The paper stratifies EHRShot test patients into quartiles by 1-gram repetition rate and by standard deviation of inter-event times, then measures CLMBR-T-BASE's average brier score across all 14 prediction tasks in each quartile. Brier score rises monotonically from 0.0647 (least repetitive) to 0.0805 (most repetitive) and from 0.0683 (least irregular) to 0.0777 (most irregular). The most irregular patients show a 14% higher brier score than the least irregular, a figure the authors highlight in the abstract as a representative example of the degradation pattern.
Evidence
correlational
Key metric
brier score 0.0647 (Q1) to 0.0805 (Q4) by 1-gram repetition rate; 0.0683 (Q1) to 0.0777 (Q4) by inter-event time standard deviation; 14% higher brier loss for most vs least irregular patients
Caveat
The authors note the repetition and irregularity metrics are 'fairly simple and could be improved,' and that the pretraining dataset was sourced from a single institution, which may limit generalizability.