血液透析治療中有許多併發症，其中包含透析中低血壓和靜脈針脫落漏血。透析中低血壓 (IDH)的發生率高達25%。醫護人員每30-60分鐘才能量測血壓值，無法有效預防患者出現透析中低血壓。血液透析 (HD)期間發生的靜脈針脫位 (VND) 也是醫療保健系統中重要的全球性問題，隨著人工智能 (AI)和物聯網 (IoT)技術的進步和成熟，已多項研究導入臨床決策與照護。本研究的主要目標為運用AIoT技術來建立透析中低血壓預測和漏液偵測的多床監控預警系統，並驗證對於併發症減少之臨床效益。
關於透析中低血壓預測模型，本研究使用來自台南市安南醫院門診洗腎室相關資料，收集穿戴裝置、血壓相關及透析機的連續數據以及來自病歷資料分析之離散數據進行訓練模型。模型包含二元分類與數值預測，使用具有時間序列差值、智慧穿戴裝置連續量測心率與血氧整合血液透析資訊共116項特徵，運用Random forest、LightGBM、XGBoost、LSTM、GRU、CNN+LSTM和CNN + GRU等機器學習與深度學習模型進行預測。藉由多項特徵及RFECV特徵等進行模型優化之比較，對於二元分類模型與對於數值預測模型成效評估。
漏血偵測系統的方法設計上，漏血檢測裝置包括處理單元、漏液貼片、藍牙模塊和報警單元。我們進行假手臂模擬實驗來評估漏液偵測貼片的性能，其中測試了不同體積、流速與方向等漏血方式。在臨床試驗設計方面，於台南市立安南醫院共招募71位透析病患，統計其漏液監控系統偵測之準確度、靈敏度及特異度等來判定漏液偵測之可行性。
此研究前兩個部份整合成一AIoT多床監控平台後，運用 Acusense AIoT平台。並經過實際上線醫護人員使用後，並統計來驗證其低血壓與漏血併發症下降比率。
在這項透析中低血壓預測模型研究中，在二元分類方面，通過XGBoost 的特徵提取，由LSTM進行預測。結果顯示AUC為 0.966。此外運用數值預測方面，運用時間序列差值特徵，最佳的機器學習法Xgboost之R2為0.609。在漏血偵測系統臨床試驗中，在不同漏液方向上皆能進行感測。將此血液感測裝置實際運行於臨床上，結果顯示靈敏度為 90.9%，準確度為 99.7%，精確度為100%。此外，該臨床研究顯示出良好的效能，沒有明顯的副作用。
當整合成一AIoT多床監控平台後，在洗腎低血壓預測系統的臨床試驗中，建立透析低血壓多床監測預警系統。醫護人員使用AIoT系統八個月後，低血壓的發生率顯著下降。透析低血壓發生率從最初的11% 下降到 6.48%，其代表總體發生率下降比率達到41%。在導入漏液監控系統之後，重度和中度漏液數量大幅減少了72.7%。而整體的漏血率為3.3%，採用標準化臨床程序和漏血監控系統後，漏液發生率率為2.1%，整體漏血發生率下降36.4%。
本論文針對兩種主要的透析中併發症事件，運用AIoT整合系統來驗證達到併發症下降之效益，未來可增加其他適應症。 因此，我們相信本整合系統具有巨大的臨床潛力，能降低透析患者併發症的發生率，並且寧從醫院內擴增到居家透析使用。

Introduction
There are many complications in dialysis, such as intradialytic hypotension and venous needle dislodgement. The incidence of intradialytic hypotension (IDH) is as high as 25%. It is difficult to prevent patients from suffering intradialytic hypotension by checking their vital signs every 30-60 minutes. Intradialytic hypotension causes uncomfortableness in the whole body and various vascular complications. Venous needle dislodgement (VND), which occurs during hemodialysis (HD), is also a significant global issue in the healthcare system. As the technology of artificial intelligence (AI) and the internet of things (IoT) advances and becomes mature, it can support clinical decisions for preventing intradialytic hypotension and venous needle dislodgement. The main goal of this study is to use AIoT technology to establish multi-bed monitoring and early warning system for both hypotension prediction and leakage detection during dialysis. Furthermore, to verify the clinical benefits of reducing complication rates.

Method
In order to construct the intradialytic hypotension prediction model, data was collected from an AIoT platform at the Tainan Municipal Annan Hospital outpatient clinic. Using smart wearable devices to measure heart rate and blood oxygen continuously and integrating with blood pressure and hemodialysis parameters to deduce their time series differences with a total of 116 features as inputs to train the models, including binary classification and numerical prediction. Machine learning models including random forest, XGBoost, LightGBM, LSTM, GRU, CNN+LSTM and CNN+GRU were chosen for comparison. We compared the performance of these models and also the effects of different numbers of features through RFECV.
For the blood leakage detection system, the complete device consisted of a processing unit, a leakage-detection patch, a Bluetooth module and an alarm unit. We conducted a phantom arm simulation experiment to evaluate the performance of the leakage detection patch, where different volumes of blood were tested. For clinical trial design, 71 hemodialysis patients from Tainan Municipal An-Nan Hospital in Tainan City, Taiwan, were chosen as subject candidates. The accuracy, sensitivity and specificity of the leak detection system were measured to determine the feasibility of leak detection.
For the AIoT multi-bed monitoring platform, this research used the Acusense AIoT platform, a real-time multi-bed monitoring system. After the real on line use by the medical staff, statistical analysis was performed to study the reduction rate of hypotension and blood leakage complications.

Results
For intradialytic hypotension prediction, the binary classification model, through XGboost feature extraction method to using LSTM prediction, resulted in an AUC of 0.966. Furthermore, for the numerical prediction, the results of LightGBM and Xgboost models with time series features showed an R-squared of 0.609.
For the blood leakage detection system, the results showed 90.9% sensitivity, 99.7% accuracy and 100% precision. In addition, the real clinical study showed good performance with no significant side effects. For the AIoT multi-bed monitoring platform, after eight months of using the AIoT system by medical staff, the incidence of intradialytic hypotension significantly decreased. The incidence[到底是prevalence還是incidence] [指的是新發的併發症人數為分母，故為incidence
]decreased from the initial 11% to 6.48%, representing a 41% reduction in the overall incidence rate.
For the VND detection part, the number of severe and moderate leakage has drastically decreased by 72.7%. In addition, the rate of blood leakage under the initial care method and after adopting the standardized procedure and implementing the blood detection system was 3.3% and 2.1%, representing a 36.4% drop in the overall incidence of blood leakage.

Conclusion
This thesis addressed two significant complications in dialysis: intradialytic hypotension and blood leakage. The preliminary results of using the AIoT system to reduce complications showed its feasibility. Furthermore, the AIoT system can be expanded to include more items in the future. Therefore, we believe that the system has a great potential for clinical applications to reduce the complications during dialysis and can further be extended from in-hospital to in-home use.

[1] National Academies of Sciences, Engineering, and Medicine, Health and Medicine Division, Board on Health Care Services, Board on Global Health, and Committee on Improving the Quality of Health Care Globally, Crossing the Global Quality Chasm: Improving Health Care Worldwide. Washington (DC): National Academies Press (US), 2018.
[2] H. Wang et al., “Global, regional, and national life expectancy, all-cause mortality, and cause-specific mortality for 249 causes of death, 1980–2015: a systematic analysis for the Global Burden of Disease Study 2015,” The Lancet, vol. 388, no. 10053, pp. 1459–1544, Oct. 2016.
[3] C. German Commercial, S. E. Fresenius, and C. KGaA, “Fresenius Annual Report 2016.” https://www.fresenius.com/sites/default/files/2022-01/Fresenius_GB_US_GAAP_2016_englisch_0.pdf.
[4] A. Shander, A. Hofmann, S. Ozawa, O. M. Theusinger, H. Gombotz, and D. R. Spahn, “Activity-based costs of blood transfusions in surgical patients at four hospitals,” Transfusion, vol. 50, no. 4, pp. 753–65, 2010.
[5] J. Kuipers et al., “Prevalence of intradialytic hypotension, clinical symptoms and nursing interventions - a three-months, prospective study of 3818 haemodialysis sessions,” BMC Nephrology, vol. 17, no. 1, Feb. 2016.
[6] L. Gabutti, M. Machacek, C. Marone, and P. Ferrari, “Predicting intradialytic hypotension from experience, statistical models and artificial neural networks,” Journal of Nephrology, vol. 18, no. 4, pp. 409–416, Jul. 2005.
[7] L. Coli et al., “Evidence of profiled hemodialysis efficacy in the treatment of intradialytic hypotension,” The International Journal of Artificial Organs, vol. 21, no. 7, pp. 398–402, Jul. 1998, Accessed: Aug. 23, 2022.
[8] J. E. Flythe, H. Xue, K. E. Lynch, G. C. Curhan, and S. M. Brunelli, “Association of Mortality Risk with Various Definitions of Intradialytic Hypotension,” Journal of the American Society of Nephrology, vol. 26, no. 3, pp. 724–734, Sep. 2014.
[9] J. J. Sands et al., “Intradialytic hypotension: Frequency, sources of variation and correlation with clinical outcome,” Hemodialysis International, vol. 18, no. 2, pp. 415–422, Jan. 2014.
[10] V. Abouie, H. Sharifian, F. Towhidkhah, V. Nafisi, and H. Abouei, “Using neural network in order to predict hypotension of hemodialysis patients,” IEEE Xplore, May 01, 2011.
[11] C.-J. Lin et al., “Intelligent system to predict intradialytic hypotension in chronic hemodialysis,” Journal of the Formosan Medical Association, vol. 117, no. 10, pp. 888–893, Oct. 2018.
[12] H. Lee et al., “Deep Learning Model for Real-Time Prediction of Intradialytic Hypotension,” Clinical Journal of the American Society of Nephrology, vol. 16, no. 3, pp. 396–406, Feb. 2021.
[13] B. Axley, J. Speranza-Reid, and H. Williams, “Venous needle dislodgement in patients on hemodialysis,” Nephrology Nursing Journal: Journal of the American Nephrology Nurses’ Association, vol. 39, no. 6, pp. 435–445; quiz 446, Nov. 2012.
[14] J.-P. Van Waeleghem, M. Chamney, E. J. Lindley, and J. Pancírova, “VENOUS NEEDLE DISLODGEMENT: HOW TO MINIMISE THE RISKS,” Journal of Renal Care, vol. 34, no. 4, pp. 163–168, Dec. 2008.
[15] A. Billie, S.-R. Joan, and W. Helen, “Venous Needle Dislodgement In Patients on Hemodialysis - ProQuest,” www.proquest.com, Nov. 2012.
[16] K. K. Tennankore, C. d’Gama, R. Faratro, S. Fung, E. Wong, and C. T. Chan, “Adverse Technical Events in Home Hemodialysis,” American Journal of Kidney Diseases, vol. 65, no. 1, pp. 116–121, Jan. 2015.
[17] J.-X. Wu, P.-T. Huang, C.-H. Lin, and C.-M. Li, “Blood leakage detection during dialysis therapy based on fog computing with array photocell sensors and heteroassociative memory model,” Healthcare Technology Letters, vol. 5, no. 1, pp. 38–44, Feb. 2018.
[18] “AAMI – Association for the Advancement of Medical Instrumentation,” Dialysis Water Solutions, Jun. 02, 2015. https://dialysiswatersolution.com/regulations-and-guidelines/ansiaami/
[19] Q. Ding and Q. Ye, “Needle dislodgement in hemodialysis patients: Progress in prevention and early intervention,” Hemodialysis International, vol. 25, no. 3, pp. 281–287, Jun. 2021.
[20] E. J. Lindley et al., “Renal Care Journal Club Discussion Summary,” EDTNA-ERCA Journal, vol. 31, no. 2, pp. 108–114, Apr. 2005.
[21] J. AHLMÉN, K.-H. GYDELL, H. HADIMERI, I. HERNANDEZ, B. ROGLAND, and U. STRÖMBOM, “A new safety device for hemodialysis,” Hemodialysis International, vol. 12, no. 2, pp. 264–267, Apr. 2008.
[22] J.-X. Wu, P.-T. Huang, C.-M. Li, and C.-H. Lin,“Bidirectional Hetero-Associative Memory Network With Flexible Sensors and Cloud Computing for Blood Leakage Detection in Intravenous and Dialysis Therapy,” IEEE Transactions on Emerging Topics in Computational Intelligence, vol. 2, no. 4, pp. 298–307, Aug. 2018.
[23] Y.-C. Du, B.-Y. Lim, W.-S. Ciou, and M.-J. Wu, “Novel Wearable Device for Blood Leakage Detection during Hemodialysis Using an Array Sensing Patch,” Sensors, vol. 16, no. 6, p. 849, Jun. 2016.
[24] Y.-K. Ou, M.-J. Wu, W.-S. Ciou, and Y.-C. Du, “A Clinical Trial of the Effect of a Blood Leakage Detection Device for Patients during Hemodialysis,” International Journal of Environmental Research and Public Health, vol. 16, no. 13, p. 2388, Jul. 2019.
[25] R. M. McAdams, R. Kaur, Y. Sun, H. Bindra, S. J. Cho, and H. Singh, “Predicting clinical outcomes using artificial intelligence and machine learning in neonatal intensive care units: a systematic review,” Journal of Perinatology, May 2022.
[26] J. Kuipers et al., “The Prevalence of Intradialytic Hypotension in Patients on Conventional Hemodialysis: A Systematic Review with Meta-Analysis,” American Journal of Nephrology, vol. 49, no. 6, pp. 497–506, 2019.
[27] S. Park et al., “Predicting intradialytic hypotension using heart rate variability,” Scientific Reports, vol. 9, no. 1, p. 2574, Feb. 2019.
[28] Y.-M. Chang et al., “Heart rate variability is an indicator for intradialytic hypotension among chronic hemodialysis patients,” Clinical and Experimental Nephrology, vol. 20, no. 4, pp. 650–659, Oct. 2015.
[29] K. V. S. S. Ganesh, S. P. S. Jeyanth, and A. R. Bevi, “IOT based portable heart rate and SpO2 pulse oximeter,” HardwareX, vol. 11, p. e00309, Apr. 2022.
[30] S. Wang, Y. Dai, J. Shen, and J. Xuan, “Research on expansion and classification of imbalanced data based on SMOTE algorithm,” Scientific Reports, vol. 11, no. 1, p. 24039, Dec. 2021.
[31] D. F. Keane, J. G. Raimann, H. Zhang, J. Willetts, S. Thijssen, and P. Kotanko, “The time of onset of intradialytic hypotension during a hemodialysis session associates with clinical parameters and mortality,” Kidney International, vol. 99, no. 6, pp. 1408–1417, Jun. 2021.
[32] P. Misra and A. S. Yadav, “Improving the Classification Accuracy using Recursive Feature Elimination with Cross-Validation,” International Journal on Emerging Technologies, vol. 11, no. 3, pp. 659–665, 2020.
[33] J. L. Speiser, M. E. Miller, J. Tooze, and E. Ip, “A comparison of random forest variable selection methods for classification prediction modeling,” Expert Systems with Applications, vol. 134, pp. 93–101, Nov. 2019.
[34] A. Ogunleye and Q.-G. Wang, “XGBoost Model for Chronic Kidney Disease Diagnosis,” IEEE/ACM Transactions on Computational Biology and Bioinformatics, vol. 17, no. 6, pp. 2131–2140, Nov. 2020.
[35] Li, Ping, Qiang Wu, and Christopher J. Burges. “Mcrank: Learning to rank using multiple classification and gradient boosting.” Advances in neural information processing systems. 2007.
[36] H. Shi, “Best-first Decision Tree Learning,” researchcommons.waikato.ac.nz, 2007.
[37] J. J. Bird, A. Ekárt, C. D. Buckingham, and D. R. Faria, “Evolutionary Optimisation of Fully Connected Artificial Neural Network Topology,”Advances in Intelligent Systems and Computing, pp. 751–762, 2019.
[38] W. Ge, J.-W. Huh, Y. R. Park, J.-H. Lee, Y.-H. Kim, and A. Turchin, “An Interpretable ICU Mortality Prediction Model Based on Logistic Regression and Recurrent Neural Networks with LSTM units,” AMIA Annual Symposium proceedings. AMIA Symposium, vol. 2018, pp. 460–469, 2018.
[39] S. Siami-Namini, N. Tavakoli, and A. S. Namin, “The Performance of LSTM and BiLSTM in Forecasting Time Series,” 2019 IEEE International Conference on Big Data (Big Data), Dec.2019.
[40] Y. Liu et al., “Bidirectional GRU networks‐based next POI category prediction for healthcare,” International Journal of Intelligent Systems, vol. 37, no. 7, pp. 4020–4040, Oct. 2021.
[41] S. A. Rahman and D. A. Adjeroh, “Deep Learning using Convolutional LSTM estimates Biological Age from Physical Activity,” Scientific Reports, vol. 9, no. 1, Aug. 2019.
[42] N. Barr Kumarakulasinghe, T. Blomberg, J. Liu, A. Saraiva Leao, and P. Papapetrou, “Evaluating Local Interpretable Model-Agnostic Explanations on Clinical Machine Learning Classification Models,” IEEE Xplore, Jul. 01, 2020.
[43] H.-W. Hu, C.-H. Liu, Y.-C. Du, K.-Y. Chen, H.-M. Lin, and C.-C. Lin, “Real-Time Internet of Medical Things System for Detecting Blood Leakage during Hemodialysis Using a Novel Multiple Concentric Ring Sensor,” Sensors, vol. 22, no. 5, p. 1988, Mar. 2022.
[44] D. Berrar, “Performance Measures for Binary Classification,” Encyclopedia of Bioinformatics and Computational Biology, pp. 546–560, 2019.
[45] J. Wu, Y. Lin, K. Lin, Y. Hu, and G. Kong, “Predicting length of stay in intensive care unit using ensemble learning methods,” Developments of Artificial Intelligence Technologies in Computation and Robotics, Aug. 2020.
[46] Y. Koda and I. Aoike, “Prevention of Intradialytic Hypotension with Intermittent Back-Filtrate Infusion Haemodiafiltration: Insights into the Underlying Mechanism,” Blood Purification, vol. 48, no. 1, pp. 1–6, 2019.
[47] M. G. W. Barnas, W. H. Boer, and H. A. Koomans, “Hemodynamic Patterns and Spectral Analysis of Heart Rate Variability during Dialysis Hypotension,” Journal of the American Society of Nephrology, vol. 10, no. 12, pp. 2577–2584, Dec. 1999.
[48] J.-B. Chen, K.-C. Wu, S.-H. Moi, L.-Y. Chuang, and C.-H. Yang, “Deep Learning for Intradialytic Hypotension Prediction in Hemodialysis Patients,” IEEE Access, vol. 8, pp. 82382–82390, 2020.
[49] M. L. Robbin et al., “Hemodialysis Arteriovenous Fistula Maturity: US Evaluation,” Radiology, vol. 225, no. 1, pp. 59–64, Oct. 2002.
[50] N. Liu et al., “Remote Monitoring Systems for Chronic Patients on Home Hemodialysis: Field Test of a Copresence-Enhanced Design,” JMIR Human Factors, vol. 4, no. 3, Aug. 2017.
[51] N. Liu et al., “Telehealth for Noncritical Patients With Chronic Diseases During the COVID-19 Pandemic,” Journal of Medical Internet Research, vol. 22, no. 8, p. e19493, Aug. 2020.