From Sentiment to Signal: Narrative–Physiology Discordance as a Testable Target for Artificial Intelligence in Critical Care

Intensive care units generate dense physiological, laboratory, imaging, microbiological and treatment data, yet much of the clinical reasoning that drives decisions is recorded only in free text. Over the past decade, clinical sentiment analysis has repeatedly shown that the affective tone of nursing and medical notes is associated with mortality. The incremental discrimination over established severity scores has nevertheless been small, of the order of 0.01 in the area under the receiver operating characteristic curve in the largest published intensive care cohort, and general-purpose sentiment tools transfer poorly to critical care text. We argue that this plateau reflects a misspecified prediction target rather than a limitation of natural language processing. Mortality at 28 days is not the question the clinician asks at the bedside. We propose that the clinically useful quantity is the narrative–physiology discordance score: the signed difference, expressed on a common calibrated scale, between the short-horizon deterioration risk implied by the documented assessment and the risk implied by time-aligned multimodal data. We specify this quantity formally, fix index time and forecast horizon, and set out the leakage, copy-forward and provenance controls without which retrospective performance is uninterpretable. We then treat alert burden as a design constraint rather than a post hoc observation, deriving the operating threshold from an explicit and context-dependent alert budget, and we outline a staged evaluation pathway aligned with TRIPOD + AI and DECIDE-AI in which discordant cases are adjudicated by clinicians. Until that adjudication has been performed, discordance is a record-level inconsistency that prompts reassessment, not evidence of missed deterioration. Clinical sentiment analysis becomes useful when it stops predicting death and starts flagging disagreement.