Discordant Care as a Computable Phenotype for Real-Time Detection of Routine Protocol Completion Without Cognitive Patient Engagement in the ICU: Retrospective Cohort Study

ABSTRACT Background Quality measurement in intensive care emphasizes task completion—whether assessments were documented and protocols followed. Electronic health record (EHR) systems capture these signals in real time, yet current metrics cannot distinguish task completion from cognitive clinical engagement. A prior analysis demonstrated that omission of orientation assessment predicted a 4.29-fold increase in hospital mortality among low-acuity ICU patients [1]. Whether combining this marker with routine task-completion data yields a computable phenotype with independent prognostic value has not been studied. Objective To define, validate, and characterize “discordant care”—a computable EHR phenotype defined as completion of ≥6 of 8 routine nursing assessments without orientation assessment documentation—as a predictor of hospital mortality, distinguishing patient-level confounding from care process signal. Methods Retrospective cohort study using MIMIC-IV v3.1 (2008–2022), including 46,004 adult ICU stays with SOFA scores 0-2 and length of stay ≥24 hours in non-neurological ICUs. Primary exposure: discordant care, computed from structured nursing flowsheet data within 24 hours of admission. Primary outcome: hospital mortality. Progressive covariate adjustment included mechanical ventilation, sedation, and diagnosis. Results Discordant care was present in 8891 patients (19.3%), with 69.7% mechanically ventilated versus 25.3% of concordant patients. Two overlapping signals were identified: a patient-level signal driven by ventilation/sedation (full adjustment OR 1.19, 95% CI 1.09–1.30) and a care process signal in non-ventilated patients (OR 2.14, 1.87–2.44; N=30,314). Among non-ventilated SOFA 0 patients, OR was 2.60 (2.13–3.18; N=16,295). The signal was present across all 7 major diagnosis categories. Quantitative bias analysis indicated unmeasured delirium could attenuate but likely not fully explain the non-ventilated signal. Conclusions Discordant care identifies two phenomena: a patient-level signal from ventilation/sedation and a care process signal where assessable patients receive routine care without cognitive engagement (OR 2.14–2.60). This care process signal is invisible to existing quality metrics and detectable in real time. Prospective validation with systematic delirium screening is needed.