(B36) Predicting HER2 Status in Non-Small Cell Lung Cancer using Machine Learning Approach for Precision Cancer Pharmacoepidemiology in Large Databases

Publication Number: 647

Background:
Adaptation of next-generation sequencing (NGS) testing and targeted therapies have improved the survival and quality of life of patients with NSCLC. To conduct valid pharmacoepidemiologic studies of targeted therapies, information on biomarker status such as HER2 alteration and mutation is critical but often not available in large databases.

Objectives:
To develop and compare prediction models for HER2 status in small cell lung cancer (NSCLC) using machine learning approaches.

Methods:
Patients with NSCLC were diagnosed between 01/2011 – 06/2021 with NGS testing performed in over 2800 patients at two NCI-designated cancer centers in New Jersey were analyzed. Demographics, behavioral, and clinical characteristics were extracted from electronic health record (EHR) data and EHR-linked cancer registry. Various machine learning models were developed to predict patient’s biomarker status for 1) HER2 mutation and 2) HER2 alteration in 1) all NSCLC patients and 2) subgroups defined by age, sex and presence of metastasis. Machine learning (ML) methods included logistic regression (LR), random forest (RF), naïve Bayes (NB), extreme gradient boosting (XGBoost) and multi-layer perceptron (MLP). Ten-fold cross validation was used for internal validation. A k-nearest neighbor method was used to impute missing values. Under-sampling, oversampling and feature selection methods were implemented to select the top ten features according to significance. Top features with the highest chi-squared statistic were selected.

Results:
Of 2,804 NSCLC patients, 775 (27%) had NGS reports available (mean age 67, 45% male, 89% non-Hispanic, 65% white). Of 775, 71% had metastasis, 46% were ≤65 years old, 9.2% had HER2 alterations, and 7.4% had HER2 mutation. Based on univariate significance tests with LR on the complete data: cerebrovascular disease, BMI of 27.5-29.9, family history of cancer, KRAS, EGFR and MET had statistically significant odds ratios (ORs) ranging from 0.1-6.3. In models predicting HER2 alteration for overall and sub-groups, area under the ROC curve (AUC) varied from 46% (LR) to 74% (MLP). A model for HER2 alteration in overall had AUC of 64% (RF), model for HER2 alteration in ≤65 age group had AUC of 69% (RF), model for HER2 alteration in females had AUC of 71% (NB), and another model for HER2 mutation in ≤65 had AUC of 74% (MLP).

Conclusion:
BMI, cerebrovascular disease, weight loss, metastasis to liver, diabetes, congestive heart failure, KRAS, EGFR, MET, RET were major predictors. Model performance was satisfactory but good in many models and subgroups, suggesting the need to consider other potential predictors. Further analyses are planned to expand predictors including coded healthcare utilization variables and geographic variables.