View a PDF of the paper titled Physiological Trajectories and Observation Patterns in ICU Mortality Prediction: Transportability Under Changing Measurement Conditions, by Benjamin T. Pike
Abstract:Objective: To examine how changes in clinical measurement practices affect the transportability of predictive information from physiological trajectories and observation patterns in longitudinal ICU records.
Methods: In-hospital mortality was predicted using measurements from the first 48 hours of ICU admission in MIMIC-III, MIMIC-IV, and eICU. Nested representations captured the latest physiological state (R1), trajectory summaries (R2), and observation timing and frequency (R3). Models trained on 7,762 MetaVision patients were evaluated on 21,682 MIMIC-IV and 43,878 eICU patients from 204 hospitals. Restriction experiments tested measurement timing at matched counts and reduced availability.
Results: Physiological trajectories improved prediction beyond latest values across settings, including transfer to eICU (R2-R1 log-loss difference -0.01446, 95% CI -0.01701 to -0.01202). Observation patterns improved prediction following transfer to MIMIC-IV (R3-R2 -0.00564) and in locally developed eICU models (-0.00402, 95% CI -0.00517 to -0.00283), but not clearly following transfer from MetaVision to eICU (+0.00106, 95% CI -0.00087 to +0.00355). At matched observation counts, retaining measurements spread throughout the 48-hour window reduced log loss relative to random retention, more for R3 than R2 (0.04324 versus 0.00721). R3 was also more sensitive to the loss of recent measurements. Retraining under reduced availability restored its advantage over R2.
Conclusions: Physiological trajectories transferred consistently, while observation patterns offered less reliable benefits and greater sensitivity to measurement timing and availability. Their transferability may depend on diversity of measurement practices during training. Clinical prediction models should be evaluated across measurement conditions as well as patient populations.
Submission history
From: Benjamin T. Pike [view email]
[v1]
Fri, 30 Oct 2026 17:56:10 UTC (557 KB)
[v2]
Thu, 19 Nov 2026 00:34:48 UTC (556 KB)
