三種資料探勘法及邏輯斯迴歸預測效能和預測因子之比較－以肝切除手術肝癌患者長期預後為例

摘要研究目的過去肝癌和許多癌症的研究目的相同，多著墨在預後因子的研究及探討，鮮少建構預測模型來做為診斷與預後的預測工具，且資料探勘在肝癌上之應用尚未有系統性之探討。故本研究欲探討三種資料探勘技術【類神經網路(Artificial Neural Networks)、決策樹(Decision Trees)及約略集合論(Rough Set Theory)】、與傳統統計方法【邏輯斯迴歸(Logistic Regression)】預測效能及重要預測因子之差異，並探討影響肝切除肝癌病患長期預後之顯著預測因子及其適用性。期望能透過預測模型之建構，能輔佐臨床醫師對於肝癌病患臨床決策之制定及預後情形之評估與預測，並做為其他相關研究和臨床應用工具之參考。研究方法本研究為縱貫性回溯性方式之研究設計，以南部兩家區域級以上之教學醫院，自2000年1月至2007年12月止在個案醫院確診為原發性肝癌(ICD9=155.0)，並且以初次接受肝切除手術的病患為基礎，排除條件後共有482名研究樣本，且每位病患最少有一年的追蹤期。而後利用統計分析的方法擷取出顯著因子，以作為預測模型之建構，並透過驗證樣本及參數之校正檢驗預測模型是否正確。最終對於不同預測模型之預測效能及重要預測因子進行探討和比較。資料來源主要包括病歷審查及內政部死亡檔。統計分析利用Cox Proportional Hazard Model Univariate Analysis分別投入個別可能影響的因子，得到未調整風險比率後，再以Log-Rank test來檢定其差異，以擷取出影響整體存活及無復發存活之單變量顯著因子。而後分析訓練樣本及驗證樣本之人口學特質及疾病特質、手術過程與手術預後等變項，類別變項以次數及百分比(%)呈現樣本型態並以χ2 test以及Fisher exact test進行檢定，P value＜0.05為達到統計學上顯著之差異。而本次研究評估預測模型預測能力的指標主要包含兩大部份：第一部份為預測準確率；第二部份為ROC曲線(Receiver Operating Characteristic Curves)。研究結果預測效能方面，在內部驗證中整體存活及無復發存活預測準確度平均依序為RST(0.997)、ANN(0.972)、DT(0.813)及LR(0.777)，AUROC平均依序為ANN(0.982)、LR(0.805)及DT(0.777)，敏感度平均依序為RST(0.997)、ANN(0.974)、DT(0.720)及LR(0.669)，特異度平均依序為RST(0.976)、ANN(0.960)、DT(0.724)及LR(0.682)；在外部驗證中預測準確度平均依序為ANN(0.763)、LR(0.757)、DT(0.744)及RST(0.715)，AUROC平均依序為ANN(0.86)、LR(0.760)及DT(0.667)，敏感度平均依序為ANN(0.761)、RST(0.634)、DT(0.598)及LR(0.587)，特異度平均依序為ANN(0.680)、RST(0.668)、LR(0.656)及DT(0.568)。重要預測因子方面，在預測整體存活中，重要預測因子數分別為ANN(9-15)、LR(4-8)、DT(7-14)、RST(9-12)；在預測無復發存活中則分別為ANN(11-17)、LR(1-4)、DT(4-13)、RST(14-15)。結論與建議從預測效能之結果看來，可得知類神經網路等資料探勘演算法適用臨床表徵等資料來預測肝切除肝癌病患長期預後之情形。就普遍通用性內部驗證及外部驗證來說ANN及RST之預測效能皆優於LR，而DT和LR則是擁有相似之預測效能，並無太大之差異。若比較ANN、DT及RST之預測效能來說，ANN及RST內推性能力相似且皆優於DT，外推性能力則屬ANN最佳，其次為RST，最差為DT。在重要預測因子差異之情形上，ANN、DT及RST等三種資料探勘演算法皆比LR能擷取出更多的重要預測因子。但不同的預測模型因各有其優缺點，如LR能找出變項之方向性及重要性，DT及RST則是能產生決策規則或關聯規則。不同之預測模型皆具有其特殊的優劣性及適用性，因此並無相互之替代性，故建議臨床醫師於制定臨床決策及評估與預測病患預後情形時，應多以資料探勘之預測模型來加以輔佐傳統統計之分析方法。同理，亦建議未來後續相關之醫學研究應須考量到資料特性及研究目的，選取適當之預測模型來加以應用，也建議後續研究者將資料探勘等方法應用於其它臨床醫學或是癌症相關之疾病以進行系統性之探討。關鍵詞：類神經網路、邏輯斯迴歸、決策樹、約略集合論、預測效能、重要預測因子

關鍵字

類神經網路；邏輯斯迴歸；決策樹；約略集合論；預測效能；重要預測因子

並列摘要

Abstract Objective In the past, the main aims of HCC (Hepatocellular carcinoma) were similarly with other cancer, it were more likely to discuss the significant factors of prognosis, and seldom to construct the predictive model of long-term prognosis. In addition, the application of data mining in HCC hasn’t had systematic discussion in the past. In present study, we want to compare three data mining methods (artificial neural networks, decision trees and rough set theory) with traditional method (logistic regression) about predicting ability and predicting important factor, then discussing the significant factors of HCC patients’ long-term prognosis and the factors’ suitable. We are looking forward to construct the predicting model can assist the clinical physician in decision making, and be other clinical research reference materials. Methods A series patients with HCC (ICD9=155.0) who underwent liver resection during a eight year period (from January 2000 to December 2007) at a medical center and a regional medical teaching hospital in the South of Taiwan, were recruited for this study. Among the recruited cases, we excluded patients who underwent the exclusion criteria. Finally, 482 patients were included in our empirical analysis. The period follow-up of each patient has at least one year. Then the statistics was used to retrieve the significant factors to construct predicting model of prognosis, the validation sample and parameter were used to test the model. Finally, we compare different predicting model on predicting ability and predicting important factor. The data were collected from medical chart review and mortality data bank established by the Statistics Office, Department of Health, Taiwan. Statistical analysis For evaluation of outcome predicator, the Cox proportional hazard model was used to retrieve the significant factors of overall survival and disease-free survival, and differences in factors were examined using the log-rank test. Then, the χ2 analysis and Fisher’s exact test were used to analysis the training sample and validation sample, and to test for differences in variables. A p-value less than 0.05 was considered significant. The main evaluated indicators of predictive ability were accuracy and receiver operating characteristic curves (ROC curves). Results In predictive ability, in internal validation, the ordinal average of accuracy were RST(0.997), ANN(0.972), DT(0.813) and LR(0.777), respectively. The ordinal average of AUROC were ANN(0.982), LR(0.805) and DT(0.777), respectively. The ordinal average of sensitivity were RST(0.997), ANN(0.974), DT(0.720) and LR(0.669), respectively. The ordinal average of specificity were RST(0.976), ANN(0.960), DT(0.724) and LR(0.682), respectively; In external validation, the ordinal average of accuracy were ANN(0.763), LR(0.757), DT(0.744) and RST(0.715), respectively. The ordinal average of AUROC were ANN(0.86), LR(0.760) and DT(0.667), respectively. The ordinal average of sensitivity were ANN(0.761), RST(0.634), DT(0.598) and LR(0.587), respectively. The ordinal average of specificity were ANN(0.680), RST(0.668), LR(0.656) and DT(0.568), respectively. In predicting important factors, in predicting overall survival, the number of predicting important factors were ANN(9-15), LR(4-8), DT(7-14) and RST(9-12), respectively; In predicting disease-free survival, the number of predicting important factors were ANN(11-17), LR(1-4), DT(4-13) and RST(14-15), respectively. Conclusion and suggestion According to the result, we know the data mining methods can used the clinical features to construct the predicting model of HCC patients’ prognosis. In common internal validation and external validation, the predicting ability of ANN and RST are better than LR, but the DT is similarly with LR. If compare the three data mining methods, the ANN is similarly with RST in internal validation, and they are better than DT. In the external validation, the predicting ability of ANN is the best of them, the next id RST, and the DT is the worst of them. In predicting important factors, three data mining methods can extract more predicting important factors than LR. But different predicting model have own unique characteristic, like the LR can find the direction and significance of variables, the DT and RST can produce the decision rules or relational rules. Different predicting model have own unique characteristic and suitable. So they can’t substitute for each other. Suggesting the clinical physician can used the data mining methods to assist traditional statistics when decision making. On the other hand, suggesting the follow-up researcher can consider data feature and research purpose to use the suitable predicting model. Finally, also suggesting the follow-up researcher can apply the data mining methods in other clinical issues or the related disease of cancer. Key word：Artificial neural networks, logistic regression, decision trees, rough set theory, predicting ability, predicting important factor