REVELIO Planned Participant-Level Validation to Separate New Drivers From Repeated Sensor Windows
The REVELIO protocol plans to keep each participant’s data entirely within either training or testing during model validation. This addresses a specific risk: repeated sensor windows from one driver can make an apparent test less independent than it looks.
Many windows do not create many drivers
The planned high-frequency data will be divided into sessions, scenarios and time windows. Windows from the same person can share driving habits and physiological characteristics. If they appeared on both sides of a model evaluation, performance could partly reflect recognition of that person rather than generalization to an unseen driver.
The protocol specifies grouped evaluation
The authors propose leave-one-subject-out or equivalent grouped cross-validation. Preprocessing and feature selection are to be learned within training data and applied unchanged to held-out data. Nested procedures are planned where feasible. These are methodological commitments in a protocol, not evidence that a completed model achieved a particular accuracy.
Internal separation does not complete external validation
Holding out participants helps assess performance within the recruited sample. It does not establish transportability to new roads, devices, populations or competing causes of poor driving. The publication supplies a reviewable strategy for limiting information leakage. A future result should be assessed against what was actually implemented, while retaining the distinction between participant-level internal validation and independent external testing.
Related analysis: the source’s broader findings and limitations.
