Auditing and Enforcing Calendar-Shift Invariance in Clinical Prediction Models

Reliable temporal forecasting models should produce forecasts that remain unchanged when only an arbitrary calendar origin changes. We audit this dependence in clinical records under equivalent calendar shifts with model weights and scoring fixed. In AACR Project GENIE non small cell lung cancer data, the average patient’s log score varies by 0.098 nats per completed gap while the pooled score remains nearly unchanged. Relative-phase removes this dependence while improving held out log scores over Absolute-phase. X-align matches candidate assessment times to prior assessment patterns, improving geometric-mean probability assigned to observed assessment days by 9.5% in non small cell lung cancer and 12.1% in colorectal cancer over a baseline already using weekly timing features. In external glucose audits, randomized training anchors reduce calendar induced category changes from 56.34% to 0.42% in a published NHiTS model. A relative calendar origin eliminates high-glucose classification changes affecting 19.7% of CGMacros patients, with higher reference prediction error. These results motivate calendar consistency checks for forecasting benchmarks built from shifted or synthetic dates.