Summary
Despite guidelines recommending biomarker testing for all advanced NSCLC cases, many remain undocumented or untested. Predicting testing patterns is critical for guideline adherence, addressing care gaps, and optimizing trial site selection. We implemented a foundation model (FM) using a time-aware transformer with lightweight adaptors to predict testing patterns.
Methods
ConcertAI’s Patient360™ provided manually curated EHR data from 70,622 NSCLC pts to develop and validate the FM. The dataset was divided into training (80%; N=56,492), validation (10%; N=7,062), and test (10%; N=7,063) sets. The FM employed hierarchical attention to capture inter- and intra-variable across timeframes. Self-supervised learning masked ≈10% of tokens/iteration, and training ran for 350 epochs. Patient embeddings were generated from up to 400 tokens/sequence encompassing 85 unique tokens from 35 clinical variables. Biomarker data was withheld during training and evaluation. A four-layer MLP adaptor predicted testing patterns from embeddings, and performance metrics were compared with literature-reported testing prevalence.
Results
The FM learned embeddings from ≈22M tokens (56,492pts × 400 tokens/pt). AUCs ranged from 0.91-0.97 for ALK, BRAF, EGFR, ROS1, MET, and RET, but were 0.86 for PD-L1 testing. This reflects real-world workflows, where PD-L1 requires separate IHC testing while other biomarkers are assessed via multigene panels. The FM's PPV was 92% for PD-L1 testing (79% literature prevalence) and 97%-99% for gene-panel biomarkers (85%-88% prevalence). Compared to literature-reported untested/undocumented rates of 10–20%, FM identified 34–70% of cases—a 3–5x increase over random selection.
Conclusions
This FM effectively predicts testing patterns for common biomarkers. High PPV and NPV suggest reliable identification of tested and untested/undocumented cases. Performance differences between panel biomarkers and PD-L1 reflect workflow nuances, supporting the model’s clinical relevance.
