Assessment of electronic health record (EHR) –based machine learning (ML) in predicting risk of brain metastasis among patients with early-stage non–small-cell lung cancer (eNSCLC).

Abstract

2036 Background: Among patients (pts) with eNSCLC, development ofbrain metastasis (BM) is a poor prognostic sign, but routine surveillance for BM is not recommended post-therapy. EHR-based ML algorithms may identify pts who would benefit from active brain MRI surveillance and/or treatment intensification in the early-stage setting. Methods: ConcertAI Patient360 database, consisting of structured and curated EHR records from pts receiving care in ~900 US oncology clinics, was used to identify pts diagnosed with stage IB-IIIA NSCLC without prior evidence of BM between Jan 2010 – December 2021. Presence of BM was identified from human curation of patient documents. Gradient boosting (GB), random forest (RF), and logistic regression (LR) algorithms with 3-fold cross-validation were trained and compared to predict risk of BM at two landmarks: 18 and 24 months (mos) from initial diagnosis (index). Pts who did not develop BM and either died or were lost to follow-up prior to landmark times were removed. Feature importance was defined using Shapley Additive Explanation (SHAP)-derived marginal odds ratios (OR). Due to low prevalence of BM, AUPRC was used as performance metric. Pts in the 1st and 4th quartiles of predicted probabilities from GB model were flagged as low-risk vs. high-risk. Results: Among 7473 pts in 18 mos model, median age was 68.4 years (IQR 13.2), 50.5% were female, and 10.9% were black. Demographics were similar for 6863 pts in 24 mos model. 6.4% and 8.3% developed BM at 18 and 24 mos. Ability of GB, RF, and LR models to predict BM was similar with validation AUPRC of 0.109 at 18 mos and 0.137 at 24 mos. In the GB model, BM prevalence in high-risk vs. low-risk group was 10.3% vs. 3.1% at 18 mos and 13.4% vs. 4.4% at 24 mos. In both landmark models, N0 stage and surgery within 90 days after index diagnosis were protective against BM, while presence of EGFR targetable mutations, adenocarcinoma (AD) histology, higher platelets (PLT), and history of pneumonia were risk factors. A glucose-by-histology interaction was found: For pts with normal blood glucose (GLU), risk of BM was independent of histology, for pts with high GLU, AD conferred greater risk of BM. Conclusions: An EHR-based ML model identified risk factors for developing BM among pts with eNSCLC and may identify pts who would benefit from active brain MRI surveillance and treatment intensification. [Table: see text]