(A-049) Prevalence Predictive Modelling of HER2 status in Non-Small Cell Lung Cancer in large multicenter data using machine learning approach and application to SEER: A pilot test

Background: Challenges of diagnosing HER2 alterations and limited approved targeted therapies remain an unmet clinical need for Non-Small Lung Cancer (NSCLC) patients with HER2 alterations. For the development of targeted therapies and prediction models for patient selection understanding of prevalence of HER2 alteration in large databases is critical and is not very well recorded.

Objectives: To develop, validate and apply prediction models for HER2 in NSCLC patients using machine learning approach.

Methods: Electronic health records (EHR) and EHR-linked registry data of 3730 NSCLC patients with NGS reports, diagnosed over past 5-10 years at three large NCI designated cancer centers were analyzed. Demographics, behavioral, and clinical characteristics were extracted using automation and manual review. Machine learning approach (logistic regression (LR), random forest (RF), naïve Bayes (NB), extreme gradient boosting (XGBoost) and multi-layer perceptron (MLP)) was used to build models. Models were externally validated on datasets independently and calibrated using cross validation and Akaike Information Criterion (AIC). The best performing model based on AUC and accuracy was pilot tested on Surveillance, Epidemiology, and End Results Program (SEER) data to estimate the prevalence of HER2 alteration in NSCLC patients.

Results: Of 3730 NSCLC patients, 1468 (40%) were tested for HER2 alteration with NGS reports available. Of dataset (A) of 761 patients, 9.3% had HER2 alterations and 7.4% had HER2 mutation. Of dataset (B) with 693 patients 7.4% had HER2 alteration. In both datasets, KRAS biomarker, metastasis to other lobes or other lungs, significant weight loss, and diabetes were among the top variables of importance. Multivariable analysis on the complete patient cohort in each dataset yielded similar AUC, ranging from 59-64% with best AUC in “only females” subgroup logistic regression model at dataset B with AUC 75%. Mean AUC of externally validated models varied from 44-64% with best performing model being dataset B logistic regression model on subgroup “only females” with AUC 64%. The random forest model developed on limited variables similar to SEER on dataset A performed better with mean AUC of 60% and estimated prevalence of HER2 alteration 6% in national NSCLC population

Conclusions: Prevalence of HER2 in NSCLC is higher than the literature suggests that HER2 alterations, while relatively rare, represent a significant subset of NSCLC with distinct biological and clinical characteristics. The model performance was modest but good in certain subgroups, advocate future research based on bigger and prospective studies for larger sample in subgroups and capture missing or additional predictors such as biomarkers and healthcare utilization variables.