Validation split integrity record ================================= StatMedAnalysis | Original downloadable worksheet | 6 November 2026 Keep related observations and future information out of the wrong partition. How to use this worksheet Define the unit that must remain together, such as person, site or time series. Assign splits using the intended deployment scenario. Fit preprocessing and feature selection only within training data. Audit identifier overlap and chronology using aggregate diagnostics before evaluating the held-out set. Save a working copy and keep the original blank version. Use public, hypothetical or aggregate information only; do not enter patient identifiers, credentials or confidential records. Keep source evidence in your approved project workspace. Mark unresolved text fields UNKNOWN or NOT REPORTED with a reason; never substitute zero for an unknown count. Project/version __________________ Prepared by role/date __________________ Working fields 1. Deployment setting Response/evidence: _________________________________________________________ ___________________________________________________________________________ 2. Grouping unit Response/evidence: _________________________________________________________ ___________________________________________________________________________ 3. Training and validation periods Response/evidence: _________________________________________________________ ___________________________________________________________________________ 4. Split rule and random state Response/evidence: _________________________________________________________ ___________________________________________________________________________ 5. Group overlap count Response/evidence: _________________________________________________________ ___________________________________________________________________________ 6. Preprocessing fit locations Response/evidence: _________________________________________________________ ___________________________________________________________________________ 7. Feature availability at prediction time Response/evidence: _________________________________________________________ ___________________________________________________________________________ 8. Leakage correction and rerun Response/evidence: _________________________________________________________ ___________________________________________________________________________ Hypothetical worked case A fictional dataset has three visits per person. A random row split places the same person in training and validation, inflating apparent generalization; group-based splitting is required for the intended new-person prediction task. Decision rule and limitations Reject a validation result when the split violates the target deployment independence or uses future information. Recreate the full training pipeline after repairing leakage. Decision _______________________ Remaining uncertainty ____________________ Next action ____________________ Owner / review date ______________________ Independent check and discrepancy resolution _______________________________ Official method or reporting reference checked 6 November 2026 https://scikit-learn.org/stable/common_pitfalls.html This is an original planning aid, not an official checklist, validated instrument, clinical recommendation or certificate of expert approval. Apply the referenced guidance to the specific design and record any additional requirements.