AI-Based Prediction of Clinical Trial Database Lock Readiness
DOI: 10.66838/ma.239
|
Abstract Database lock is the point at which a clinical trial’s data are declared final and statistical analysis can begin, and the weeks between the last patient’s last visit and lock are among the most closely watched in drug development. Lock dates are planned months ahead, are frequently missed, and are usually forecast by hand from status reports of open queries, missing pages, coding and reconciliation backlogs. This paper asks whether machine learning applied to weekly snapshots of those backlogs can forecast the remaining time to lock, flag that a study will miss its plan, and declare readiness for lock with calibrated confidence. We build a reproducible simulation of a portfolio of 150 studies run by five contract research organisations, in which eight backlog categories are worked down at organisation-specific rates, interrupted by staffing dips, vendor data transfers and late safety findings, followed by a variable sign-off period, and in which the planned lock date uses only information available at the last patient visit. We compare the planned date, a burn-down extrapolation, a ridge regression, gradient-boosted trees on current counts and gradient-boosted trees with trajectory features (changes, burn rates and projected clearance times), with leave-one-organisation-out evaluation over five seeds. Across all snapshots the trajectory model had a mean absolute error of 3.8 weeks against 9.3 weeks for the plan and 5.9 for burn-down extrapolation; eight weeks before lock its error was 1.8 weeks against 4.5 and 3.9, and four weeks before lock 1.3 against 2.9 and 1.8. Two weeks before lock it was no better than the plan (2.5 against 2.0 weeks). The probability that lock is within four weeks had an area under the curve of 0.93 and an expected calibration error of 0.04; a gate that declared readiness at probability 0.8 was reached before lock in only 29% of studies and was wrong by more than two weeks in 10.6% of those. Four weeks after the last visit, a burn-down extrapolation and the learned model flagged studies that would slip with similar skill (area under the curve 0.90 and 0.89), and the plan itself has none. Prediction intervals were too narrow (70% coverage at a nominal 80% eight weeks before lock, and none two weeks before). Open queries dominated the explanations. The data are synthetic, and the study shows what such forecasts can and cannot do, not how a real portfolio would perform.
Keywords
Database lock, Clinical data management, Time-to-event forecasting, Gradient boosting, Quantile regression, Calibration, Study conduct., |