Novel machine learning algorithm to predict response to immunotherapy in patients with small cell and non-small cell lung cancer.

Abstract

e13556 Background: Immune checkpoint inhibitors (ICI) are used to manage patients with both small cell (SCLC) and non-small cell (NSCLC) lung cancer. However, ICI response rates are often low, and identifying patients that will benefit from ICIs can be challenging. The value of biomarkers used to predict ICI response, such as PD-L1, Combined Positive Score (CPS) or tumor mutational burden (TMB) have been debated. Furthermore, the resources needed to assess these biomarkers may not be available in many centres. Developing more accurate and accessible tools that predict ICI responses could enable a precision medicine approach that improves patient outcomes. This study aimed to use a novel machine-learning (ML) algorithm to predict response to different ICI therapies in patients with lung cancer based on clinically available data. Methods: 334 eligible records were cleaned and reprocessed from textual to categorical data using one hot encoding. Complete datasets were available for 161 patients. Differences in the data distribution were handled using the Synthetic Minority Oversampling Technique. Six ML algorithms were trained, including Linear regression, Support Vector Classifier, XGBoost Classifier, Random Forest, Decision Tree, and Gaussian Naive Bayes Classifier. These algorithms used 80% of the training data, were tested on 20% of validation data and used the Grid Search Cross-Validation technique for hyperparameter optimization. Results: For the 161 patients included in the final analysis, the mean age was 68 years and 48% were female. 9% of patients had SCLC and 80% had NSCLC. Patients receiving Pembrolizumab, Nivolumab and Atezolizumab comprised 62%, 11% and 25%, respectively. The artificial intelligence (AI) algorithm predicted and stratified ICI response better than PD-L1 levels. Of the ML algorithms, XGBoost Classifier predicted response with the most accuracy, 64% (0.61 F1 score). This model found that good performance status (0-1), female gender and adenocarcinoma sub-type predicted response to ICI. On the other hand, M1, N2 staging, male gender, squamous cell carcinoma sub-type and receiving Atezolizumab were predictive of disease progression. Conclusions: This study developed multiple novel ML models to predict responses to ICIs in lung cancer. XGBoost Classifier used clinically available data to show that the type of ICI a patient receives, their histopathology sub-type and their TMN staging impact ICI response. Future work will aim to improve accuracy and predict ICI toxicity by including data from multiple centres, different cancer types and additional clinical variables. [Table: see text]