Auditing clinical prediction robustness under missing measurements and incomplete follow-up
Year: 2026
Type: article
Abstract: Robustness stress tests for clinical prediction models remove measurements and compare prediction error, but the outcomes used for that comparison may themselves be incompletely observed. We examined whether incomplete follow-up can distort the estimated effect of missing measurements. A controlled simulation separated measurement loss from censoring. A semi-synthetic experiment used SUPPORT clinical records divided by patient into development, validation and test sets of 5,416, 1,739 and 1,812 eligible records, retaining original known 180-day outcomes as a reference while artificial censoring was imposed. Five fixed prediction strategies and standard censoring-weighted estimators were evaluated. For indicator logistic regression, albumin-dependent measurement loss worsened reference Brier score by 0.00483, yet scoring only outcomes still known after albumin-driven censoring suggested an improvement of 0.00384. Masking-effect estimation error was ?0.00867 [?0.01033, ?0.00696] for known-outcome-only scoring and ?0.00167 [?0.00303, ?0.00017] for marginal inverse-probability-of-censoring weighting, conditional on the fitted source model; corresponding source-refitting validation intervals included zero. Other pipelines showed exaggerated rather than reversed deterioration. Exploratory hepatocellular carcinoma analyses in 165 UCI and 366 TCGA records motivated the study; their original augmentation contrast was inconclusive. Robustness evaluations should report how outcomes were observed and distinguish fixed-model from source-refitting uncertainty. Keywords: Missing Predictors, Censoring, Brier Score, Clinical Prediction, Robustness, Reproducibility.
Authors: Hamza El Massaoudy, EL Houcine Oussi
Language: English
Subfield: Statistics and Probability
Field: Mathematics
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7236637173
