Background: Liquid biopsies based on circulating tumor DNA (ctDNA) analysis are being investigated for detection of residual disease and recurrence. Conclusive evidence for utility of ctDNA in early-stage non-small cell lung cancer (NSCLC) is awaited. Due to low ctDNA levels in early-stage disease or post-treatment, effective methods require high analytical sensitivity to detect mutant allele fractions (MAF) below 0.01%. Methods: We analysed 363 plasma samples from 88 patients with NSCLC recruited to the LUng cancer CIrculating tumour DNA (LUCID) study, with disease stage I (49%), II (28%) and III (23%). 62% were adenocarcinomas. Plasma was collected before and after treatment, and at 3, 6 and 9 months after surgery (N = 69) or chemoradiotherapy (N = 19). Additional plasma was collected at disease relapse for 17 patients. Median follow-up was 3 years, and 40 patients progressed or died of any cause. We employed the RaDaR™ assay, a highly sensitive personalized assay using deep sequencing of up to 48 tumor-specific variants. Variants identified by tumor exome analysis were tested by deep sequencing of tumor tissue and buffy coat DNA to verify somatic mutations and exclude clonal hematopoiesis. The RaDaR assay demonstrated 90% sensitivity at 0.001% MAF in analytical validation studies. Results: ctDNA was detected in 26% of samples, at median MAF of 0.047% (range: 0.0007% to > 2%), and prior to treatment in 87%, 77% and 24% for disease stage III, II and I respectively. For 62 patients, plasma was collected at a landmark timepoint, between 2 weeks and 4 months after initial treatment. ctDNA detection at the landmark timepoint was strongly predictive of clinical disease relapse, with Hazard Ratio of 20.7 (CI: 7.7-55.5, p-value < 0.0001). All 11 cases with ctDNA detected at landmark had disease progression, a median of 121 days after detection, and these included all 8 patients that relapsed within 300 days of treatment. Across 27 patients whose disease progressed during the study, ctDNA was detected at any timepoint post-treatment in 17 cases, with a median lead time of 203 days, and up to 741 days prior to clinical progression. ctDNA was detected post-treatment, in 13 of the 15 patients that progressed and had ctDNA detected prior to treatment. Conclusions: Our results support an emerging paradigm shift, by demonstrating that liquid biopsies can reliably detect recurrence of NSCLC at a preclinical stage, many months before clinical progression, thereby offering the opportunity for earlier therapeutic intervention. Clinical trial information: NCT04153526