<![CDATA[Continuous monitoring of physiological conditions in intensive care units generates complex multivariate time-series data that can support early clinical decision-making. However, conventional hybrid models combining Long Short-Term Memory networks and XGBoost generally learn correlational patterns and may experience performance degradation under non-stationary conditions, sensor artifacts, and changes in patient characteristics. Moreover, these models are typically designed for outcome prediction rather than treatment-effect estimation. Therefore, a framework capable of integrating temporal forecasting, representation disentanglement, uncertainty estimation, and treatment-response analysis is required. Purpose: This study aims to develop a Causal Multi-Task Fusion Engine (CMTFE) for forecasting future vital-sign trajectories and estimating individualized treatment effects in IoT-supported ICU monitoring. Methods/Study Design/Approach: The proposed framework integrates a Temporal Convolutional Network for multivariate vital-sign encoding, a β-Variational Autoencoder for separating outcome-relevant and confounding representations, a Transformer decoder for probabilistic trajectory forecasting, and propensity-weighted XGBoost regressors for treatment-effect estimation. Experiments were conducted using 487,000 observation windows derived from 38,200 ICU stays in the MIMIC-IV dataset. The data were chronologically divided into training, validation, and testing sets at proportions of 70%, 15%, and 15%, respectively. Forecasting performance was evaluated using Mean Absolute Error, Root Mean Squared Error, and Continuous Ranked Probability Score. Treatment-effect estimation was evaluated on a semi-synthetic test set using Precision in Estimation of Heterogeneous Effect and Area Under the Receiver Operating Characteristic Curve. Results/Findings: CMTFE achieved an MAE of 0.281, RMSE of 0.390, and CRPS of 0.218, outperforming the evaluated LSTM, TCN, Transformer, LSTM–XGBoost, and TCN–XGBoost baselines. Under the constructed distribution-shift scenario, the model exhibited a relative MAE increase of 4.2%, compared with 9.5% for TCN–XGBoost. For semi-synthetic treatment-effect estimation, CMTFE obtained a PEHE of 0.187 and an AUC of 0.842. The ablation analysis further indicated that representation disentanglement, contrastive regularization, propensity-score weighting, and temporal convolution each contributed to the model’s performance. Novelty/Originality/Value: The principal contribution of this study lies in integrating temporal forecasting, disentangled representation learning, uncertainty quantification, and propensity-adjusted treatment-effect estimation within a unified ICU-monitoring framework. Unlike conventional hybrid models that combine neural and tree-based predictions through passive aggregation, CMTFE uses learned outcome-relevant representations to support both forecasting and intervention-response analysis. Nevertheless, the treatment-effect findings are based on semi-synthetic outcomes and should be interpreted as experimental decision-support evidence rather than validated clinical recommendations. External, prospective, and clinician-centered validation is required before real-world deployment.]]>
