Good Machine Learning Practice (GMLP)

Leverage a deep, cross-functional understanding of clinical workflows, model engineering, and regulatory requirements throughout the design process. Operational silos are a primary source of safety risk.

• Integrate clinicians, data scientists, and regulatory specialists from day one.
• Ensure the team understands the intended use environment and potential misuse scenarios.
• Document the roles and responsibilities of each discipline in the development plan.

Implement rigorous software lifecycle controls, risk management, and cybersecurity practices. AI/ML components are software first and must adhere to foundational IEC 62304 principles.

• Apply defense-in-depth security strategies to protect the model pipeline.
• Conduct adversarial testing to assess vulnerability to data poisoning and evasion attacks.
• Maintain a comprehensive software bill of materials (SBOM) for all dependencies.

Ensure training, tuning, and test datasets are statistically representative of the intended patient population. Selection bias is a critical failure mode that leads to health inequity.

• Define explicit inclusion/exclusion criteria based on demographic and phenotypic factors.
• Audit datasets for hidden stratification across subpopulations.
• Document the provenance, acquisition protocols, and annotation processes for all data sources.

Maintain strict separation between training and tuning datasets to prevent overfitting and inflated performance estimates. Data leakage is a primary cause of real-world model degradation.

• Physically or logically separate data partitions at the patient level, not the sample level.
• Use a held-out lockbox dataset that is never touched during iterative development.
• Pre-register the data split methodology before any model training begins.

Construct a reliable, clinically accepted ground truth based on expert consensus or established biomarkers. A model is only as valid as the labels it learns from.

• Use adjudicated expert panels to resolve ambiguous or conflicting labels.
• Quantify and report inter-rater variability to establish the noise ceiling.
• Justify the clinical validity of the reference standard relative to the target condition.

Select model architectures and regularization techniques appropriate for the size and complexity of the available data. Over-parameterization relative to data volume invites spurious correlations.

• Prefer intrinsically interpretable models when performance is equivalent to black-box alternatives.
• Apply inductive biases that align with known physiological or physical constraints.
• Document the rationale for architectural choices in the context of data limitations.

Evaluate the performance of the clinician-model interaction, not just the standalone model. The relevant metric is the combined decision-making accuracy of the human-AI dyad.

• Design user interfaces that mitigate automation bias and alert fatigue.
• Conduct think-aloud usability studies to observe how clinicians interpret model outputs.
• Measure the impact of the model on clinical decision speed and confidence calibration.

Proactively assess model performance on demographic, clinical, and edge-case subgroups defined during the risk assessment. Aggregate metrics can mask catastrophic subgroup failure.

• Report per-subgroup sensitivity and specificity with confidence intervals.
• Define a priori the minimum acceptable performance threshold for each subgroup.
• Test on intersectional subgroups to detect compounding biases.

Provide users with a Model Card or similar document that clearly communicates the device's intended use, limitations, and performance characteristics in plain language.

• Disclose the demographic composition of the training data.
• Explain the output format, confidence calibration, and known failure modes.
• Provide actionable guidance on how to interpret and act upon model outputs safely.

Implement a Predetermined Change Control Plan (PCCP) and continuous monitoring infrastructure to detect data drift, concept drift, and performance degradation in production.

• Monitor input data distributions for covariate shift relative to the training baseline.
• Establish statistical process control limits that trigger automated alerts and human review.
• Log all model predictions and associated metadata for post-market surveillance audits.