Background: Breast cancer (BC) has contributed to the most cancer-related mortalities among females in 2020. Early detection using breast imaging methods, such as mammography and ultrasound, has significantly improved patients' survival. However, these early-screening methods suffer from high false-positive rates, resulting in many unnecessary biopsies which add to patients' discomfort. Cell-free DNA (cfDNA) fragmentomics assays have recently illustrated prominent abilities for detecting various cancer. In this multi-center prospective cohort study, we aim to develop a non-invasive liquid biopsy assay utilizing cfDNA fragmentomics and machine learning for detecting early-stage BC patients. Methods: We recruited 402 female patients (234 with BC and 168 with benign nodules [BN]) from the Yantai Yuhuangding Hospital (training cohort, N = 193, 91 BC and 102 BN) and the Cancer Hospital of the Chinese Academy of Medical Sciences (independent validation cohort, N = 209, 143 BC and 66 BN). Surgical or core needle biopsies were performed for all 402 patients following positive breast imaging results by mammography and ultrasound. Each patient's malignant or benign status was then pathologically confirmed using the tissue specimen. Women with negative biopsy results were excluded from breast cancer through a 6-month follow-up. The fragmentomics profiles, which were generated using plasma cfDNA WGS data (~5X), were used by our machine-learning algorithm. An automated machine learning approach was deployed to optimize base models before being ensembled into the final predictive model. Results: Our predictive model showed excellent performance in discriminating BCs from BNs, reaching an Area Under the Curve (AUC) of 0.816 and 0.811 in the training cohort (5-fold cross-validation) and the independent validation cohort, respectively. It achieved 74.1% sensitivity (95% CI: 66.1-81.1%) at 81.8% specificity (95% CI: 70.4-90.2%) in the validation cohort. In a potential clinical setting, our model could detect 53.0% (95% CI: 40.3-65.4%) false positives from the breast imaging methods at a targeted 90% sensitivity (89.5%, 95% CI:83.3–94.0%), which could theistically reduce more than half of the unnecessary biopsies. Moreover, our model illustrated great predictive power (AUC = 0.830) among small nodules (≤ 1cm), which were difficult to distinguish by the traditional imaging methods (AUC = 0.690). Finally, our model maintained its predictive power during a down-sampling process, even using 2X WGS data in the independent validation cohort (mean AUC: 0.789, 0.781-0.798). Conclusions: Our non-invasive liquid biopsy assay, which utilized low-depth cfDNA fragmentomics profiling and machine learning, can distinguish malignant nodules from benign nodules. The highlight of our model is its ability to prevent unnecessary biopsies among patients with false positive breast imaging results.