAN, we recognize that AI has the potential to transform the practice of neurology from the bench to the bedside with appropriate awareness and cautions. We also recognize that AI is a broad term, encompassing many technologies, both new and established, some of which are already in use; and many of which are mostly potential without the reliability, trust, and accuracy required for patient-centered care.

We encourage members to learn more about AI. This page will be updated with resources, references, and commentary about AI relevant to neurology practice and which uphold our principles as an organization.

Artificial Intelligence Definition

A broad term that refers to automation of complex tasks (Luger 2004) with a range of technologies that range from deterministic (known input leads to known output) to probabilistic (outputs are not predictable from inputs).

Artificial Intelligence in Neurology Community

Get connected with AAN members to stay up to date on the latest AI discussions.

Join Today

Introduction to AI

AI In Neurology

AAN Involvement

Learn More about AI

Neurology Publications

Find the latest artificial intelligence research, publications, and news.

Neurology

The official journal of the American Academy of Neurology

Neurology Today

The official news source of the AAN, reporting breaking news, issues, and trends in the practice and science of neurology.

AI in the News

Artificial Intelligence FAQs

AI Terms

Machine learning (ML) 

A broad term for the processes by which computers “learn” from examples to draw conclusions or make decisions without being explicitly programmed to do so. Basic ML models take a given input and usually result in the same output. Specific ML uses are increasingly embedded in medical devices and software.

Deep learning (neural networks) 

A form of machine learning using a model with layers of interconnected networks. Each layer could be considered a “specialist” in one part of the problem. The neural network is how the layers are connected to each other, often in a hierarchy (subspecialists to specialists) to produce a final output. Deep learning is widely used for processing large, complex data. This type of AI is particularly established in natural language processing and image analysis tools.

Natural language processing (NLP)

A specific application of deep learning which enables computers to process and analyze text and spoken words. As a form of deep learning, NLP models require training on the nuances of language (e.g., accents, styles, language, domain expertise, context) to make an application effective and accurate in its task. Accuracy has been gradually increasing over decades of development and often commercially available in sofware.

Generative AI

A complex algorithm that can generate novel content (e.g., text or images) in response to prompts. A large language model (LLM) is an extremely large application of deep learning in natural language tasks and is a widely used example of generative AI. LLMs are trained on large amounts of text data from online articles, books, and other content, and then can generate human-like text by predicting the next set of words based on plain language prompts. Follow-up prompts and responses can occur in a conversation like fashion.

Clinical algorithms 

A series of rules that process a set of input data and outputs a result. Predictive models use clinical algorithms as the engine to predict a clinical outcome, often a classification model (e.g., what diagnosis is most likely) or a number (e.g., percent likelihood of a no-show or length-of-stay). The range of actual technology used in clinical algorithms ranges from simple if-then statements or flowcharts (which may be considered AI when embedded in computers to assist in automating complex tasks) to complex deep learning models and large language models (which is currently often what is meant by AI).

Considerations for Generative AI Platforms/Large Language Models (LLM)

  1. Define the problem that generative AI/LLM will solve. Define how generative AI and large language models (LLM) can fit into the practice strategy. Recognize safe integration strategies with electronic health records (EHR), call management, chart review and documentation, assessment and care planning, and education. Here is additional information regarding the principles of AI models in medicine and neurology clinical care.
  2. Understand the limitations. Problems with generative AI/LLM can include hallucinations (fabricated erroneous facts) from datasets, factual mistakes, faulty logic, bias (ex. racial/ethnic biases), and memorization. Fully understanding the limitations can ensure safe execution of these models.
  3. Consider the barriers. Before implementation, consider the barriers to successful implementation within a clinical setting. These can include patient safety issues, data quality, data variability, data security, discrimination and bias, and ongoing model training. Learn more about the potential barriers to implementation.
  4. Review compliance. Before implementation of generative AI/LLM models within clinical practice, it is vital to review compliance with regulations such as HIPAA and GDPR. Collaboration with departmental/institutional leadership is foundational for success.
  5. Consider silent pilots. Prospectively measure accuracy and drift to ensure that generative AI models work sufficiently well and satisfy the institutional regulatory requirements. Consider re-training generative AI models to meet your institutional and regulatory needs based on collected data. Silent pilots have low risk and may help build experience and trust in an AI tool before implementation.