Table 1.
Proposed registration information for early stage clinical AI algorithms in healthcare
| Itema | Description | Generative AIb | Non-Generative AIb |
|---|---|---|---|
| Name, Version, and AI Model Type | Name of the system, its version, and the type of AI model used (e.g., deep learning, decision tree, etc.) | ✓ | ✓ |
| Training and validation population | Demographics (e.g., age, gender, ethnicity) of the patient population on which the algorithm was trained | ✓ | ✓ |
| Clinical context | Model application (e.g. used for administrative purpose or for instance to predict a specific illness) | ✓ | ✓ |
| Performance Metrics | Performance of the AI system in preclinical development/validation and prior clinical studies (e.g., model discrimination and calibration) | ✓ | ✓ |
| Input Data Types | Types of data used as inputs by the AI system (e.g., images, clinical notes, lab results, etc.) | ✓ | ✓ |
| Data Acquisition and Processing | Process of data acquisition, the steps required for input data entry, the pre-processing procedures applied, and the methodologies employed for handling missing or low-quality data | ✓ | ✓ |
| Output Types and Presentation | Types of outputs generated by the AI system (e.g., predictions, recommendations, etc.) and how these outputs are presented to the users | ✓ | ✓ |
| Registrant Information | Name, affiliation, and contact information of the person or organization that registered the AI system | ✓ | ✓ |
| Foundation model-specific information | Type and version of the foundation model used (e.g., LLM, version: GPT-4 or PaLM 2) | ✓ | |
| Manufacturer or company that developed the foundation model (e.g., OpenAI, Microsoft, Google etc.) | ✓ | ||
| Fine-tuning or grounding process used on the foundation model | ✓ |
AI Artificial Intelligence, LLM Large Language Model, GPT-4 Generative Pre-trained Transformer 4, PaLM 2 Pathways Language Model.
aItems to be registered have been adapted from DECIDE-AI and CONSORT-AI guidelines.
bThe terminology “Generative AI” and “Non-Generative AI” specifies if certain data or criteria are relevant to generative AI models, non-generative AI models, or both. The presence of a checkmark (✓) in a column signals that the mentioned data or criteria pertain to that AI category. Generative AI models are capable of producing new content, such as the way Large Language Models (LLMs) can craft text that mimics human writing. On the other hand, Non-Generative AI models are designed to interpret and learn from pre-existing data to make forecasts or decisions. For instance, these models analyze and learn from existing data to make predictions or decisions.