Retrospective single-institution application of a deep learning–based radiomic score in metastatic NSCLC: Potential impact on first-line treatment decisions and outcomes.

Abstract

e20608 Background: Immune checkpoint inhibitors (ICIs) targeting the PD-1/PD-L1 axis are standard of care for metastatic non-small cell lung cancer (mNSCLC). Yet a minority of patients achieve durable benefit from ICI monotherapy (ICI MT), underscoring the limitations of current predictive biomarkers such as PD-L1 tumor proportion score (TPS). Lung Clinical Management (LCM) is a novel deep learning-based radiomic biomarker that classifies patients as LCM-high or LCM-Low (likelihood of ICI MT benefit). We retrospectively evaluated if LCM could guide the selection of first-line ICI MT vs chemo-immunotherapy (IT) in mNSCLC. Methods: We conducted a single-institution retrospective chart review of 27 patients who received anti-PD-1 ICI treatment (alone or in combination with chemo) for mNSCLC between 2012 and 2025. Patients were excluded if they lacked documented PD-L1 TPS, received <1 full dose of disease-directed therapy, or participated in a clinical trial. The LCM score was generated using a radiomics pipeline comprising imaging quality control, preprocessing, expert lesion annotation, deep learning feature extraction, and Cox-based survival modeling. Baseline CT/PET-CT scans and select variables (age, sex, and line of therapy) served as inputs. Model development was blinded to treatment decisions and outcomes. We compared the first-line therapy chosen without LCM to the LCM recommendation. Endpoints included treatment selection, subsequent therapies, treatment-related adverse events, and mortality. Results: The treatment recommended by a clinician with access to LCM aligned with the original treatment in 14 patients (51.85%) and differed in 13 (48.15%). Among these 13 patients, 3 classified as LCM-high received upfront chemo (2 alone, 1 combined with IT); 2 of the 3 developed severe toxicities. Had these patients received ICI MT as suggested by LCM, they might have avoided chemo-related toxicity. 7 of the 13 had PD-L1 TPS ≥50% but were classified as LCM-low, indicating a low likelihood of ICI MT benefit; yet they received ICI MT. 3 of the 7 ultimately received chemo, and all 7 died (4 within 1 year of diagnosis). The remaining 3 of 13 received chemo alone without IT. Conclusions: The prognostic information provided by LCM identified patients who were likely to respond to ICI MT but were instead treated with chemo, potentially exposing them to preventable treatment-related adverse effects. LCM also identified patients unlikely to respond to frontline ICI MT but received it, with some patients subsequently receiving chemo and others expiring before further therapy. These findings suggest that LCM could refine first-line treatment decisions, helping avoid unnecessary toxicity in patients likely to benefit from ICI MT and guiding earlier use of chemo-IT in those less likely to respond to MT.