A deep-learning model for immunotherapy efficacy prediction directly from H&E slides in head-neck squamous cell carcinoma (HNSCC).

Abstract

e18006 Background: Immunotherapy (IO) has emerged as the standard of care for recurrent and metastatic HNSCC. However, conventional PD-L1 expression levels offer limited utility for predicting therapeutic response. This study evaluates the application of a deep-learning framework, ENLIGHT-DP, to predict response to IO in HNSCC directly from common H&E slide scans. Methods: IO efficacy and pre-treatment tumor H&E slides were retrospectively collected from 113 HNSCC cases seen between 2020 and 2025 at the Taipei Veterans General Hospital (VGHTPE). On this data we trained, in cross-validation, an attention-based deep-learning model (ENLIGHT-DP) to predict a continuous response score to IO - the ENLIGHT Matching Score (EMS). For clinical practice, we also categorized all patients into three prediction classes: “EMS- H” (top 30% EMS), “EMS-I” (middle 40%) and “EMS-L” (bottom 30%). We defined "PDL1-I" as CPS 1-19, TC 1-9 if CPS is missing, and TPS 1-45 if both are missing. “PDL1-L” and “PDL1-H” were defined by values below and above these thresholds, respectively. Finally, we validated the model on two previously published cohorts collected at Hadassah Medical Center (HMC, n=25) and the BIO2 study from UHN, Toronto (BIO2, n=15). Results: The VGHTPE cohort had a median age of 57.8 years, with 90% males. Primary site distribution was 55% oral cavity, 17% oropharynx (5/19 cases HPV-associated), 15% hypopharynx, 8% larynx, 4% others. 15% of the cases were stage I/II, 9% stage III, 76% stage IV. 85% of cases had received prior radiation therapy, 48% had locoregional disease only, and 55% were treatment-naïve for the recurrent/metastatic disease. The cohort had an ORR of 34.5%, PFS of 5.6 months (3.3-7.9), and OS of 28.1 months (12.7-43.4). The EMS-H group had an ORR of 52.5%, 37% higher than the baseline. The EMS-I group had an ORR of 33.3%, and the EMS-L group 17.5%, 49.3% lower than baseline. The 35% ORR difference between EMS-H and EMS-L more than doubled the 16.3% ORR difference between PDL1-H and PDL1-L (41.3% and 25% respectively). ENLIGHT-DP achieved good stratification of the patient population with respect to PFS (HR: 0.447, p=0.005 for EMS-H vs. rest; HR: 0.343, p=0.0003 for EMS-H vs. EMS-L). This is superior to the PFS separation using PD-L1 (HR: 0.69, p=0.14 for PDL1-H vs. rest; HR: 0.523, p=0.0055 for PDL1-H vs. PDL1-L). ENLIGHT-DP was also borderline significant in stratifying with respect to OS (HR:0.55 , p=0.07 for EMS-H vs rest; HR: 0.577, p=0.16 for EMS-H vs. EMS-L), while PD-L1 was not predictive of OS (HR: 1.12, p=0.71 for PDL1-H vs. rest; HR: 1.1, p = 0.8 for PDL1-H vs PDL1-L). ENLIGHT-DP generalizes to the external cohorts with a ROC AUC of 0.70 in the HMC cohort and 0.67 in the BIO2 cohort, vs. 0.68 in the VGHTPE cohort. Conclusions: ENLIGHT-DP demonstrated a superior prediction performance to IO ORR, PFS, and OS in HNSCC, suggesting a potential clinical utility for treatment guidance.