Pathologic stage is an important predictor of prostate cancer outcome and several ANN models have been developed to predict pathologic stage from clinical variables. In 1998, Tewari and Narayan reported a pilot study using an ANN to develop a staging tool for clinically localized prostate cancer.33
Data from 1200 men from four institutions were used to develop the model to predict three possible outcomes: positive surgical margins, seminal vesicle involvement and lymph node involvement. The input variables used included race, age, DRE findings, tumor size by ultrasound, serum PSA level, biopsy Gleason sum, perineural infiltration and biopsy staging findings, such as the number of positive cores. The resulting ANN model had a sensitivity of 81%, specificity of 75% and an AUROC of 0.79 for predicting positive margins. For predicting seminal vesicle involvement, the model produced a sensitivity of 100%, specificity of 72.1% and an AUROC of 0.80. The ANN's performance on predicting lymph node spread was 73% and 83% for sensitivity and specificity, respectively. The AUROC for predicting lymph node spread was 0.77. The authors concluded that the ANN model was accurate enough to miss less than 10% of patients with positive margins, less than 2% with positive lymph nodes, and none with seminal vesicle involvement. Further, they concluded that implemen tation of such a model could preclude unnecessary staging tests at a significant cost saving.
In 2000, Crawford and coworkers developed an ANN staging model to predict lymph node spread in men with clinically localized prostate cancer using data from 6454 patients from two institutions.34 The inputs used were clinical stage, biopsy Gleason sum and PSA level. Two validation sets from separate institutions were used. The model's sensitivities on each of the validation sets were 64% and 44%, specificities were 82% and 81%, and the AUROCs were 0.81 and 0.77. The authors concluded that their results suggest a role for ANNs in the accurate staging of prostate cancer. Further work by this group includes a model to predict the risk of non-organ-confined prostate cancer in men who are clinically localized.35 This work resulted in a model with a sensitivity of 72% and a specificity of 67%. The AUROC was 0.76. Another ANN model by this group was able to predict capsular penetration (CP) in 83% of those patients with CP and had a false-negative rate of 16%.36 Several working ANN models published by Crawford and coworkers have been made available on the internet for use by patients and physicians (www.annsincap.org).
Was this article helpful?