Background: We evaluated the usefulness of artificial neural networks (ANNs) for survival prediction in patients with uterine cervical cancer treated by radiotherapy.
Methods: We used data from 134 patients with uterine cervical cancer treated by combined external and high-dose-rate remote afterloading intracavitary radiotherapy between 1978 and 1993. The ANNs were trained using the data from 67 randomly selected patients. Using the trained ANNs, we predicted the 5-year survival in the remaining 67 patients, and compared it with the known 5-year survival. The performance of the ANNs was evaluated using a receiver operating characteristic (ROC) curve and was compared using the area under the ROC curve (Az).
Results: When fundamental factors, such as age, performance status, hemoglobin, total protein, International Federation of Gynecology and Obstetrics (FIGO) stage, and histological type were used as inputs in the ANNs, Az was 0.5483 ± 0.0145 (mean ± SD). When the histological grading of radiation effect determined by periodic biopsy examination was used in addition to the fundamental factors, Az was highest (0.7782 ± 0.0105). When the cytological grading of radiation effect by the periodic smear was used in addition to the fundamental factors, Az was 0.5523 ± 0.0135, which was not significantly different from that when only the fundamental factors were used.
Conclusion: ANNs allow us to evaluate the importance of prognostic factors, and make it possible to predict the survival of each patient. Using ANNs, the combination of histological grading of radiation effect determined by periodic biopsy examination, in addition to the fundamental factors, is the most effective for prediction of survival in patients with uterine cervical cancer.