| Summary: | 碩士 === 臺北醫學大學 === 醫學資訊研究所 === 102 === There are many treatment methods to patients with hepatocellular carcinoma (HCC) according to cancer stage. Hepatic resection is a favorite treatment method in patients with early stage HCC without liver dysfunction. But most patients have hepatitis or cirrhosis which makes liver resection is not
feasible .Therefore, radiofrequency ablation (RFA) has become the effective, alternative and important treatment for liver cancer or waiting for liver transplantation. High recurrence rate after treatment of HCC is still a very important issue. Early recurrence affects the long-term survival of patients .For patients on the transplantaion waiting list, adequate treatment to HCC is also important due to the limit number of organs .Therefore, we try to use biological neural network to analyze non-linear data by artificial neural network (ANN) prediction model. By using liver cancer related multivariable factors to provide accurate prediction of early (1- year，2- year) disease-free survival after radiofrequency ablation in patients with hepatocellular carcinoma.
This study is a retrospective study, tracking from January 2009 to April 2012 using computer tomography-guided percutaneous radiofrequency ablation for the treatment of patients with hepatocellular carcinoma, collating disease-free survival of patients with a total of 252 patients in the 1-year, with the 2-year disease-free survival of patients with a total of 179 patients, and applied 15, 8 and 6 items which correlated with hepatocellular carcinoma as the input variables respectively. To analyze of the effectiveness of prediction model was based on internal assessment of retrospective data and external data validation group, the effectiveness of its analysis models were tested by simulation prospective study assessment. To build eight types of artificial neural network prediction model and analyze the performance and assess its model.
The results showed that 15 input variables have better predictive power in all prediction models .The 1-year disease-free survival in the retrospective data analysis, the accuracy of the prediction performance group were 85.0%, sensitivity 75.0%, specificity 87.5%, AUC 0.84 in internal validation. And in external validation group, the accuracy of prediction performance were 70%, sensitivity 63.6%, specificity 71.8%, AUC 0.77, all of them had good predictive performance. The retrospective data of input 15 variables in 2-year disease-free survival, the internal validation the accuracy of the prediction performance of the group of 67.9%, sensitivity of 50.0%, specificity of 85.7%, AUC of 0.75 and external validation group the accuracy of the prediction performance of the group 63.9% , sensitivity 56.3%, specificity 70.0%, AUC0.72 both performance were reached medium predictive power probably due to small number of training samples.
This study utilize artificial intelligence data mining techniques to construct a good prediction model. This study showed that the development of predictive models have moderate predictive capabilities which could be used as assisting clinicians to improve clinical decisions and prognostic evaluation.
|
|---|
