本研究延續先前研究成果，旨在解決傳統骨盆底肌肉訓練中，因受試者疲勞導致的施力變化無法被現有評分模型正確評估的問題。先前之模型雖可辨識理想波形與連續波形，但面對受試者肌力隨時間變化、非動作錯誤導致的波形差異，無法進行個人化評估。
本研究透過 LSTM 模型預測受試者未來波形的波峰與左右波谷，並根據預測結果動態調整原有評分模型之映射區間，使評分能更貼近受試者真實復健狀況。具體方法為將波形切割後提取波峰與波谷特徵，並利用三組 LSTM 模型分別預測下一次波峰、左波谷與右波谷的位置，再以預測結果與當前評分模型進行映射修正。
研究實作部分選定 6 號受試者為主要測試對象，其具備充足數量的規律波形資料，並呈現下降趨勢，適合進行預測與調整。模型訓練與測試資料比例為 7:3，並使用 RMSE 指標驗證預測精度。
驗證結果顯示，調整後的評分模型能有效降低原模型因固定區間而產生的誤判情形，提升對實際施力狀況的評分準確性。混淆矩陣結果亦顯示調整模型具備良好的一致性與準確度。
未來可根據使用者錯誤波形出現的規律，或是非線性下降運動循環的時序特性，進行運動資料的時間段切割，將不同區段的動作獨立建模，提升預測準確度與評分辨識能力。

Pelvic floor muscle training (PFMT) is a widely recommended non-invasive intervention for urinary incontinence. However, conventional evaluation models often fail to reflect a patient's true rehabilitation progress, especially when muscle strength fluctuates due to fatigue rather than improper posture. This study extends previous research efforts to address the limitations of traditional PFMT evaluation models, particularly in cases where fatigue, rather than incorrect movements, causes variability in exertion.

This study proposes a dynamic evaluation system using Long Short-Term Memory (LSTM) networks to improve the accuracy of home-based PFMT scoring. After segmenting the waveforms and extracting peaks and valleys from pressure signals, three LSTM models predict the next peak and the surrounding valleys. These predictions are used to adjust the threshold intervals of the existing scoring model, making the scores more reflective of actual rehabilitation status.
In the implementation, participant 6 was selected as the main test subject due to a large amount of consistent waveform data and a downward trend suitable for prediction and adjustment. The training and testing data were split in a 7:3 ratio, and RMSE was used to evaluate prediction accuracy.

Results showed that the adjusted scoring model effectively reduced misjudgments caused by the fixed interval mechanism in the original model and improved the accuracy of force evaluations. The confusion matrix results also demonstrated enhanced consistency and accuracy of the adjusted model.

To further enhance prediction accuracy and scoring reliability, future work may explore segmenting exercise data based on patterns in erroneous waveforms or nonlinear temporal decline trends, allowing for separate modeling of different movement segments.

目錄
第一章 緒論	1
第二章 文獻探討	3
第一節 骨盆底肌肉訓練評分模型	3
第二節 運動分群	4
第三節 時間序列模型與LSTM應用	4
第三章 研究方法	5
第一節 資料收集	6
第二節 資料預處理	9
壹、 規律波形分類	9
貳、 運動類型分類	10
參、 建立資料集	13
第三節 預測波形位置	14
第四節 分析評分模型	15
第四章 實作驗證	19
第一節 資料收集	19
第二節 資料預處理	20
第三節 分析評分模型	22
第五章 結論與未來展望	25
參考文獻		26
附錄一 IRB同意臨床試驗 / 研究證明書 / 研究持續審查案同意證明	28
附錄二 受試者同意書	31
附錄三 學術倫理修課證明 （中文）	43
附錄四 學術倫理修課證明 （英文）	45
附錄五 骨盆腔器官脫垂-日常生活影響問卷調查表 (POPDI-6)	47
附錄六 尿失禁症狀困擾量表 (UDI-6)	47
附錄七 尿失禁症狀影響量表 (IIO-7)	48
附錄八 部分脫垂與性生活功能評估 (PISQ-IR)	48
 
圖目錄
圖2.1 骨盆底肌肉運動儀器歷屆版本設計	3
圖3.1 骨盆底肌肉運動復健系統的應用情境流程圖	5
圖3.2 研究流程圖	6
圖3.3 運動壓力數據	8
圖3.4 骨盆底肌肉收縮波形圖	8
圖3.5 規律運動波形與無規律運動波形	9
圖3.6 運動分類方式示意圖	10
圖3.7 運動分類及其特徵	11
圖3.8 以三個特徵進行K-MEANS運動分類	11
圖3.9 波形資料分為頂點、左側、右側資料集	13
圖3.10 資料集切分示意圖	14
圖3.11 LSTM模型訓練架構	15
圖3.12 評分模型區域表示圖	16
圖3.13 評分區域映射流程圖	16
圖4.1 規律波形數量圓餅圖	20
圖4.2 各受試者規律運動資料統計	21
圖4.3 K-MEANS分群結果	21
圖4.4 波形預測模型預測結果	22
 
表目錄
表3.1 混淆矩陣	17
表4.1 原始評分模型混淆矩陣	23
表4.2 動態調整後評分模型混淆矩陣	23
表4.3 原始模型與動態調整後分數判斷比較	24

[1]	Q. Wang, Y.-Z. Que, X.-Y. Wan, and C.-Q. Lin, “Prevalence, Risk Factors, and Impact on Life of Female Urinary Incontinence: An Epidemiological Survey of 9584 Women in a Region of Southeastern China,” Risk Manag. Healthc. Policy, vol. 16, pp. 1477–1487, Aug. 2023, doi: 10.2147/RMHP.S421488.
[2]	Jensen LCG, Axelsen S. Urinary incontinence and work capacity. Dan Med J. 2023 Sep 25;70(10):A04230271. PMID: 37897392.
[3]	P.-M. Chow, Y.-C. Chuang, K. C. P. Hsu, Y.-C. Shen, and S.-P. Liu, “Impact of Female Stress Urinary Incontinence on Quality of Life, Mental Health, Work Limitation, and Healthcare Seeking in China, Taiwan, and South Korea (LUTS Asia): Results from a Cross-Sectional, Population-Based Study,” Int. J. Womens Health, vol. 14, pp. 1871–1880, Dec. 2022, doi: 10.2147/IJWH.S383651.
[4]	王俞君（2025）。AI卷積模型在尿失禁患者骨盆底肌訓練中的即時評估應用〔碩士論文，淡江大學〕。華藝線上圖書館。https://doi.org/10.6846/tku202500010
[5]	Constantin-Teodosiu D, Constantin D. Molecular Mechanisms of Muscle Fatigue. International Journal of Molecular Sciences. 2021; 22(21):11587. https://doi.org/10.3390/ijms222111587
[6]	Chen J-F, Horng H-C, Lin W-C, Tsai K-H. Noninvasive Wireless Sensor PFMT Device for Pelvic Floor Muscle Training. International Journal of Distributed Sensor Networks. 2012;8(5). doi:10.1155/2012/658724
[7]	Horng, Huann-Chenga,b,c; Chao, Wei-Tingb,c,d,e; Chen, Jui-Faf; Chang, Chia-Peib,c; Wang, Peng-Huib,c,d,e; Chang, Po-Luna,*. Home-based noninvasive pelvic floor muscle training device to assist women in performing Kegel exercise in the management of stress urinary incontinence. Journal of the Chinese Medical Association 85(4):p 484-490, April 2022. | DOI: 10.1097/JCMA.0000000000000660
[8]	F. F. Fitz, L. M. Paladini, L. de A. Ferreira, M. M. Gimenez, M. A. T. Bortolini, and R. A. Castro, “Ability to contract the pelvic floor muscles and association with muscle function in incontinent women,” Int. Urogynecology J., vol. 31, no. 11, pp. 2337–2344, Nov. 2020, doi: 10.1007/s00192-020-04469-0.
[9]	P. Chiarelli, B. Murphy, and J. Cockburn, “Women’s knowledge, practises, and intentions regarding correct pelvic floor exercises,” Neurourol. Urodyn., vol. 22, no. 3, pp. 246–249, 2003, doi: 10.1002/nau.10119.
[10]	A. Farooq, M. Seyedmahmoudian and A. Stojcevski, "A Wearable Wireless Sensor System Using Machine Learning Classification to Detect Arrhythmia," in IEEE Sensors Journal, vol. 21, no. 9, pp. 11109-11116, 1 May1, 2021, doi: 10.1109/JSEN.2021.3062395.
[11]	Liu, K.; Liu, Y.; Ji, S.; Gao, C.; Fu, J. Estimation of Muscle Forces of Lower Limbs Based on CNN–LSTM Neural Network and Wearable Sensor System. Sensors 2024, 24, 1032. https://doi.org/10.3390/s24031032
[12]	C. -C. Hsiao, P. -C. Yu, R. -G. Lee and H. Jiang, "Deep-Learning LSTM Mechanism and Wearable Devices based Virtual Fitness-Coach Information System for Barbell Bench Press," 2020 IEEE International Conference on Systems, Man, and Cybernetics (SMC), Toronto, ON, Canada, 2020, pp. 3880-3885, doi: 10.1109/SMC42975.2020.9283422.