Deep-learning prediction of progression-free survival to EGFR inhibitors from H&E tissue slides in advanced EGFR -mutated non-small cell lung cancer

Background: Third-generation tyrosine kinase inhibitors (TKIs) like osimertinib are standard for untreated advanced non-small cell lung cancer (aNSCLC) with common EGFR mutations (exon 19 deletion/L858R). Combination with chemotherapy, bispecific antibodies or anti-angiogenic agents prolong survival but raises toxicity. No biomarker identifies patients benefiting from combinations versus osimertinib alone, yet. Methods: We trained a multiple-instance model to predict progression-free survival (PFS) to first line EGFR-TKIs from hematoxylin-eosin (H&E) tissue slides, adjusting to the TKI type. Diagnostic slides from primary tumors of aNSCLC with common EGFR mutations were digitized at 20× at three independent institutions, and retrospectively included with clinical data. Performance was evaluated with C-index for continuous PFS and AUC for 1-year PFS, using bootstrapping. Results: In total, 141 patients were included (n=73 osimertinib; n=68 erlotinib/gefitinib). We trained on Gustave Roussy (GR, France) and inferred on Campus Bio-Medico (CBM, Italy) and University Hospital of Udine (U, Italy). Patients were mostly female (65%, 62%, 72%), with no smoking habit (63%, 59%, 55%), had adenocarcinoma (92%, 94%, 100%) and exon 19 deletions (56%, 56%, 55%) across GR, CBM, U cohorts. Median (IQR) age was 65 (56-74), 71 (62-78), 68 (63-73) years in GR, CBM, U, respectively. Median true PFS and model performances are presented in the table. Conclusion: This is the first study to predict PFS to EGFR-TKIs from H&E slides. Despite limited sample size, these findings support further investigation of digital pathology for PFS prediction in EGFR -mutated NSCLC. Results on additional cohorts, combination of the model with clinical variables, and interpretability will be presented at the congress. Model's performance across cohorts Cohort Subset #Patients True PFS median (95%CI) C-index continous PFS Mean (95%CI) AUC 1-year PFS Mean (95%CI) All TKIs 78 11.9 (9.5, 15.4) NA-Training cohort NA-Training cohort GR Osimertinib 42 12.6 (9.8, 16.9) NA-Training cohort NA-Training cohort Erlotinib/gefitinib 36 9.9 (7.0, 15.4) NA-Training cohort NA-Training cohort All TKIs 34 9.5 (6.4, 16.3) 0.63 (0.53, 0.73) 0.73 (0.57, 0.86) CBM Osimertinib 16 13.6 (5.5, 23.3) 0.63 (0.49, 0.77) 0.62 (0.40, 0.86) Erlotinib/gefitinib 18 7.4 (2.1, 9.5) 0.59 (0.44, 0.74) 0.77 (0.57, 0.94) All TKIs 29 13.0 (9.4, 17.9) 0.78 (0.70, 0.84) 0.83 (0.69, 0.95) U Osimertinib 15 23.0 (6.7, 50.1) 0.84 (0.70, 0.94) 0.90 (0.72, 1.00) Erlotinib/gefitinib 14 12.7 (8.0, 17.5) 0.74 (0.61, 0.85) 0.73 (0.46, 0.93)