ICH M14: Real-World Data for Safety Assessment

Only the statistical significance of the estimate The difference between the target and study populations and the resulting uncertainty about generalisability That the study is invalid because any database represents fewer than all patients That large sample size makes population differences irrelevant

Immortal-time or other time-related bias from incoherent exposure and follow-up definitions Confounding by indication that is necessarily solved by matching Loss of external validity caused by using a prescription record A data-provenance problem that affects only the final report

The algorithm is automatically valid because predictive value is an intrinsic property of the code The prior evidence should be assessed for applicability to the new source, coding system, population and clinical context The algorithm should be rejected because validation studies cannot be reused Outcome validation is unnecessary if the study includes sensitivity analyses

RWD are routinely collected patient or healthcare data; RWE is evidence generated by analysing such data in a study RWD are data from randomised trials; RWE is data from routine clinical practice RWD are individual case reports; RWE is the total number of reports in a database RWD are data that have been validated; RWE is any unanalysed healthcare dataset

They are interchangeable terms for database size Fit-for-use addresses whether data suit the question; QA/QC controls help prevent or detect errors during study processes QA/QC establishes clinical relevance, while fit-for-use applies only to programming A fit-for-use assessment removes the need to control data transformations

Codes have identical meanings whenever their labels match The meaning may depend on how the data were generated, recorded, transformed and used operationally A diagnosis code is reliable only if it was entered by a regulator Provenance is relevant only to digital-health data, not healthcare records

Every recommendation is automatically a statutory obligation It indicates a recommendation in the guideline, while binding obligations may arise separately from applicable legislation or PASS procedures It means the recommendation is optional even when another EU requirement makes the activity mandatory It has no significance for regulatory interpretation

A common model guarantees identical clinical practice and source-data provenance at every site Underlying differences in coding, care, data quality and ascertainment may remain despite standardised structure Federated studies cannot use standardised analytic programs Data harmonisation makes site-level quality checks unnecessary

Use it without further assessment because machine learning adapts automatically to any source Describe the algorithm and its development and assess whether its performance is valid in the proposed study context Exclude all machine-learning methods from non-interventional studies Rely on the model's complexity instead of documenting training data or validation

Feasibility is prohibited from using any patient counts or outcome-frequency information The process has allowed comparative outcome information to influence protocol choices before the study is finalised Feasibility must be performed only after the final report is submitted A protocol cannot be amended after a feasibility assessment