8578 Background: In development of effective treatment such as immune checkpoint inhibitors (ICIs) among patients with small-cell lung cancer (SCLC), the absence of biomarkers is critical. Tumor-infiltrating lymphocytes are the main activator of antitumor immunity and could be a promising biomarker if TIL can be objectively assessed throughout the whole tumor immune microenvironment (TIME). To evaluate TIME, pathological assessment is one of the easiest ways. Recently, several studies showed that machine learning analysis can assess type of lymphocytes and their localization in pathological images. Here, we aimed to develop a novel biomarker to predict efficacy of ICI in SCLC using machine learning of pathological images. Methods: This study was a biomarker analysis of the APOLLO study which was 32-centered, prospective cohort study of patients with extensive-stage SCLC who received chemo-immunotherapy as the first-line treatment between September 2019 and September 2020. The patient who can provide sufficient tumor tissue sample from the primary tumor were enrolled. We trained a classifier which predicts 365-day progression-free survival (PFS) by all three types of pathological images (hematoxylin and eosin, programmed death-ligand 1, and CD8_FoxP3) and patient information, and developed the patient information model, pathological image model, and combination model. We used the area under the curves (AUC) to evaluate the machine learning models. Results: Of 78 patients, the median age was 78 (interquartile range, 48-87), 65 patients (83%) were male, 67 patients (86%) had a performance status of 0 or 1, and three patients (3.8%) treated with steroid therapy. Among all patients, the median PFS and the 365-day PFS rates were 145 days and 10.3%. The mean AUC of these models was 0.789 (range, 0.571-0.982) in the patient information model, 0.782 (range, 0.750-0.911) in the pathological image model, and 0.868 (range, 0.786-0.929) in the combination model, respectively. According to the median precision model, the median PFS was longer for the high efficacy group than the low efficacy group (the patient information model; hazard ratio (HR) 0.468, 95% confidence intervals (CI) 0.287-0.762. the pathological image model; HR 0.334, 95%CI 0.117-0.628. the combination model; HR 0.353, 95%CI 0.195-0.637). Conclusions: Using machine learning by pathological images, we could predict the efficacy of immunotherapy in SCLC. This study demonstrated the potential of machine learning to help the biomarker development in SCLC by assessing TIME. Clinical trial information: UMIN000038064 .