Artificial neural network model analysis in T790M detection after failure of first- and second-generation epidermal growth factor receptor tyrosine kinase inhibitors.

Abstract

e21041 Background: Despite osimertinib availability, first or second generation (1st/2nd gen.) epidermal growth factor receptor tyrosine kinase inhibitors (EGFR TKIs) are commonly used in many countries. Clinical factors identified before first line EGFR TKIs treatment that correlate with T790M occurrence are not well-known. In this single center retrospective study we define the artificial intelligence-based model that support T790M detection in daily practice. Methods: In 80 patients (pts) consecutively progressed on 1st/2nd generation EGFR TKIs T790M mutation was assessed. In multivariate logistic regression factors predictive to T790M mutation were identified. Based on clinical and molecular factors present before EGFR TKIs initiation multi-layer perceptron (MLP) neural network was used for building model classifying patients according to T790M status. The artificial neural network consisted of three layers: input, hidden and output layers of 36, 6 (optimally due to the result of the neural network) and 1 neurons, respectively. During the learning process the weight links between the neurons were modified using Broyden-Fletcher-Goldfarb-Shanno algorithm. To calculate the diagnostic quality of the artificial neural network, we divided at random of all patients for two groups: training and testing. The training group contained 65 and testing group 15 cases. Sensitivity and specificity for both models were assessed. TIBCO Software Inc., Statistica (data analysis software system), version 13 software was used for all calculations. Results: In 49/80 pts (61.3%) T790M was present. In logistic regression model: del19 EGFR mutation, chest or liver metastases and 1st gen. TKI treatment correlated with T790M presence. Parameters included in building neural network model were: del19/L858R/other EGFR mutation status, sex, age, 1st/2nd gen. TKI in first line, clinical stage, metastases location, ECOG PS, smoking history. Learning and testing phase models detected T790M with sensitivity-94.7%, specificity – 91.3% and sensitivity-77.8%, specifity-83.3%. In testing set (15 pts) proposed neural network model achieved 80% accuracy in predicting correctly T790M status. Clinical stage, TKI generation in first line, metastases location, EGFR mutation were the most valuable entrance parameters. After conducting a sensitivity analysis for all 8 the analysed variables the observed error quotients were above 1, therefore, they are all significance and should be entered as input to the developed network. Conclusions: The proposed artificial neural network model used parameters identified before start of TKI treatment and achieved high accuracy in T790M status prevalence. Although this model need to be validated in large cohort it can be an useful tool to support clinicians decision.