預測肝細胞癌病患發生同時性與異時性第二癌的風險及其預後

Predicting the Risk of Synchronous and Metachronous Second Primary Cancers and Prognosis in Patients with Hepatocellular Carcinoma

王昱翔(Yu-Xiang Wang)

指導教授 ： 張啟昌

中山醫學大學/健康管理學院/醫學資訊學系/碩士(2022年)

https://doi.org/10.6834/csmu202200275

摘要

背景: 肝細胞癌病患存活率的提升使得第二癌逐漸成為臨床管理重要的議題。本研究針對肝細胞癌病患發生第二癌相關的危險因子進行整合性分析，比較不同預測模式的預測能力與臨床效益，提供臨床風險管理。

方法: 資料來源為五家醫院癌症登記資料庫的肝細胞癌病患，有效資料共10,741筆。依據文獻與臨床專家意見共選取18個危險因子，使用統計與機器學習分類器進行特徵選取的差異比較。第二癌預測分類模式為CART (Classification And Regression Tree)、C4.5 (C4.5 Decision tree)、RF (Random Forest)與C5.0 (C5.0 Decision tree)使用準確性、敏感度、特異度、F度量評分、Kappa、Matthews相關係數和曲線以下面積(AUC) 作為績效評估，並利用決策曲線分析(Decision Curve Analysis)和臨床影響曲線(clinical impact curve)比較各預測模式的臨床效益。

結果: 特徵選取的結果顯示整體前三名為：甲型胎兒蛋白、根治性治療、總膽紅素；同時性第二癌與異時性第二癌共同危險因子為:總膽紅素、根治性治療、腫瘤大小；手術切除危險因子為化學治療；射頻消融危險因子為腫瘤大小；經動脈導管肝臟腫瘤注射化學栓塞治療危險因子為總膽紅素。預測模式結果顯示在整體、異時性、手術切除與射頻消融的AUC值最高均為RF；同時性的AUC值最高為CART；經動脈導管肝臟腫瘤注射化學栓塞治療的AUC值最高為C4.5。決策曲線分析結果顯示，整體肝細胞癌以BC和C5.0的預測臨床效益最高；同時性第二癌以C5.0的預測臨床效益最高；異時性第二癌以RF的預測臨床效益最高；手術切除以C5.0和RF的預測臨床效益最高；射頻消融治療以C4.5的預測臨床效益最高；經動脈導管肝臟腫瘤注射化學栓塞治療以CART的預測臨床效益最高。

討論/結論: 肝細胞癌倖存者面臨最大的臨床挑戰是第二癌的發生。本研究透過台灣癌症登記資料庫數據集，分析肝細胞癌病患發生第二癌危險因子為甲型胎兒蛋白、根治性治療與腫瘤大小；肝纖維化是同時性第二癌獨立危險因子；異時性第二癌危險因子為腫瘤大小、根治性治療、BCLC期別；化學治療是手術切除獨立危險因子；射頻消融治療危險因子為B型肝炎、腫瘤大小和BMI；經動脈導管肝臟腫瘤注射化學栓塞治療獨立危險因子為總膽紅素。此外，本研究使用決策曲線分析和臨床影響曲線比較各預測模式的臨床效益和準確度，用於協助醫師對於不同風險門檻值的最佳預測模式及其風險解釋，為病患提供最大的臨床益處。

關鍵字

肝細胞癌 ； 同時性與異時性的二癌 ； 決策曲線分析 ； 臨床影響曲線

並列摘要

Background: As the survival rate in patients with hepatocellular carcinoma (HCC) has increased, second primary cancer (SPC) has become an increasingly important issue in clinical management. In this study, risk factors associated with HCC were analyzed along with risk prediction models, and clinical risk management was provided.

Methods: A total of 10,741 valid data were gathered from five hospital cancer registries. We analyzed 18 risk factors based on the literature and the opinions of clinical experts with statistical and machine learning methods. For prediction classification, CART (Classification And Regression Tree), C4.5 (C4.5 Decision Tree), RF (Random Forest), and C5.0 (C5.0 Decision Tree) are used. Area Under Curves (AUCs), Kappa, Matthews correlation coefficients, and F1 scores are used for evaluating performance. A clinical benefit comparison is conducted using Decision Curve Analysis (DCA), and clinical impact curve.

Results: Alpha-Fetoprotein (AFP), curative treatment, and tumor size were the top three important features for HCC patients with SPC. A common risk factor for both synchronous and metachronous SPC patients were total bilirubin, curative treatment, and the size of the tumor. The risk factor of surgical resection, Radiofrequency Ablation Therapy (RFA) and Trans-Arterial Chemo-Embolization (TACE) were chemotherapy, tumor size and total bilirubin, respectively. As a result of the prediction model results, RF has the highest AUC value overall, metachronous SPC, surgical resection, and RFA. In synchronous SPC, CART was the highest AUC value. In TACE, C4.5 was the highest AUC value. DCA results showed that BC and C5.0 was most clinically beneficial to overall HCC patients. C5.0 provided the greatest clinical benefit for synchronous SPC patients, RF provided the greatest clinical benefit for metachronous SPC patients, C5.0 and provided the greatest clinical benefit for surgical resection patients, C4.5 provided the greatest clinical benefit for RFA patients, and CART provided the greatest clinical benefit for TACE patients.

Discussion/Conclusion: One of the greatest challenges facing HCC survivors is SPC. Our study found that AFP, curative treatment, and tumor size were the most common risk factors for SPC development. Liver fibrosis is an independent risk factor for synchronous SPC, while, tumor size, curative treatment, BCLC stage (Barcelona clinic liver cancer staging) and BMI (Body mass index) are risk factors for metachronous SPC. chemotherapy is an independent risk factor for surgical resection while, HBV (Hepatitis B virus), tumor size, and BMI were risk factors for RFA, and total bilirubin is an independent risk factor for TACE. Additionally, this study uses decision curve analysis and clinical impact curve to compare the clinical benefit and accuracy of each prediction models, which is used to assist physicians in determining the best predictive model and risk interpretation for different risk thresholds to maximize patient benefit.

並列關鍵字

hepatocellular carcinoma ； synchronous and metachronous second primary cancers ； decision curve analysis ； clinical impact curve

參考文獻

Allan, J. M., Travis, L. B. (2005). Mechanisms of therapy-related carcinogenesis. Nat Rev Cancer, 5(12), 943-955.

Bewick, V., Cheek, L., Ball, J. (2004). Statistics review 13: receiver operating characteristic curves. Crit Care, 8(6), 508-512.

Breiman, L. (2001). Random forests. Machine learning, 45(1), 5-32.

Brown, M. (2018). rmda: Risk Model Decision Analysis. https://CRAN.R-project.org/package=rmda

Chang, C. C., Chen, C. C., Cheewakriangkrai, C., Chen, Y. C., Yang, S. F. (2021). Risk Prediction of Second Primary Endometrial Cancer in Obese Women: A Hospital-Based Cancer Registry Study. Int J Environ Res Public Health, 18(17).