Causal Multi-Task Fusion Engine with Disentangled Representations for IoT-Based Clinical Recommendation in ICU Monitoring

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Continuous monitoring through Internet of Things (IoT) sensors provides valuable physiological data for clinical decision support in intensive care units. However, conventional hybrid models, such as LSTM–XGBoost, primarily learn correlations and often fail to distinguish stable causal mechanisms from spurious patterns caused by patient heterogeneity, sensor artifacts, and distribution shifts. Consequently, their forecasting accuracy and treatment recommendations may become unreliable under non-stationary clinical conditions. Purpose: This study aims to develop and evaluate a Causal Multi-Task Fusion Engine (CMTFE) for robust vital-sign forecasting, individualized treatment-effect estimation, and clinical intervention recommendation in IoT-based ICU monitoring. Methods/Study Design/Approach: This study employed an experimental quantitative design using the MIMIC-IV dataset. After preprocessing, 487,000 observation windows from 38,200 ICU stays were divided chronologically into training, validation, and testing sets. The proposed CMTFE integrates a Temporal Convolutional Network for encoding multivariate vital-sign sequences, a disentangled variational autoencoder for separating causal and confounding representations, a Transformer decoder for probabilistic trajectory forecasting, and a propensity-weighted multi-task XGBoost model for counterfactual treatment-effect estimation. Performance was evaluated using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Continuous Ranked Probability Score (CRPS), Precision in Estimation of Heterogeneous Effect (PEHE), and Area Under the Curve (AUC). Results/Findings: CMTFE achieved an MAE of 0.281, RMSE of 0.390, and CRPS of 0.218, reducing MAE by 18.3% and RMSE by 16.7% compared with the LSTM–XGBoost baseline. Under distribution shift, CMTFE recorded only a 4.2% relative MAE increase, compared with 9.5% for TCN–XGBoost. For treatment-effect estimation, the model achieved a PEHE of 0.187 and an AUC of 0.842, representing a 31.5% reduction in PEHE compared with the best causal baseline. Ablation results further confirmed the contribution of each architectural component. Novelty/Originality/Value: The novelty of this study lies in (i) integrating causal disentanglement into IoT-based ICU time-series modeling, (ii) jointly optimizing probabilistic forecasting and individualized treatment-effect estimation, and (iii) transforming passive hybrid prediction into an active clinical recommendation framework. The proposed approach provides a robust foundation for developing reliable clinical decision-support systems under non-stationary ICU conditions.