CBT-ICU transformer: a clinician-behaviour-aware multimodal deep learning framework for early detection of critical care adverse events in MIMIC-IV

Abstract

Early identification of clinical deterioration in intensive care units (ICUs) remains a major challenge despite the availability of large-scale electronic health records. Most existing models focus primarily on physiological measurements such as vital signs and laboratory results, while overlooking implicit clinical signals embedded in care delivery patterns. In this study, we propose a novel multi-modal framework, the CBT-ICU Transformer (Clinician Behaviour–Temporal ICU Transformer), for predicting early composite critical care adverse events (CCAE) using the MIMIC-IV v3.1 dataset. The proposed model integrates three complementary information sources: structured clinical features, 24-hour physiological time-series data, and clinician monitoring intensity. A composite deterioration label is defined as vasopressor initiation within 72 h or in-hospital mortality, enabling the model to capture early instability and severe progression. Temporal physiological signals are modelled using a Transformer encoder to capture dynamic dependencies, while monitoring intensity is encoded through a dedicated convolutional behaviour pathway to represent implicit clinical concern. Structured laboratory data are processed using a missingness-aware representation to preserve informative patterns of test ordering. These representations are fused within a unified architecture for end-to-end learning. Experimental evaluation demonstrates strong predictive performance (ROC-AUC: 0.87, PR-AUC: 0.81) with balanced precision and recall. By explicitly modelling clinician behaviour alongside patient physiology, the proposed approach offers a behaviour-aware perspective for early prediction of ICU deterioration and provides a more clinically contextualised early warning signal. External multi-centre validation is required to confirm generalisability across diverse ICU settings. The contribution of the clinician behaviour branch is that of an improvement in PR-AUC (ΔPR-AUC = − 0.0071, p = 0.02), while there is no significant contribution to ROC-AUC ( p = 0.36); such a range of contributions is clearly defined as one of the limitations of the behavioural signal.