Components and Drivers
|
Description and purpose
|
Functionalities
|
Target Groups
|
Data Acquisition Package (DAP)
Joachim Krois
| High-quality data are required to train and evaluate AI solutions. DAP coordinates the compilation of such data and ensures the availability of metadata and (where relevant) patient consent information.
| Facilitates data compilation for many modalities; registers data and metadata; ingests data from DP; and manages patient consent information.
| Manufacturers and medical experts
|
Data Storage Package (DP)
Andrea Romaoli Garcia
| The basis for all subsequent packages is an orderly processing and storage of data and metadata. Medical data storage requires adherence to strict guidelines that preserve the privacy and safety of patients. DP provides data storage guidelines that consider these constraints.
| Provides safe and secure storage of medical data; serves as an interface for other packages (AP, PP, and EP) to access this data; and offers data governance that complies with data protection laws.
| Manufacturers, medical experts, and notified regulatory bodies
|
Annotation Package (AP)
Marc Lecoultre
| High-quality annotated data provide the basis for supervised learning. Unfortunately, production is challenging and labor intensive. Certain features must be considered when evaluating an annotation: the quality of labels, the number of expert opinions, and the handling of non-unanimous decisions. AP brings together leading health experts across the globe to produce the highest-quality annotations at maximum efficiency.
| Provides an annotation interface for many modalities; includes collaboration features; develops a network of annotation experts; and creates notifications of pending annotation tasks
| Manufacturers and notified regulatory bodies
|
Prediction Package (PP)
Luis Oala,
Alixandro Werneck
| For many health cases (e.g., the detection of breast cancer tissue or the classification of skin irritations), multiple parameters and AI models can provide viable solutions. PP allows various models to be evaluated in parallel for a benchmarking result.
| Loads AI models; manages models undergoing prediction tasks (considers versioning, meta-data, etc.); and orchestrates computations
| Manufacturers and notified regulatory bodies
|
Evaluation Package (EP)
Luis Oala,
Alixandro Werneck
| Model performance is dependent on the choice of metric and possible parameters. Thus, it is of utmost importance to have a framework that allows for the comparison of the performance of different AI models. EP provides meaningful, state-of-the-art metrics that promise high expressibility.
| Offers testing measures and methods for different quality dimensions including interpretation, bias, uncertainty, and robustness; questionnaires provide qualitative evaluation
| Manufacturers and notified regulatory bodies
|
Reporting Package (RP)
Pradeep Balachandran
| RP delivers a standardized format for communicating and reporting the properties, features, intended use, and limitations of AI for health to help connect different stakeholders.
| A customizable reporting interface that presents the results of EP
| Manufacturers, notified regulatory bodies, users of AI for health (doctors, patients), and vendors of AI for health
|