透析低血壓(Intradialytic hypotension, IDH)是透析最常見的併發症，盛行率甚至可以高達40%，且可能導致心肌梗塞、中風、猝死，是透析病患需面對的急迫問題。我們在研究中透過機器學習建立IDH的預測系統，利用透析機台各項偵測器與時序性的生命徵象數據時序地預測下個小時的IDH發生，讓醫護團隊有機會事前介入調整洗腎設定，減少IDH的嚴重程度也利於透析毒素清除。
現今醫療模式受到醫學研究的發展及檢測儀器進步的影響，逐漸走向依照個體的差異去研擬不同治療計畫的方向，也就是「個人化醫療 」。然而，現行的時序性模型僅以量測到病患的身體數據做為判斷，忽略了不同病患間個體的差異。這些差異不計其數，若全部蒐集齊全並整理將耗費大量的人力以及時間，冗長且繁複的蒐集程序亦會導致資料的數量驟減，因此我們在研究中透過讓模型學習病患過去資料的時序性特徵來辨別不同的病患，進而達到個人化預測的目的。
本研究提出個人化模組，利用預訓練模型抓取該位病患過去透析資料中曾發生透析中低血壓的時序性特徵，與該位病患當前透析資料做比對，計算相似度後做為主要模型的輸入，改變時序性資料的權重，目的是讓主要模型透過個人化模組辨別相同數據對於不同病患代表的意義也會不同。主要模型則是分為三個部分，時間感知注意力主要負責處理時序性資料，特徵選擇網路負責處理非時序性資料，最後再將室內外溫度做為隱藏態的權重，三個模組同時進入時序性模型做最後的預測。
本研究設計了三個實驗用以評估本研究所提出框架。在成大醫院提供的資料集中，我們使用三個國際公認評斷有無發生透析中低血壓的標準，分別為NADIR-90 (IDH-1)、Fall-40 (IDH-2)以及Fall-20 (IDH-3)，本研究提出的框架在IDH-2及IDH-3上表現優於其他比較方法，IDH-1則與表現最好的模型接近。本研究除了與其他深度學習的方法做比較以外，我們邀請五位腎臟科醫師針對IDH-1及IDH-2進行評分測驗，本研究提出的框架在表現上亦具有優勢，足以做為輔助醫師的工具。最後是使用成大醫院斗六分院的資料進行外部驗證，以證實模型的強健性。在本研究的三個實驗中，本研究提出的模型表現在多個國際公認的標準上皆呈現優異的表現、實用性以及強健性。

Intradialytic hypotension (IDH) is the most common complication encountered during hemodialysis. If we can utilize time-series data from hemodialysis machine to build a deep learning-based system for sequentially predicting IDH in the upcoming hour, the doctors would have the opportunity to intervene and adjust dialysis settings before the onset of IDH.
The current medical approach is shifting towards personalized medicine, which tailors treatment plans to the unique characteristics of each individual. Historically, time-series models only relied on the measured physiological data of patients for prediction, ignoring the individual differences among patients. Therefore, our aim is to utilize the temporal features of the patient's past medical records to enable the model to differentiate between different patients, thus achieving the goal of personalized prediction.
This thesis proposes a personalized module that uses a model trained with pretext task to capture the temporal features of past dialysis data in which IDH occurred for a specific patient. The personalized module compares the current dialysis data of the patient with the captured features and calculates the similarity as the input for the main model to adjust the weights of the temporal data. The main model is designed to distinguish the different meanings of the same data for different patients through the personalized module. The primary model consists of three modules: the time-aware attention module handles the temporal data, the feature selection network handles the non-temporal data, and the indoor and outdoor temperature are used as weights for the hidden state. These three modules are integrated into the time-series model for the final prediction.
This study designed three experiments to evaluate the proposed framework. In the dataset provided by National Cheng Kung University Hospital (NCKUH), we evaluated the proposed framework using three internationally recognized criteria, specifically NADIR-90 (IDH-1), Fall-40 (IDH-2), and Fall-20 (IDH-3) for assessing the occurrence of hypotension during dialysis. Our framework outperformed other methods for comparison on IDH-2 and IDH-3 and showed comparable performance on IDH-1. In addition to comparing with other deep learning methods, this study also invited digital nephrology specialists to conduct rating tests on IDH-1 and IDH-2. The proposed framework also showed advantages in performance, which is sufficient to serve as a tool to assist physicians. Finally, external validation was conducted using the dataset from Douliu Branch of NCKUH to verify the robustness. In the three experiments conducted in this study, the proposed model demonstrated excellent performance, practicality, and robustness in multiple internationally recognized standards.

Reference
1. Choi, E., et al., Retain: An interpretable predictive model for healthcare using reverse time attention mechanism. Advances in neural information processing systems, 2016. 29.
2. Luo, J., et al. Hitanet: Hierarchical time-aware attention networks for risk prediction on electronic health records. in Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. 2020.
3. Ye, M., et al. Lsan: Modeling long-term dependencies and short-term correlations with hierarchical attention for risk prediction. in Proceedings of the 29th ACM International Conference on Information & Knowledge Management. 2020.
4. Tan, Q., et al. Data-gru: Dual-attention time-aware gated recurrent unit for irregular multivariate time series. in Proceedings of the AAAI Conference on Artificial Intelligence. 2020.
5. Ren, H., et al. Rapt: Pre-training of time-aware transformer for learning robust healthcare representation. in Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining. 2021.
6. Zhang, Y. and J. Yan. Crossformer: Transformer utilizing cross-dimension dependency for multivariate time series forecasting. in The Eleventh International Conference on Learning Representations. 2022.
7. Du, W., D. Côté, and Y. Liu, Saits: Self-attention-based imputation for time series. Expert Systems with Applications, 2023. 219: p. 119619.
8. Zerveas, G., et al. A transformer-based framework for multivariate time series representation learning. in Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining. 2021.
9. Cho, K., et al., Learning phrase representations using RNN encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078, 2014.
10. Vaswani, A., et al., Attention is all you need. Advances in neural information processing systems, 2017. 30.
11. Halonen, J.I., et al., Relationship between outdoor temperature and blood pressure. Occupational and environmental medicine, 2011. 68(4): p. 296-301.
12. Lim, B., et al., Temporal fusion transformers for interpretable multi-horizon time series forecasting. International Journal of Forecasting, 2021. 37(4): p. 1748-1764.
13. Assimon, M.M. and J.E. Flythe. Definitions of intradialytic hypotension. in Seminars in dialysis. 2017. Wiley Online Library.
14. Chen, T. and C. Guestrin. Xgboost: A scalable tree boosting system. in Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining. 2016.
15. Suykens, J.A. and J. Vandewalle, Least squares support vector machine classifiers. Neural processing letters, 1999. 9: p. 293-300.
16. Liaw, A. and M. Wiener, Classification and regression by randomForest. R news, 2002. 2(3): p. 18-22.
17. Paszke, A., et al., Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 2019. 32.
18. Kingma, D.P. and J. Ba, Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.
19. Pedregosa, F., et al., Scikit-learn: Machine learning in Python. the Journal of machine Learning research, 2011. 12: p. 2825-2830.
20. LaValle, S.M., M.S. Branicky, and S.R. Lindemann, On the relationship between classical grid search and probabilistic roadmaps. The International Journal of Robotics Research, 2004. 23(7-8): p. 673-692.
21. Lee, H., et al., Deep learning model for real-time prediction of intradialytic hypotension. Clinical Journal of the American Society of Nephrology, 2021. 16(3): p. 396-406.