多重器官功能障礙症候群（MODS）是 ICU 中常見導致重症病人死亡的原因之一。 ICU 病人生理狀態具高度動態性，可能在短時間即有多器官系統惡化的可能性，若未能及早辨識並介入，可能導致 MODS。如何從複雜快速的 ICU 臨床時間序列資料中及早偵測器官惡化趨勢，是智慧醫療早期預警的重要課題。現有模型多著重於單一臨床終點，雖可提供風險預測，但缺乏器官層級資訊。多任務學習可以同時預測多個器官系統的惡化狀態，較符合 MODS 照護中對器官層級預警的需求。然而不同器官惡化模式具有異質性，若模型只依賴共同的病人生理表徵，可能無法充分保留各器官獨有的惡化訊號，而若跨器官資訊未經篩選便直接傳遞，也可能將不相關或帶有雜訊的訊號引入預測過程。
為解決上述問題，本研究提出 Structure-Aware Disentangled Graph Network Multi-Task（SADG-MT）模型，用於 ICU 病人六個 SOFA 定義器官系統之逐小時惡化預測。 SADG-MT建立於 IOC-MT 架構之上，並針對特徵同質化與跨器官雜訊傳遞問題來改進。首先，本研究設計 Organ-Query disentanglement 機制，使各器官透過可學習查詢向量從共享時間序列表徵中擷取相關生理特徵，保留器官特定惡化訊號。其次，本研究提出 Edge-Gated Graph Attention 模組，在器官間圖注意力機制中加入邊層級可學習閘門，調節訊息傳遞強度，降低弱相關訊號造成的干擾。
本研究以 MIMIC-IV 為主要開發資料集，並以 MIMIC-III、eICU-CRD 及去識別化本地 ICU cohort 作為驗證。結果顯示，SADG-MT 相較於 IOC-MT 將 macro AUROC 由 0.834 提升至 0.894 ，macro AUPRC 由 0.306 提升至 0.368 ，於六項器官系統中皆取得較佳表現。進一步分析來看，模型在 CNS、Liver 與 Renal 等依賴間接或異質生理訊號的器官任務中具有較明顯提升。此外，SADG-MT 在 MIMIC-IV 時間對齊誤差情境下維持較佳穩健性，並於 MIMIC-III、eICU-CRD 與本地 ICU cohort 的外部驗證中亦展現穩定表現。除二元器官惡化發生預測外，本模型亦可延伸至三分類器官趨勢預測與未來最大 SOFA 分數估計。
整體而言，結構感知的器官特定表徵解耦與選擇性跨器官資訊聚合，有助於提升 ICU 病人動態多器官惡化預測之表現。SADG-MT 不僅能提供整體惡化風險判斷，更能產生符合 SOFA 定義的器官層級預警資訊，具備作為 Human Healthcare Digital Twin 架構下 ICU 智慧預警模型之應用潛力。

Multiple Organ Dysfunction Syndrome (MODS) is a severe clinical condition in intensive care units and remains a major contributor to mortality among critically ill patients. Early identification of organ-level deterioration is therefore essential for timely intervention. However, existing multi-organ prediction models may suffer from feature homogenization and noisy inter-organ information propagation, limiting their ability to capture organspecific deterioration patterns.
This thesis proposes SADG-MT, a Structure-Aware Disentangled Graph Multi-Task model for hourly deterioration prediction across six SOFA-defined organ systems. Built upon IOC-MT, SADG-MT introduces an Organ-Query mechanism to extract organ-specific physiological representations and an Edge-Gated Graph Attention module to regulate inter-organ message passing.
This study developed SADG-MT on MIMIC-IV and validated it on MIMIC-III, eICUCRD, and a de-identified local ICU cohort. Compared with IOC-MT, SADG-MT improved macro AUROC from 0.834 to 0.894 and macro AUPRC from 0.306 to 0.368 across all six organ systems. It also showed stable external validation, robustness to temporal alignment errors, and extensibility to organ trend prediction and future maximum SOFA estimation. These results suggest that structure-aware organ disentanglement and selective inter-organ aggregation can improve dynamic multi-organ deterioration prediction in ICU patients.

[1] G. Halpern et al., “Advances and Utility of Digital Twins in Critical Care and Acute Care Medicine: A Narrative Review,” Journal of Yeungnam Medical Science, 2024.
[2] X. Feng, S. Zhu, Y. Shen, H. Zhu, M. Yan, G. Cai, and G. Ning, “Multi-organ spatiotemporal information aware model for sepsis mortality prediction,” Artificial Intelligence in Medicine, vol. 147, Art. no. 102746, 2024.
[3] M. J. Murray and D. B. Coursin, “Multiple Organ Dysfunction Syndrome,” The Yale Journal of Biology and Medicine, vol. 66, no. 5, pp. 501–510, 1993.
[4] N. Mahmud et al., “Risk prediction scores for acute on chronic liver failure development and mortality,” Liver International, vol. 40, no. 5, pp. 1159–1167, 2020.
[5] R. Malhotra et al., “A risk prediction score for acute kidney injury in the intensive care unit,” Nephrology Dialysis Transplantation, vol. 32, no. 5, pp. 814–822, 2017.
[6] F. Li et al., “Prediction model of in-hospital mortality in intensive care unit patients with heart failure: machine learning-based, retrospective analysis of the MIMIC-III database,” BMJ Open, vol. 11, no. 7, Art. no. e044779, 2021.
[7] S. Roy et al., “Multitask prediction of organ dysfunction in the intensive care unit using sequential subnetwork routing,” Journal of the American Medical Informatics Association, vol. 28, no. 9, pp. 1936–1946, 2021.
[8] Z. Zeng et al., “Inter-organ correlation based multi-task deep learning model for dynamically predicting functional deterioration in multiple organ systems of ICU patients,” BioData Mining, vol. 18, no. 1, Art. no. 27, 2025.
[9] F. Li, S. Wang, Z. Gao, M. Qing, S. Pan, Y. Liu, and C. Hu, “Harnessing artificial intelligence in sepsis care: advances in early detection, personalized treatment, and real-time monitoring,” Frontiers in Medicine, vol. 11, Art. no. 1510792, 2025.
[10] N. Tomašev et al., “A clinically applicable approach to continuous prediction of future acute kidney injury,” Nature, vol. 572, no. 7767, pp. 116–119, 2019.
[11] I. Bendavid et al., “A novel machine learning model to predict respiratory failure and invasive mechanical ventilation in critically ill patients,” Scientific Reports, vol. 12, Art. no. 10577, 2022.
[12] I. Bendavid, L. Statlender, L. Shvartser, S. Teppler, R. Azullay, R. Sapir, and P. Singer, “A novel machine learning model to predict respiratory failure and invasive mechanical ventilation in critically ill patients suffering from COVID-19,” Scientific Reports, vol. 12, Art. no. 10573, 2022.
[13] A. Bhattacharyya, S. Sheikhalishahi, H. Torbic, W. Yeung, T. Wang, J. Birst, A. Duggal, L. A. Celi, and V. Osmani, “Delirium prediction in the ICU: designing a screening tool for preventive interventions,” JAMIA Open, vol. 5, no. 2, Art. no. ooac048, 2022.
[14] Y. Tan, M. Dede, V. Mohanty, J. Dou, H. Hill, E. V. Bernstam, and K. Chen, “Forecasting acute kidney injury and resource utilization in ICU patients using longitudinal, multimodal models,” Journal of Biomedical Informatics, vol. 154, Art. no. 104648, 2024.
[15] H. Lee, H.-L. Yang, H. G. Ryu, C.-W. Jung, Y. J. Cho, S. B. Yoon, H.-K. Yoon, and H.-C. Lee, “Real-time machine learning model to predict in-hospital cardiac arrest using heart rate variability in ICU,” npj Digital Medicine, vol. 6, Art. no. 215, 2023.
[16] J. Lei, J. Zhai, Y. Zhang, J. Qi, and C. Sun, “Supervised Machine Learning Models for Predicting Sepsis-Associated Liver Injury in Patients With Sepsis: Development and Validation Study Based on a Multicenter Cohort Study,” Journal of Medical Internet Research, vol. 27, Art. no. e66733, 2025.
[17] R. Caruana, “Multitask Learning,” Machine Learning, vol. 28, pp. 41–75, 1997.
[18] S. Ruder, “An overview of multi-task learning in deep neural networks,” arXiv preprint arXiv:1706.05098, 2017.
[19] M. B. A. McDermott, B. Nestor, E. Kim, W. Zhang, A. Goldenberg, P. Szolovits, and M. Ghassemi, “A Comprehensive Evaluation of Multi-task Learning and Multi-task Pre-training on EHR Time-series Data,” arXiv preprint arXiv:2007.10185, 2020.
[20] S. Liu, E. Johns, and A. J. Davison, “End-to-End Multi-Task Learning with Attention,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019.
[21] J. Ma, Z. Zhao, X. Yi, J. Chen, L. Hong, and E. H. Chi, “Modeling Task Relationships in Multi-task Learning with Multi-gate Mixture-of-Experts,” in Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD), 2018.
[22] J. Hu, L. Shen, and G. Sun, “Squeeze-and-Excitation Networks,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018.
[23] E. Perez, F. Strub, H. de Vries, V. Dumoulin, and A. Courville, “FiLM: Visual Reasoning with a General Conditioning Layer,” in Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2018.
[24] E. Choi, M. T. Bahadori, L. Song, W. F. Stewart, and J. Sun, “GRAM: Graph-based Attention Model for Healthcare Representation Learning,” in Proc. 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2017, pp. 787–795.
[25] E. Rocheteau et al., “Temporal pointwise convolutional networks for length of stay prediction in the intensive care unit,” in Proc. Conference on Health, Inference, and Learning, 2021, pp. 58–68.
[26] Y. Xu et al., “Time-Aware Context-Gated Graph Attention Network for Clinical Risk Prediction,” IEEE Transactions on Knowledge and Data Engineering, vol. 35, no. 7, pp. 6793–6805, 2023.
[27] H. O. Boll et al., “Graph neural networks for clinical risk prediction based on electronic health records: A systematic review,” Journal of Biomedical Informatics, vol. 151, Art. no. 104616, 2024.
[28] P. Veličković et al., “Graph Attention Networks,” in Proc. International Conference on Learning Representations, 2018.
[29] J. Chen and H. Chen, “Edge-Featured Graph Attention Network,” arXiv preprint arXiv:2101.07671, 2021.
[30] X. Bresson and T. Laurent, “Residual Gated Graph ConvNets,” in Proc. International Conference on Learning Representations (ICLR), 2018.
[31] C. Chen, S. Hou, Y. Yuan, G. Zhong, and L. Qiao, “Edge-Gated Graph Attention Network for Graph Classification,” in Proc. IEEE International Conference on Big Data and Smart Computing (BigComp), 2026, pp. 358–361.
[32] J. Chen, K. P. Seng, L. M. Ang, J. Smith, and H. Xu, “Towards a Gated Graph Neural Network with an Attention Mechanism for Audio Features with a Situation Awareness Application,” Electronics, vol. 14, no. 13, Art. no. 2621, 2025.
[33] C. Guo, G. Pleiss, Y. Sun, and K. Q. Weinberger, “On Calibration of Modern Neural Networks,” in Proc. 34th International Conference on Machine Learning, 2017, pp. 1321–1330.
[34] T.-Y. Lin, P. Goyal, R. Girshick, K. He, and P. Dollár, “Focal Loss for Dense Object Detection,” in Proc. IEEE International Conference on Computer Vision, 2017, pp. 2980–2988.
[35] Y. Cui, M. Jia, T.-Y. Lin, Y. Song, and S. Belongie, “Class-Balanced Loss Based on Effective Number of Samples,” in Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019, pp. 9268–9277.
[36] A. Johnson, L. Bulgarelli, T. Pollard, B. Gow, B. Moody, S. Horng, L. A. Celi, and R. Mark, “MIMIC-IV,” PhysioNet, version 3.1, 2024, doi: 10.13026/kpb9-mt58.
[37] A. Johnson, T. Pollard, and R. Mark, “MIMIC-III Clinical Database,” PhysioNet, version 1.4, 2016, doi: 10.13026/C2XW26.
[38] T. Pollard, A. Johnson, J. Raffa, L. A. Celi, O. Badawi, and R. Mark, “eICU Collaborative Research Database,” PhysioNet, version 2.0, 2019, doi: 10.13026/C2WM1R.