| Summary: | 碩士 === 臺北醫學大學 === 醫學資訊研究所 === 97 === Objective: An artificial neural network (ANN) was developed to predict the pathologic stage of prostate cancer more effectively than regression models based on the combined use of pelvic magnetic resonance imaging (pMRI), prostate specific antigen (PSA) and biopsy Gleason score in patients ongoing receiving radical prostatectomy (RP).
Materials and methods: One hundred and twenty-four patients undergoing retropubic RP or robotic assisted LRP with pelvic lymphadenectomy were evaluated. An ANN was developed using two randomly selected training and validation sets for predicting pathologic stage. Predictive study variables included age, body mass index (BMI), preoperative serum PSA, pathology biopsy Gleason score 1 and Gleason score 2, transrectal ultrasound (TRUS) findings, and digital rectal examination (DRE). The predicted result was a pathological stage of prostate cancer (T2 or T3) after receiving radical surgery. The predicted ability of ANN was compared with those of logistic regression analysis and “Partin Tables” by area under the receiving operating characteristic curve (AUC) analysis.
Results: Of the participants, 40 were prostate cancer with capsule invasion (32.25%) and 84 were prostate cancer without capsule invasion (67.74%). In this model, the hyperbolic and logistic functions were used as an activation function in the hidden and output layers respectively. The LR analysis showed that only PSA and Biopsy pathology Gleason score 1 of the independent variables had a statistically significant influence on prostate cancer with capsule invasion. The classification threshold for predicted values was optimally set to 0.2477. The overall accuracy rate of ANN was 65%, which is higher than that of LR (60%). As to the traditional evaluation tool of prostate cancer, MRI revealed relatively lower predictive ability to previous ANN and LR models. The ANN overall outperformed LR overall significantly (0.795±0.023 versus 0.746±0.025; p＝ 0.016). The ANN testing performed better than LR testing (0.735±0.051 versus 0.65±0.055; p = 0.093). Therefore, we applied each patient of the total data set (n = 124) to the Partin table, the performance of the clinical predictable model showed the AUC of 0.688. The clinically practicable model has worse performance of predictability than ANN and LR models.
Conclusions: ANN was superior to logistic regression and Partin Tables to predict accurately final pathologic result of extracapsular invasion. Artificial neural network models can be developed and used to better predict final pathologic stage of extracapsular invasion when preoperative pathologic and clinical features are known.
|
|---|
