ARTIFICIAL NEURAL-NETWORK MODEL OF SURVIVAL IN PATIENTS TREATED WITH IRRADIATION WITH AND WITHOUT CONCURRENT CHEMOTHERAPY FOR ADVANCED-CARCINOMA OF THE HEAD AND NECK

Purpose: This study was performed to investigate the feasibility of pr edicting survival in squamous cell carcinoma of the head and neck (SCC HN) with an artificial neural network (ANN), and to compare ANN perfor mance with conventional models. Methods and Materials: Data were analy zed from a Phase III trial in which patients with locally advanced SCC HN received hyperfractionated irradiation with or without concurrent c isplatin and 5-fluorouracil. Of the 116 randomized patients, 95 who ha d 2-year follow-up and all required data were evaluated. ANN and logis tic regression (LR) models were constructed to predict 2-year total su rvival using round-robin cross-validation. A modified staging model wa s also examined. Results: The best LR model used tumor size, nodal sta ge, and race to predict survival. The best ANN used nodal stage, tumor size, stage, and resectability, and hemoglobin. Treatment type did no t predict 2-year survival and was not included in either model. Using the respective best feature sets, the area under the receiver operatin g characteristic curve (A(z)) for the ANN was 0.78 +/- 0.05, showing m ore accurate overall performance than LR (A(z) = 0.67 +/- 0.05, p = 0. 07). At 70% sensitivity, the ANN was 72% specific, while LR was 54% sp ecific (p = 0.08). At 70% specificity, the ANN was 72% sensitive, whil e LR was 54% sensitive (p = 0.07). When both models used the five pred ictive variables best for an ANN, A(z) for LR decreased [A(z) = 0.61 /- 0.06, p < 0.01 (ANN)]. The models performed equivalently when using the three variables best for LR. The best ANN also compared favorably with staging [A(z) = 0.60 +/- 0.07, p = 0.02 (ANN)]. Conclusions: An ANN modeled 2-year survival in this data set more accurately than LR o r staging models and employed predictive variables that could not be u sed by LR. Further work is planned to confirm these results on larger patient samples, examining longer follow-up to incorporate treatment t ype into the model. (C) 1998 Elsevier Science Inc.