RESUMO
OBJECTIVES: Stereotactic radiosurgery (SRS) is a minimally invasive modality for the treatment of trigeminal neuralgia (TN). Outcome prediction of this modality is very important for proper case selection. The aim of this study was to create artificial neural networks (ANN) to predict the clinical outcomes after gamma knife radiosurgery (GKRS) in patients with TN, based on preoperative clinical factors. PATIENTS AND METHODS: We used the clinical findings of 155 patients who were underwent GKRS (from March 2000 to march 2015) at Iran Gamma Knife center, Teheran, Iran. Univariate analysis was performed for a long list of risk factors, and those with P-Value < 0.2 were used to create back-propagation ANN models to predict pain reduction and hypoesthesia after GKRS. Pain reduction was defined as BNI score 3a or lower and hypoesthesia was defined as BNI score 3 or 4. RESULTS: Typical trigeminal neuralgia (TTN) (P-Value = 0.018) and age>65 (P-Value = 0.040) were significantly associated with successful pain reduction and three other variables including radiation dosage >85 (P-Value = 0.098), negative history of diabetes mellitus (P-Value = 0.133) and depression (P-Value = 0.190). On the other hand, radio dosage>85 (P-Value = 0.008) was significantly associated with hypoesthesia, other related risk factors (with p-Value<0.2), were history of multiple sclerosis (P-Value = 0.106), pain duration more than 10 years before GKRS (P-Value = 0.115), history of depression (P-Value = 0.139), history of percutaneous ablative procedures (P-Value = 0.148) and history of diabetes mellitus (P-Value = 0.169).ANN models could predict pain reduction and hypoesthesia with the accuracy of 84.5% and 91.5% respectively. By mutual elimination of each factor in this model we could also evaluate the contribution of each factor in the predictive performance of ANN. CONCLUSIONS: The findings show that artificial neural networks can predict post operative outcomes in patients who underwent GKRS with a high level of accuracy. Also the contribution of each factor in the prediction of outcomes can be determined using the trained network.
