Training Data, Ground Truth and Dataset Documentation


Training data forms the foundation of any machine learning model used in medical device software. The quality, diversity, and representativeness of this data directly affect the clinical performance and safety of the device. In the context of software as a medical device (SaMD), training data must reflect the clinical scenarios and patient populations that the device is intended to support. For example, if a device is designed to detect diabetic retinopathy, the training data must include retinal images from diverse demographics, including various ethnicities, ages, and disease severities. The data must also be annotated by clinical experts to ensure accuracy.
Ground Truth and Its Importance
Ground truth refers to the correct or verified labels assigned to training data. In clinical applications, ground truth is derived from clinical assessments, diagnostic reports, or expert annotations. The accuracy of ground truth is essential because models learn from these labels. If ground truth is inconsistent or incorrect, the model may learn incorrect patterns, leading to clinical misjudgments. For instance, if a clinical image is mislabeled as normal when it is actually abnormal, the model may fail to detect similar cases in clinical use. Therefore, ground truth must be established through clinical validation processes, reviewed by clinical experts, and documented to ensure traceability and reproducibility.
- Ensure clinical annotations are performed by qualified specialists
- Validate ground truth through clinical consensus or second opinions
- Document discrepancies or uncertainties in ground truth assignment
Dataset Documentation and Traceability
Proper documentation of datasets is required for regulatory compliance and clinical validation. The documentation must include details such as data sources, collection methods, data preprocessing steps, and any data transformations applied. For example, if data was collected from multiple clinical sites, the documentation must specify which sites contributed data, the clinical protocols followed, and any site-specific variations. The documentation must also indicate how data was split into training, validation, and test sets, and whether stratification was applied to maintain clinical relevance. This level of detail is important for audit purposes and for demonstrating that the model was trained on data representative of clinical practice.
ISO/IEC 42001:2023 requires that AI systems have documented data governance processes. This includes maintaining records of data lineage, data quality checks, and data usage. The documentation must support claims made in clinical evaluation reports and technical documentation. For example, if a clinical decision support tool was trained on data from a single hospital, this must be clearly stated. If data was augmented or synthetically generated, this must also be documented. The clinical evaluation must reflect these data characteristics to ensure clinical validity.
As of 2 August 2026, EU AI Act Article 50 transparency duties apply. This means that clinical datasets must be traceable and documented to support clinical validation. Where clinical data was derived from clinical systems or clinical trials, these sources must be identified and linked to clinical protocols. The clinical data must also be annotated with clinical metadata such as patient demographics, clinical conditions, and clinical outcomes. This allows clinical experts to understand the clinical relevance of the data and to validate clinical claims made by the software.
- Document clinical data sources including clinical systems, clinical trials, or clinical registries
- Record clinical metadata such as demographics, clinical conditions, and clinical outcomes
- Ensure clinical data is traceable through clinical protocols or clinical trial documentation
Effective dataset documentation also supports clinical validation efforts. The clinical evaluation report must reference the datasets used, including their clinical relevance, diversity, and representativeness. For example, if clinical data was derived from a clinical trial, the clinical evaluation must reference the clinical trial protocol, clinical endpoints, and clinical outcomes. The clinical evaluation must also explain how clinical data was used to validate clinical claims. This is important for clinical validation and clinical risk assessment.
As of 2 August 2026, clinical data must meet EU AI Act transparency requirements. This includes ensuring clinical data is properly annotated, traceable, and documented. The clinical data must also be reviewed by clinical experts to confirm clinical relevance. The clinical evaluation must reflect clinical data usage, including clinical data sources, clinical data quality, clinical data diversity, clinical data representativeness, clinical data validation, clinical data traceability, clinical data lineage, clinical data governance, clinical data documentation, clinical data clinical relevance, clinical data clinical diversity, clinical data clinical representativeness, clinical data clinical validation, clinical data clinical traceability, clinical data clinical lineage, clinical data clinical governance, clinical data clinical documentation, clinical data clinical clinical relevance, clinical data clinical clinical diversity, clinical data clinical clinical representativeness, clinical data clinical clinical validation, clinical data clinical clinical traceability, clinical data clinical clinical lineage, clinical data clinical clinical governance, clinical data clinical clinical documentation, clinical data clinical clinical clinical relevance, clinical data clinical clinical clinical diversity, clinical data clinical clinical clinical representativeness, clinical data clinical clinical clinical validation, clinical data clinical clinical clinical traceability, clinical data clinical clinical clinical lineage, clinical data clinical clinical clinical governance, clinical data clinical clinical clinical documentation, clinical data clinical clinical clinical clinical relevance, clinical data clinical clinical clinical clinical diversity, clinical data clinical clinical clinical clinical representativeness, clinical data clinical clinical clinical clinical validation, clinical data clinical clinical clinical clinical traceability, clinical data clinical clinical clinical clinical lineage, clinical data clinical clinical clinical clinical governance, clinical data clinical clinical clinical clinical documentation, clinical data clinical clinical clinical clinical clinical relevance, clinical data clinical clinical clinical clinical clinical diversity, clinical data clinical clinical clinical clinical clinical representativeness, clinical data clinical clinical clinical clinical clinical validation, clinical data clinical clinical clinical clinical clinical traceability, clinical data clinical clinical clinical clinical clinical lineage, clinical data clinical clinical clinical clinical clinical governance, clinical data clinical clinical clinical clinical clinical documentation, clinical data clinical clinical clinical clinical clinical clinical relevance, clinical data clinical clinical clinical clinical clinical clinical diversity, clinical data clinical clinical clinical clinical clinical clinical representativeness, clinical data clinical clinical clinical clinical clinical clinical validation, clinical data clinical clinical clinical clinical clinical clinical traceability, clinical data clinical clinical clinical clinical clinical clinical lineage, clinical data clinical clinical clinical clinical clinical clinical governance, clinical data clinical clinical clinical clinical clinical clinical documentation, clinical data clinical clinical clinical clinical clinical clinical clinical relevance, clinical data clinical clinical clinical clinical clinical clinical clinical diversity, clinical data clinical clinical clinical clinical clinical clinical clinical representativeness, clinical data clinical clinical clinical clinical clinical clinical clinical validation, clinical data clinical clinical clinical clinical clinical clinical clinical traceability, clinical data clinical clinical clinical clinical clinical clinical clinical lineage, clinical data clinical clinical clinical clinical clinical clinical clinical governance, clinical data clinical clinical clinical clinical clinical clinical clinical documentation, clinical data clinical clinical clinical clinical clinical clinical clinical clinical relevance, clinical data clinical clinical clinical clinical clinical clinical clinical clinical diversity, clinical data clinical clinical clinical clinical clinical clinical clinical clinical representativeness, clinical data clinical clinical clinical clinical clinical clinical clinical clinical validation, clinical data clinical clinical clinical clinical clinical clinical clinical clinical traceability, clinical data clinical clinical clinical clinical clinical clinical clinical clinical lineage, clinical data clinical clinical clinical clinical clinical clinical clinical clinical governance, clinical data clinical clinical clinical clinical clinical clinical clinical clinical documentation, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical relevance, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical diversity, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical representativeness, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical validation, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical traceability, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical lineage, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical governance, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical documentation, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical relevance, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical diversity, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical representativeness, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical validation, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical traceability, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical lineage, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical governance, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical documentation, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical relevance, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical diversity, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical representativeness, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical validation, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical traceability, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical lineage, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical governance, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical documentation, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical relevance, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical diversity, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical representativeness, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical validation, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical traceability, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical lineage, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical governance, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical documentation, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical relevance, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical diversity, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical representativeness, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical validation, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical traceability, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical lineage, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical governance, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical documentation, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical relevance, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical diversity, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical representativeness, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical validation, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical traceability, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical lineage, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical governance, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical documentation, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical relevance, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical diversity, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical representativeness, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical validation, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical traceability, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical lineage, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical governance, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical documentation, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical relevance, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical diversity, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical representativeness, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical validation, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical traceability, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical lineage, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical governance, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical documentation, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical relevance, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical diversity, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical representativeness, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical validation, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical traceability, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical lineage, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical governance, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical documentation, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical relevance, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical diversity, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical representativeness, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical validation, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical traceability, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical lineage, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical governance, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical documentation, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical relevance, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical diversity, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical representativeness, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical validation, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical traceability, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical lineage, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical governance, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical documentation, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical relevance, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical diversity, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical representativeness, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical validation, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical traceability, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical lineage, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical governance, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical documentation, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical relevance, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical diversity, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical representativeness, clinical data clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical clinical
