Asymmetric Evidence and Belief Revision in ICU Deterioration Forecasting

• Rafael Mendes

  Departamento de Computação Aplicada, Universidade Federal do Pampa, Avenida Tiarajú 810, Bairro Ibirapuitã, Alegrete, Brazil

  Author

• Thiago Ribeiro

  Departamento de Sistemas de Informação, Universidade Estadual do Norte do Paraná, Rodovia PR-160 Km 0, Jardim Panorama, Cornélio Procópio, Brazil

  Author

Intensive care forecasting is frequently described as the problem of predicting near-future clinical deterioration from recent electronic health record data. That description is correct at a high level, yet it suppresses the computational asymmetry that makes the task difficult. The structured record and the documentary record do not play equivalent roles. Vital signs and laboratory values form a sparse and irregular trace of monitored physiology, while notes, reports, and bedside assessments provide compressed interpretations whose semantic richness exceeds their temporal precision. A practical deterioration model must therefore do more than combine modalities. It must maintain a revisable internal belief about patient state while deciding how strongly numeric trajectories and narrative context should influence that belief at each moment. This paper develops a computer-science-centered framework for ICU deterioration prediction built around representational asymmetry, belief revision, and support-aware priority formation. The proposed view treats structured signals as primary carriers of short-horizon physiologic dynamics and clinical text as a higher-level contextual operator that sharpens, softens, or redirects latent hypotheses rather than replacing the temporal stream outright. A dual-timescale memory model is introduced to separate rapidly changing monitored evidence from slowly decaying contextual evidence, followed by a belief-update rule that explicitly represents contradiction, support strength, and residual uncertainty. The discussion then turns to training objectives that prevent modality collapse, maintain a useful upper tail under severe class imbalance, and link calibrated prediction to constrained review capacity. The result is a framework in which forecasting, confidence, and triage are understood as coupled problems of sequential representation rather than as isolated steps in a standard classification pipeline.

• PDF

Copyright (c) 2026 International Journal of Scientific and Applied Computing Research

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.