目的 : 為了預測質子治療頭頸癌患者發生放射性皮膚炎情形，本研究透過監督式機器學習方法建立放射性皮膚炎預後模型，並與傳統邏輯斯回歸比較。
材料與方法 : 本研究收集了57 位頭頸癌患者的回溯性資料，所有患者皆已完成強度調控質子治療 (IMPT)計畫的處方療程。使用監督式機器學習 (ML)演算法，包括隨機森林 (RF)、XGBoost 和CatBoost，對這些資料進行併發症預測建模。預測輸出將 ≥2 級的放射性皮膚炎定義為嚴重併發症。預後因子的候選特徵包括患者的臨床因子和治療計劃劑量因子。通過皮爾森相關性排除共線因子後，進行LASSO 正則化對特徵挑選，並建立邏輯斯迴歸 (LR)模型進行比較。此外，我們還使用機器學習結合沙普利加解釋 (SHAP)來挑選特徵作為LR 輸入進行建模，並比較了相同特徵輸入至邏輯斯、隨機森林、XGBoost、CatBoost 之效能差異。ML 算法部分則將資料集切分為訓練、驗證和測試集，並對訓練驗證集進行10折交叉驗證，以網格搜尋法調整超參數作為穩健優化方法。最後輸出各模型在測試集上的評估指標，包括準確性 (ACC)、受試者特性曲線下面積(AUC)、陽性預測值 (PPV)、陰性預測值 (NPV)以及F1 分數，作為預測模型的性能量化指標。
結果 : 四種特徵挑選方法之LR 模型中以LASSO 方法表現最佳，於測試集AUC 為0.87，因子挑選出吸菸史、N 分期、AJCC 分期、性別、D100_3mm、V60_5mm
六個特徵。而在ML 模型部分最後得到測試集預測性能為RF 的AUC 0.828、XGBoost 為0.818 以及CatBoost 為0.805。
結論 : 透過比較LR 與ML 方法來評估ML 方法對於併發症預測領域帶來的主要優勢包括 (1) 對非線性關係建模能力; (2) 對於特徵選擇的高度自適應結構; (3) 選擇性多元的參數調整和模型優化，這些優勢提高了對併發症預測的泛用性與便利性，並且除去特徵挑選流程能夠減少人為疏失。而透過LASSO 與多變量邏輯斯回歸發現吸菸可能是影響放射皮膚炎的重要因素之一，未來於臨床質子治療頭頸癌之放射性皮膚炎風險評估中可以納入為具有預測效力之因子。

Purpose : For the purpose of predicting the occurrence of radiation dermatitis in head and neck cancer patients undergoing proton therapy, this study employs supervised machine learning methods to establish a prognostic model for radiation dermatitis, and compares it with traditional logistic regression.
Materials and methods : This study collected data from 57 head and neck cancer patients treated with Intensity-Modulated Proton Therapy (IMPT). Supervised machine learning (ML) algorithms, including Random Forest (RF), XGBoost (XGB), and CatBoost (Cat), were used for complication prediction modeling and feature selection. These features were then used to establish a Logistic Regression (LR) model. Additionally, predictive models using RF, XGB, and Cat were developed and compared with the LR model. The comparison also included the ML models themselves. Clinical and dose factors were considered, defining severe radiation dermatitis as grade ≥ 2. Feature selection was done through LASSO regularization and Pearson correlation. The dataset was divided into training, validation, and test sets, with 10-fold cross-validation and grid search for hyperparameter tuning. Evaluation metrics included accuracy (ACC), AUC, PPV, NPV, and F1 score.
Results : Among the four feature selection methods for the LR model, the LASSO method performed the best, with an AUC of 0.87 in the test set, selecting six features including smoking history, N stage, AJCC stage, gender, D100_3mm, and V60_5mm. As for the ML models, the final predictive performance on the test set was an AUC of 0.828 for RF, 0.818 for XGBoost, and 0.805 for CatBoost.
Conclusions : By comparing LR and ML methods, the main advantages that ML methods bring to the field of complication prediction include (1) the ability to model non-linear relationships; (2) a highly adaptive structure for feature selection; (3) selective multi-variable parameter tuning and model optimization. These advantages enhance the generalizability and convenience of complication prediction and reduce human oversight by eliminating the feature selection process. Through LASSO and multivariate logistic regression, smoking was found to be one of the significant factors affecting radiation dermatitis. In the future, it can be included as a factor with predictive power in the risk assessment of radiation dermatitis in clinical proton therapy for head and neck cancer.

摘要 i
Abstract iii
致謝 v
目錄 vi
表目錄 viii
圖目錄 ix
符號說明 x
縮寫符號 xi
第一章 緒論 1
1.1 動機 1
1.2 目的 4
1.3 相關文獻探討 5
第二章 材料與方法 7
2.1 前言 7
2.2 患者資料 9
2.3 治療儀器及治療計畫系統 13
2.4 資料預處理 14
2.5 特徵挑選方法 15
2.6 預測模型建立方法 17
2.6.1邏輯斯迴歸 17
2.6.2 隨機森林 18
2.6.3 XGBoost 20
2.6.4 CatBoost 21
2.7 網格搜尋法 22
2.8 十折交叉驗證法 23
2.9 模型效能評估指標 24
2.10 沙普利加法解釋法 25
第三章 結果 27
3.1 前言 27
3.2 特徵挑選結果 28
3.2.1 LASSO挑選結果 28
3.2.2 機器學習挑選結果 32
3.2.2.1 模型特徵重要程度 32
3.2.2.2 機器學習模型特徵SHAP值 34
3.3 邏輯斯迴歸模型結果 37
3.3.1 LASSO特徵-建模結果 37
3.3.2 機器學習特徵-建模結果 39
3.4 機器學習模型超參數優化 40
3.5 機器學習模型預測結果 43
3.5.1 全因子建模結果 43
3.5.2 特徵挑選後-建模結果 44
第四章 討論 45
4.1 機器學習方法討論 45
4.2 機器學習訓練成效探討 47
4.3 預測模型探討 49
4.4 建模演算法比較探討 51
第五章 結論與未來展望 54
5.1結論 54
5.2未來研究 55
參考文獻 56

[1]F. Hegedus, L. M. Mathew, and R. A. Schwartz, "Radiation dermatitis: an overview," International journal of dermatology, vol. 56, no. 9, pp. 909-914, 2017.
[2]N. Salvo et al., "Prophylaxis and management of acute radiation-induced skin reactions: a systematic review of the literature," Current Oncology, vol. 17, no. 4, pp. 94-112, 2010.
[3]T. Wang et al., "Development of a normal tissue complication probability (NTCP) model using an artificial neural network for radiation-induced necrosis after carbon ion re-irradiation in locally recurrent carcinoma," Annals of Translational Medicine, 2021 Jan 2021, doi: 10.21037/atm-20-7805.
[4]D. W. Wen et al., "Normal tissue complication probability (NTCP) models for predicting temporal lobe injury after intensity-modulated radiotherapy in nasopharyngeal carcinoma: A large registry-based retrospective study from China," Radiotherapy and Oncology, vol. 157, pp. 99-105, Apr 2021, doi: 10.1016/j.radonc.2021.01.008.
[5]M. Chen et al., "Predictive performance of different NTCP techniques for radiation-induced esophagitis in NSCLC patients receiving proton radiotherapy," Scientific Reports, vol. 12, no. 1, Jun 2022, Art no. 9178, doi: 10.1038/s41598-022-12898-8.
[6]S. M. Bentzen et al., "Quantitative Analyses of Normal Tissue Effects in the Clinic (QUANTEC): an introduction to the scientific issues," International Journal of Radiation Oncology* Biology* Physics, vol. 76, no. 3, pp. S3-S9, 2010.
[7] A. Jakulin and I. Bratko, "Analyzing attribute dependencies," in European conference on principles of data mining and knowledge discovery, 2003: Springer, pp. 229-240.
[8]P. Samant et al., "Machine learning for normal tissue complication probability prediction: Predictive power with versatility and easy implementation," Clinical and Translational Radiation Oncology, vol. 39, Mar 2023, Art no. 100595, doi: 10.1016/j.ctro.2023.100595.
[9]T. F. Lee et al., "Using Multivariate Regression Model with Least Absolute Shrinkage and Selection Operator (LASSO) to Predict the Incidence of Xerostomia after Intensity-Modulated Radiotherapy for Head and Neck Cancer," Plos One, vol. 9, no. 2, Feb 2014, Art no. e89700, doi: 10.1371/journal.pone.0089700.
[10]C. Lilla et al., "Predictive factors for late normal tissue complications following radiotherapy for breast cancer," Breast cancer research and treatment, vol. 106, pp. 143-150, 2007.
[11]W. Jaschke, M. Schmuth, A. Trianni, and G. Bartal, "Radiation-induced skin injuries to patients: what the interventional radiologist needs to know," CardioVascular and Interventional Radiology, vol. 40, no. 8, pp. 1131-1140, 2017.
[12]X. Yang, H. Ren, X. Guo, C. Hu, and J. Fu, "Radiation-induced skin injury: pathogenesis, treatment, and management," Aging (Albany NY), vol. 12, no. 22, p. 23379, 2020.
[13]J. Kang, R. Schwartz, J. Flickinger, and S. Beriwal, "Machine learning approaches for predicting radiation therapy outcomes: a clinician's perspective," International Journal of Radiation Oncology* Biology* Physics, vol. 93, no. 5, pp. 1127-1135, 2015.
[14]X. Liang et al., "Prognostic factors of radiation dermatitis following passive-scattering proton therapy for breast cancer," Radiation Oncology, vol. 13, no. 1, pp. 1-8, 2018.
[15]K. C. Fang et al., "Acute radiation dermatitis among patients with nasopharyngeal carcinoma treated with proton beam therapy: Prognostic factors and treatment outcomes," International Wound Journal, vol. 20, no. 2, pp. 499-507, 2023.
[16]K.-C. Fang et al., "Dosimetric Parameters Related to Acute Radiation Dermatitis of Patients with Nasopharyngeal Carcinoma Treated by Intensity-Modulated Proton Therapy," Journal of Personalized Medicine, vol. 12, no. 7, p. 1095, 2022.
[17] G. Abdurrahman and M. Sintawati, "Implementation of xgboost for classification of parkinson’s disease," in Journal of Physics: Conference Series, 2020, vol. 1538, no. 1: IOP Publishing, p. 012024.
[18]F. Zhou et al., "Fire prediction based on catboost algorithm," Mathematical Problems in Engineering, vol. 2021, pp. 1-9, 2021.
[19]T. M. Deist et al., "Erratum: â Machine learning algorithms for outcome prediction in (chemo) radiotherapy: An empirical comparison of classifiersâ [Med. Phys. 45 (7), 3449â 3459 (2018)]," 2019.
[20]A. P. Chen et al., "Grading dermatologic adverse events of cancer treatments: the Common Terminology Criteria for Adverse Events Version 4.0," Journal of the American Academy of Dermatology, vol. 67, no. 5, pp. 1025-1039, 2012.
[21]A. Graves and A. Graves, Supervised sequence labelling. Springer, 2012.
[22] X. Wan, "Influence of feature scaling on convergence of gradient iterative algorithm," in Journal of physics: Conference series, 2019, vol. 1213, no. 3: IOP Publishing, p. 032021.
[23]A. Fernández, V. López, M. Galar, M. J. Del Jesus, and F. Herrera, "Analysing the classification of imbalanced data-sets with multiple classes: Binarization techniques and ad-hoc approaches," Knowledge-based systems, vol. 42, pp. 97-110, 2013.
[24]P. Komarek, Logistic regression for data mining and high-dimensional classification. Carnegie Mellon University, 2004.
[25]I. Guyon and A. Elisseeff, "An introduction to variable and feature selection," Journal of machine learning research, vol. 3, no. Mar, pp. 1157-1182, 2003.
[26]V. A. Profillidis and G. N. Botzoris, Modeling of transport demand: Analyzing, calculating, and forecasting transport demand. Elsevier, 2018.
[27]J. Li et al., "Feature selection: A data perspective," ACM computing surveys (CSUR), vol. 50, no. 6, pp. 1-45, 2017.
[28]R. Tibshirani, "Regression shrinkage and selection via the lasso," Journal of the Royal Statistical Society: Series B (Methodological), vol. 58, no. 1, pp. 267-288, 1996.
[29]I. El Naqa et al., "Multivariable modeling of radiotherapy outcomes, including dose–volume and clinical factors," International Journal of Radiation Oncology* Biology* Physics, vol. 64, no. 4, pp. 1275-1286, 2006.
[30]J. R. Quinlan, "Induction of decision trees," Machine learning, vol. 1, pp. 81-106, 1986.
[31] R. J. Lewis, "An introduction to classification and regression tree (CART) analysis," in Annual meeting of the society for academic emergency medicine in San Francisco, California, 2000, vol. 14: Citeseer.
[32]L. Breiman, "Random forests," Machine learning, vol. 45, pp. 5-32, 2001.
[33]L. Breiman, "Bagging predictors," Machine learning, vol. 24, pp. 123-140, 1996.
[34]V. K. Ayyadevara and V. K. Ayyadevara, "Gradient boosting machine," Pro machine learning algorithms: A hands-on approach to implementing algorithms in python and R, pp. 117-134, 2018.
[35] Y. Freund and R. E. Schapire, "Experiments with a new boosting algorithm," in icml, 1996, vol. 96: Citeseer, pp. 148-156.
[36]L. Prokhorenkova, G. Gusev, A. Vorobev, A. V. Dorogush, and A. Gulin, "CatBoost: unbiased boosting with categorical features," Advances in neural information processing systems, vol. 31, 2018.
[37]A. V. Dorogush, V. Ershov, and A. Gulin, "CatBoost: gradient boosting with categorical features support," arXiv preprint arXiv:1810.11363, 2018.
[38] S. Han, C. Qubo, and H. Meng, "Parameter selection in SVM with RBF kernel function," in World Automation Congress 2012, 2012: IEEE, pp. 1-4.
[39]D. Berrar, "Cross-Validation," ed, 2019.
[40]J. M. Lobo, A. Jiménez‐Valverde, and R. Real, "AUC: a misleading measure of the performance of predictive distribution models," Global ecology and Biogeography, vol. 17, no. 2, pp. 145-151, 2008.
[41] C. Goutte and E. Gaussier, "A probabilistic interpretation of precision, recall and F-score, with implication for evaluation," in Advances in Information Retrieval: 27th European Conference on IR Research, ECIR 2005, Santiago de Compostela, Spain, March 21-23, 2005. Proceedings 27, 2005: Springer, pp. 345-359.
[42]D. S. Watson et al., "Clinical applications of machine learning algorithms: beyond the black box," Bmj, vol. 364, 2019.
[43]S. M. Lundberg and S.-I. Lee, "A unified approach to interpreting model predictions," Advances in neural information processing systems, vol. 30, 2017.
[44] H. Deng and G. Runger, "Feature selection via regularized trees," in The 2012 International Joint Conference on Neural Networks (IJCNN), 2012: IEEE, pp. 1-8.
[45]Z. Kovacic, "Early prediction of student success: Mining students' enrolment data," 2010.
[46]I. Khosravi and Y. Jouybari-Moghaddam, "Hyperspectral imbalanced datasets classification using filter-based forest methods," IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 12, no. 12, pp. 4766-4772, 2019.
[47]T. G. Dietterich, "An experimental comparison of three methods for constructing ensembles of decision trees: Bagging, boosting, and randomization," Machine learning, vol. 40, pp. 139-157, 2000.
[48]Z. Cui and G. Gong, "The effect of machine learning regression algorithms and sample size on individualized behavioral prediction with functional connectivity features," Neuroimage, vol. 178, pp. 622-637, 2018.
[49]S. Gao, V. D. Calhoun, and J. Sui, "Machine learning in major depression: From classification to treatment outcome prediction," CNS neuroscience & therapeutics, vol. 24, no. 11, pp. 1037-1052, 2018.
[50]D. K. Smith, H. Clark, A. Hovan, and J. Wu, "Neural network and spline-based regression for the prediction of salivary hypofunction in patients undergoing radiation therapy," Radiation Oncology, vol. 18, no. 1, p. 77, 2023.
[51]P. S. Satheeshkumar, M. El-Dallal, and M. P. Mohan, "Feature selection and predicting chemotherapy-induced ulcerative mucositis using machine learning methods," International Journal of Medical Informatics, vol. 154, p. 104563, 2021.
[52]J. Dean et al., "Incorporating spatial dose metrics in machine learning-based normal tissue complication probability (NTCP) models of severe acute dysphagia resulting from head and neck radiotherapy," Clinical and translational radiation oncology, vol. 8, pp. 27-39, 2018.
[53]D. Jarrett, E. Stride, K. Vallis, and M. J. Gooding, "Applications and limitations of machine learning in radiation oncology," The British journal of radiology, vol. 92, no. 1100, p. 20190001, 2019.
[54]Z. Zhang, Y. Tian, L. Bai, J. Xiahou, and E. Hancock, "High-order covariate interacted Lasso for feature selection," Pattern Recognition Letters, vol. 87, pp. 139-146, 2017.
[55]K. b. halvorsen. "Why is multicollinearity not checked in modern statistics/machine learning." https://stats.stackexchange.com/questions/168622/why-is-multicollinearity-not-checked-in-modern-statistics-machine-learning/168631 (accessed.
[56]P. D. Allison, Logistic regression using SAS: Theory and application. SAS institute, 2012.
[57]J.-z. Feng, Y. Wang, J. Peng, M.-w. Sun, J. Zeng, and H. Jiang, "Comparison between logistic regression and machine learning algorithms on survival prediction of traumatic brain injuries," Journal of critical care, vol. 54, pp. 110-116, 2019.
[58] X. Zhou, K.-Y. Liu, G. Li, and S. Wong, "Model the relationship between gene expression and TFBSs using a simplified neural network with bayesian variable selection," in International Symposium on Neural Networks, 2005: Springer, pp. 719-724.
[59]S. Hwang, G. Yoon, E. Baek, and B.-K. Jeon, "A Sales Forecasting Model for New-Released and Short-Term Product: A Case Study of Mobile Phones," Electronics, vol. 12, no. 15, p. 3256, 2023.