Multimodal Prediction of Early Clinical Deterioration in Adult Intensive Care Units: A Multi-Center Study

Abstract

Large intensive care units generate dense streams of physiological observations, laboratory values, orders, and narrative documentation, yet the translation of these heterogeneous data into stable early warning signals remains uneven across hospitals. In China, this challenge is intensified by cross-hospital variation in documentation practice, measurement frequency, and information system architecture. We present an empirical study of early deterioration prediction using a multimodal learning framework designed for adult intensive care units in a multi-center Chinese setting. The study uses de-identified electronic health records from six tertiary hospitals spanning northern, eastern, and southwestern China, with hourly prediction windows constructed from vital signs, laboratory trajectories, medication and device indicators, and Mandarin clinical notes. The target outcome is a composite deterioration event within the next 24 hours, defined as ICU death, initiation of invasive mechanical ventilation, vasopressor initiation, or emergency continuous renal replacement therapy. The proposed model combines a missingness-aware temporal transformer for structured signals, a domain-adapted Chinese clinical language encoder for free text, and a hospital-invariant fusion module trained to preserve predictive content while reducing site-specific bias. Across 5,931,284 prediction windows from 118,406 ICU stays, the multimodal model achieves an internal test AUROC of 0.867 and AUPRC of 0.352, outperforming structured-only, text-only, logistic regression, gradient boosting, and recurrent baselines. On a fully held-out external hospital, performance remains stable with AUROC 0.846 and AUPRC 0.327. Calibration improves materially after temperature scaling, and operational simulation shows that ranking by model score can concentrate a substantial fraction of future deterioration events within a manageable alert budget. The findings support multimodal, cross-hospital, and calibration-aware modeling as practical directions for ICU risk estimation in Chinese critical care systems.