本研究應用機器學習(ML)與深度學習(DL)模型來預測患者在麻醉誘導後25分鐘內是否發生誘導後低血壓(Post-induction Hypotension, PIH)，我們比較了這些模型在評估指標和解釋性上的差異，找出了影響PIH的重要因子，並使用SHAP深入解釋特徵與PIH之間的關係。此外，本研究延伸了過去文獻中的藥物劑量建議，為麻醉師提供有用的決策資訊，並找出何種手術適合使用泛化模型或專屬模型。資料集由K醫院提供。實驗結果顯示，機器學習模型整體上比深度學習模型表現更佳，其中LightGBM模型表現最佳，AUC指標為75.2%。模型的評估指標也證明了時序特徵之測量時間間隔更長的可行性，以及深度模型預測PIH的可行性。

This study applies machine learning (ML) and deep learning (DL) models to predict whether patients will experience post-induction hypotension (PIH) within 25 minutes after anesthesia induction. We compared these models in terms of evaluation metrics and interpretability, identifying key factors influencing PIH, and used SHAP to deeply explain the relationship between features and PIH. Additionally, this study extends the drug dosage recommendations from previous literature to provide useful decision-making information for anesthesiologists and identifies which types of surgeries are suitable for using generalized models or dedicated models. The dataset was provided by K Hospital. Experimental results show that ML models overall perform better than DL models, with the LightGBM model performing best, achieving an AUC score of 75.2%. The evaluation metrics of the models also demonstrate the feasibility of longer measurement intervals for time-series features, as well as the feasibility of using deep models to predict PIH.

論文審定書 i
摘要 ii
Abstract iii
目錄 iv
圖次 vii
表次 x
第一章 緒論 1
1.1 研究背景 1
1.2 研究動機 2
1.3 研究目的 3
第二章 文獻探討 4
2.1 PIH的預測因子研究 4
2.2 深度時序模型介紹 5
2.2.1 長短期記憶(Long Short-Term Memory, LSTM) 5
2.2.2 GRU(Gate Reccurent Unit): 5
2.2.3 卷積神經網路(Convolutional Neural Network,CNN) 6
2.2.4 一維卷積神經網路(1D Convolutional Neural Network, 1D CNN): 6
2.2.5 時間卷積網路(Temporal Convolution Network, TCN) 7
2.3 預測血壓的深度模型 7
2.4 預測PIH事件之模型 8
2.5 基於ML和DL模型的劑量建議 10
第三章 研究方法 13
3.1 資料集建立 14
3.1.1 資料收集和預處理 14
3.1.2 低血壓事件標註 16
3.1.3 訓練、測試集建立流程 17
3.2 資料集中不平衡的處理 17
3.3 模型演算法與架構 18
3.3.1 ML模型演算法 18
3.3.2 DL模型架構 19
3.4 特徵選取 21
3.5 模型解釋性 22
3.6 Simulation 23
3.7 手術專屬模型 24
第四章 實驗結果 25
4.1 評估指標 25
4.2 處理不平衡的策略 26
4.3 升壓劑的影響 28
4.4 模型的評估指標 31
4.4.1 ML模型評估指標 31
4.4.2 DL模型評估指標 32
4.4.3 ML和DL模型的評估指標差異 35
4.5 ML模型的重要特徵 35
4.6 模型的特徵選取 38
4.6.1 實驗一 38
4.6.2 實驗二 39
4.6.3 特徵選取結論 42
論文審定書 i
摘要 ii
Abstract iii
目錄 iv
圖次 vii
表次 x
第一章 緒論 1
1.1 研究背景 1
1.2 研究動機 2
1.3 研究目的 3
第二章 文獻探討 4
2.1 PIH的預測因子研究 4
2.2 深度時序模型介紹 5
2.2.1 長短期記憶(Long Short-Term Memory, LSTM) 5
2.2.2 GRU(Gate Reccurent Unit): 5
2.2.3 卷積神經網路(Convolutional Neural Network,CNN) 6
2.2.4 一維卷積神經網路(1D Convolutional Neural Network, 1D CNN): 6
2.2.5 時間卷積網路(Temporal Convolution Network, TCN) 7
2.3 預測血壓的深度模型 7
2.4 預測PIH事件之模型 8
2.5 基於ML和DL模型的劑量建議 10
第三章 研究方法 13
3.1 資料集建立 14
3.1.1 資料收集和預處理 14
3.1.2 低血壓事件標註 16
3.1.3 訓練、測試集建立流程 17
3.2 資料集中不平衡的處理 17
3.3 模型演算法與架構 18
3.3.1 ML模型演算法 18
3.3.2 DL模型架構 19
3.4 特徵選取 21
3.5 模型解釋性 22
3.6 Simulation 23
3.7 手術專屬模型 24
第四章 實驗結果 25
4.1 評估指標 25
4.2 處理不平衡的策略 26
4.3 升壓劑的影響 28
4.4 模型的評估指標 31
4.4.1 ML模型評估指標 31
4.4.2 DL模型評估指標 32
4.4.3 ML和DL模型的評估指標差異 35
4.5 ML模型的重要特徵 35
4.6 模型的特徵選取 38
4.6.1 實驗一 38
4.6.2 實驗二 39
4.6.3 特徵選取結論 42
4.7 模型的可解釋性 43
4.7.1 最佳ML模型 43
4.7.2 最佳DL模型 47
4.8 藥物劑量的Simulation 50
4.9 手術專屬模型 53
4.9.1 手術一、治療性導管植入術 54
4.9.2 手術二、深部複雜創傷處理 57
4.9.3 手術三、部分乳房切除術 67
4.9.4 手術專屬模型結論 70
第五章 結論 71
參考文獻 73
附錄 76

[1] J.-L. Vincent et al., "Mean arterial pressure and mortality in patients with distributive shock: a retrospective analysis of the MIMIC-III database," Annals of intensive care, vol. 8, pp. 1-10, 2018.
[2] K. Maheshwari et al., "The relationship between ICU hypotension and in-hospital mortality and morbidity in septic patients," Intensive care medicine, vol. 44, pp. 857-867, 2018.
[3] V. Salmasi et al., "Relationship between intraoperative hypotension, defined by either reduction from baseline or absolute thresholds, and acute kidney and myocardial injury after noncardiac surgery: a retrospective cohort analysis," Anesthesiology, vol. 126, no. 1, pp. 47-65, 2017.
[4] S. Südfeld et al., "Post-induction hypotension and early intraoperative hypotension associated with general anaesthesia," BJA: British Journal of Anaesthesia, vol. 119, no. 1, pp. 57-64, 2017.
[5] S. Zhao, "Prediction of Post-induction Hypotension by Machine Learning," 2023.
[6] S. Czajka, Z. Putowski, and Ł. J. Krzych, "Post-induction hypotension and intraoperative hypotension as potential separate risk factors for the adverse outcome: a cohort study," Journal of Anesthesia, pp. 1-9, 2023.
[7] A. R. Kang et al., "Development of a prediction model for hypotension after induction of anesthesia using machine learning," PloS one, vol. 15, no. 4, p. e0231172, 2020.
[8] D. L. Reich et al., "Predictors of hypotension after induction of general anesthesia," Anesthesia & Analgesia, vol. 101, no. 3, pp. 622-628, 2005.
[9] B. Chen, Q.-Y. Pang, R. An, and H.-L. Liu, "A systematic review of risk factors for postinduction hypotension in surgical patients undergoing general anesthesia," European Review for Medical & Pharmacological Sciences, vol. 25, no. 22, 2021.
[10] O. Jor et al., "Hypotension after induction of general anesthesia: occurrence, risk factors, and therapy. A prospective multicentre observational study," Journal of anesthesia, vol. 32, pp. 673-680, 2018.
[11] Y. Aissaoui et al., "Prediction of post-induction hypotension by point-of-care echocardiography: a prospective observational study," Anaesthesia Critical Care & Pain Medicine, vol. 41, no. 4, p. 101090, 2022.
[12] S. Abdullah Mohamed, M. Yousry Helmy, S. A. Khattab, A. M. Hossam, and M. S. Arafa, "Perfusion index as a predictor of hypotension after induction of general anesthesia in elderly patients–a prospective observational study," Egyptian Journal of Anaesthesia, vol. 39, no. 1, pp. 619-625, 2023.
[13] S. Aktas Yildirim, Z. T. Sarikaya, L. Dogan, H. Ulugol, B. Gucyetmez, and F. Toraman, "Arterial Elastance: A Predictor of Hypotension Due to Anesthesia Induction," J Clin Med, vol. 12, no. 9, Apr 27 2023, doi: 10.3390/jcm12093155.
[14] S. Hochreiter and J. Schmidhuber, "Long short-term memory," Neural computation, vol. 9, no. 8, pp. 1735-1780, 1997.
[15] J. Chung, C. Gulcehre, K. Cho, and Y. Bengio, "Empirical evaluation of gated recurrent neural networks on sequence modeling," arXiv preprint arXiv:1412.3555, 2014.
[16] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, "Gradient-based learning applied to document recognition," Proceedings of the IEEE, vol. 86, no. 11, pp. 2278-2324, 1998.
[17] S. Bai, J. Z. Kolter, and V. Koltun, "An empirical evaluation of generic convolutional and recurrent networks for sequence modeling. arXiv 2018," arXiv preprint arXiv:1803.01271, vol. 2, 1803.
[18] H. Eom et al., "End-to-end deep learning architecture for continuous blood pressure estimation using attention mechanism," Sensors, vol. 20, no. 8, p. 2338, 2020.
[19] H. Mou and J. Yu, "CNN-LSTM prediction method for blood pressure based on pulse wave," Electronics, vol. 10, no. 14, p. 1664, 2021.
[20] K. Zhou, Z. Yin, Y. Peng, and Z. Zeng, "Methods for continuous blood pressure estimation using temporal convolutional neural networks and ensemble empirical mode decomposition," Electronics, vol. 11, no. 9, p. 1378, 2022.
[21] S. Kendale, P. Kulkarni, A. D. Rosenberg, and J. Wang, "Supervised Machine-learning Predictive Analytics for Prediction of Postinduction Hypotension," Anesthesiology, vol. 129, no. 4, pp. 675-688, 2018, doi: 10.1097/aln.0000000000002374.
[22] J. Lee et al., "Comparative analysis on machine learning and deep learning to predict post-induction hypotension," Sensors, vol. 20, no. 16, p. 4575, 2020.
[23] S. A. Yildirim, L. Dogan, Z. T. Sarikaya, H. Ulugol, B. Gucyetmez, and F. Toraman, "Hypotension after Anesthesia Induction: Target-Controlled Infusion Versus Manual Anesthesia Induction of Propofol," J Clin Med, vol. 12, no. 16, Aug 14 2023, doi: 10.3390/jcm12165280.
[24] N. Asai et al., "Proposal of anesthetic dose prediction model to avoid post-induction hypotension using electronic anesthesia records," in 2019 Twelfth International Conference on Mobile Computing and Ubiquitous Network (ICMU), 2019: IEEE, pp. 1-4.
[25] N. Miyaguchi, K. Takeuchi, H. Kashima, M. Morita, and H. Morimatsu, "Predicting anesthetic infusion events using machine learning," Sci Rep, vol. 11, no. 1, p. 23648, Dec 8 2021, doi: 10.1038/s41598-021-03112-2.
[26] R. B. Schonberger et al., "Association of propofol induction dose and severe pre-incision hypotension among surgical patients over age 65," Journal of Clinical Anesthesia, vol. 80, p. 110846, 2022.
[27] X. Cai, J. Chen, Y. Zhu, B. Wang, and Y. Yao, "Towards Safe Propofol Dosing during General Anesthesia Using Deep Offline Reinforcement Learning," arXiv preprint arXiv:2303.10180, 2023.
[28] K. Maheshwari et al., "The association of hypotension during non‐cardiac surgery, before and after skin incision, with postoperative acute kidney injury: a retrospective cohort analysis," Anaesthesia, vol. 73, no. 10, pp. 1223-1228, 2018.
[29] X. Huang, A. Khetan, M. Cvitkovic, and Z. Karnin, "Tabtransformer: Tabular data modeling using contextual embeddings," arXiv preprint arXiv:2012.06678, 2020.
[30] K. Qin, W. Huang, and T. Zhang, "Deep generative model with domain adversarial training for predicting arterial blood pressure waveform from photoplethysmogram signal," Biomedical Signal Processing and Control, vol. 70, p. 102972, 2021.
[31] H. Eom et al., "End-to-End Deep Learning Architecture for Continuous Blood Pressure Estimation Using Attention Mechanism," Sensors (Basel), vol. 20, no. 8, Apr 20 2020, doi: 10.3390/s20082338.
[32] A. T. Garmendia et al., "Towards personalised early prediction of Intra-Operative Hypotension following anesthesia using Deep Learning and phenotypic heterogeneity," medRxiv, p. 2023.01. 20.23284432, 2023.