Background: Although small cell lung cancer (SCLC) is managed as a single cancer type, new evidence supports that subtypes (high vs low neuroendocrine) of SCLC acquire diverse transcriptional and epigenetic states. Furthermore, distinct SCLC subtypes acquire unique therapeutic vulnerabilities, such as the low-neuroendocrine highly inflamed SCLC subtype that has been linked to immunotherapy sensitivity. We have previously shown that DELFI (DNA evaluation of fragments for early interception) can be used to non-invasively distinguish SCLC from non-small cell lung cancer (NSCLC). Here, we report a preliminary analysis of using the DELFI approach for distinguishing among SCLC subtypes. Methods: Circulating cell-free DNA (cfDNA) was isolated from plasma samples of patients diagnosed with relapsed SCLC and treated with durvalumab plus olaparib in a phase II trial (NCT02484404). Over 200 cfDNA libraries were prepared for whole genome sequencing in batches with internal controls. To infer tumor gene expression profiles in cfDNA, we investigated genome-wide signals of tissue-specific transcription factors differentially regulated in SCLC, and applied a novel DELFI-based approach to inform SCLC molecular subtypes. Clinical information, patient-matched tissue transcriptome (RNA-seq), and chromatin accessibility data (ATAC-seq) were examined in orthogonal analyses. Results: The DELFI classifier detected SCLC cases (N = 47) with a median DELFI score of 1.0 (95% CI 0.99-1), significantly higher than previously reported scores for other lung cancer subtypes. Comparison of DELFI fragmentome signals to publicly available tumor transcriptomes shows subtype-level concordance (r = 0.78, p < 0.001, Pearson), particularly in pre-treatment SCLC cases separating high- vs low- neuroendocrine subtypes. High-neuroendocrine SCLCs exhibited a decrease in aggregate cfDNA fragment coverage at ASCL1 transcription factor binding sites relative to low-neuroendocrine SCLCs (r = 0.90, p < 0.001, Pearson). Additionally, low-neuroendocrine SCLCs revealed a significant decrease in aggregate cfDNA fragment coverage at genomic binding sites regulated by hematopoietic transcription factors, reflecting the inflamed phenotype of this SCLC subtype. Integration of these approaches provided a cfDNA fragmentation-based machine learning model that distinguished SCLC subtypes with high performance. Conclusions: Genome-wide cfDNA fragmentome analyses can differentiate high- and low-neuroendocrine SCLC subtypes. Given the challenges for performing SCLC biopsies in a clinical setting, we believe this approach could be a viable method of subtyping SCLC in a non-invasive manner.