Validation of HistoTME-predicted immune subtypes and immunotherapy outcomes using human interpretable features (HIFs) from H&E images in non-small cell lung cancer.

8583 Background: Tumor microenvironment (TME) plays a critical role in tumor progression and response to treatment, especially in improving the response to immune checkpoint inhibitors (ICIs) in non-small cell lung cancer (NSCLC). However, current methods are costly and not feasible for routine clinical use for characterizing the TME. Addressing this, we recently developed Histo-TME, an AI-powered tool that accurately characterizes TME subtypes and ICI responses from routine hematoxylin-eosin (H&E) scanned images. This study aims to validate and interpret HistoTME-predicted subtypes using a series of machine learning (ML) models (PathExplore, PathAI, Boston, MA) that output human interpretable features (HIFs) to quantitatively characterize the TME. Methods: We analyzed 1375 H&E images from 689 NSCLC patients using PathExplore algorithm, which yielded 171 HIFs spatially characterizing tumor-immune cell interactions. Unsupervised k-means clustering (UMAP) identified distinct patient subgroups based on these HIFs. We compared these subgroups to Histo-TME classifications from the same cohort and evaluated their association with immunotherapy response using Kaplan-Meier (KM) and Cox proportional hazards analyses. Results: Five distinct clusters with varying immune infiltration were identified using the 171 HIFs in UMAP clustering. Three out of five clusters characterized by the abundance of macrophages, plasma cells, lymphocytes and fibroblasts within proximity of tumor cells (i.e. 40µm radius), resembled the "Immune Inflamed" subgroup as predicted by HistoTME. There was no survival difference between HIF- and HistoTME-predicted immune inflamed and immune-desert clusters in KM analysis (p>0.05). Both HIF-defined clusters and the HistoTME subtypes had similar median overall survival times (4.33 vs. 4.18 years for "Immune Inflamed", p>0.05; 1.84 vs. 2.62 years for "Immune Desert", p>0.05) in KM analysis. Multivariate analysis adjusted for AJCC stage, age, smoking, ECOG, and CCI demonstrated that both methods have comparable predictive values for overall survival (HIF clusters; HR: 0.69, CI: 0.58-0.81; p<0.001 vs. HistoTME subtypes; HR: 0.80, CI: 0.68-0.94; p=0.008). Conclusions: This study independently validates our published HistoTME method, confirming its ability to accurately identify patients who are likely to respond to ICI therapy using H&E scanned images. Our findings underscore the importance of incorporating TME characteristics using AI-based approaches in routine histopathology and clinical decision-making workflows.