A clinical decision support system to predict the efficacy for EGFR-TKIs based on artificial neural network

Abstract

BackgroundThe e cacy of epidermal growth factor receptor (EGFR) -tyrosine kinase inhibitor (TKI) was affected by numerous factors. We developed and validated an arti cial neural network (ANN) system based on clinical characteristics and next-generation sequencing (NGS) to support clinical decisions. MethodsA multicenter retrospective non-interventional study was conducted. 196 untreated patients from three hospitals with advanced non-small cell lung cancer (NSCLC) and EGFR mutation were tested by NGS before the rst treatment. All patients received formal EGFR-TKIs treatment. Five different models were individually trained to predict the e cacy of EGFR-TKIs based on an independent cohort. Others from two independent cohorts were collected for external validation. ResultsCompared with logistic regression, four machine learning methods show better predicting abilities for EGFR-TKIs. The inclusion of mutations further improved the predictive power in all models. ANN performed best on the dataset with mutations TP53, rb1 and pik3ca. The prediction accuracy on the test set reached 84%, the recall on the test set was about 95% for poor e cacy and ROC was 0.67 In the external validation, ANN still showed the best performance and differentiated patients with poor outcomes. Finally, a clinical decision support software based on ANN was developed and provides a visualization interface for clinicians ConclusionThis study provides an approach to assess the e cacy of NSCLC patients with rst-line EGFR-TKI treatment. A software is developed to support the clinical decision. inhibitors (TKIs) as a standard rst-line regimen for the treatment of advanced lung cancer patients with EGFR mutation. EGFR-TKIs have extended the survival for NSCLC.Drug resistance greatly limits the usefulness of EGFR-TKIs [4]. Although the response rate of EGFR-TKIs as rst-line treatment was even up to 70% and the progression-free survival (PFS) was around 10 months[5], EGFR-TKIs are still less effective or even ineffective for some patients. EGFR-TKIs combined with chemotherapies or anti-angiogenic therapies have been proved to prolong the PFS and increase toxicity in some patients [6,7]. The PFS in some studies has also translated into an increased overall survival (OS) [7]. Screening ineffective patients and bringing earlier interventions have great clinical implications.Next-generation sequencing (NGS) is widely available. It is revealing the high heterogeneity of lung cancer at a molecular level. With massive sequencing, more resistance mutations in EGFR mutated patients were discovered. Many mutations have been proved to be associated with poor e cacy of EGFR-TKIs [8]. Recently, deep learning (DL) has been widely used in different elds of medicine. Models based on DL have been demonstrated to predict prognosis and risks of various diseases [9]. At the present stage, there are some excellent predictive models to identify the e cacy of drugs [10,11]. Most of these models are based on radiomics methods. It increases the di culty of clinical translation...