Endometrial cancer - cancer of the uterine lining - is the most common gynecologic cancer in developed countries. When caught early at stage I, the five-year survival rate is over 80%. By stages III and IV, that figure drops to below 42%. Early detection is therefore critical, but current methods fall short: transvaginal ultrasound is widely used but produces many false positives requiring additional testing, while hysteroscopy is accurate but invasive.
A 2024 study published in BMC Medicine presents a promising new approach: a liquid biopsy based on analyzing tiny fragments of DNA that tumor cells shed into the bloodstream. This DNA, called cell-free DNA or cfDNA, carries patterns that differ between cancer patients and healthy individuals. By using machine learning to detect these patterns from a simple blood draw, researchers developed a test capable of detecting endometrial cancer with remarkable accuracy.
The key innovation is not just detecting DNA mutations, but rather analyzing the fragmentomics - the size, distribution, and structural patterns of cfDNA fragments. These patterns reflect how DNA is packaged in cells and are altered in cancer, providing a rich signal that the machine learning model can use to distinguish cancer from healthy tissue.
The study enrolled 111 women with confirmed endometrial cancer and 111 healthy women at Fujian Cancer Hospital in China. Blood samples were collected before any treatment, and cfDNA was extracted from the plasma (the liquid portion of blood). Low-coverage whole-genome sequencing was then performed at a relatively low depth of around 5x - meaning the entire genome was read roughly five times - which keeps costs manageable while still capturing the needed patterns.
From this sequencing data, researchers extracted three types of features. Copy number variation (CNV) measures gains and losses of DNA segments across the genome - cancers frequently show abnormal patterns here. Fragment size distribution (FSD) captures the sizes of cfDNA fragments across all chromosome arms in 5-base-pair steps, because cancer-derived fragments tend to have different size distributions than normal cfDNA. Nucleosome footprint (NF) analysis examines which regions of the genome are accessible or wrapped around protein spools called nucleosomes - a pattern that reflects gene activity and is disrupted in cancer cells.
A two-tier machine learning framework was built. First, five different algorithms (including Gradient Boosting, Random Forest, and XGBoost) were trained separately on each feature type. Then the best-performing models were combined into an ensemble model that averaged predictions across all three feature types. This approach is more robust than any single model because it captures complementary information from different dimensions of the cfDNA data.
The ensemble model demonstrated outstanding performance. In the training cohort (133 participants), it achieved an AUC of 0.991 - meaning it was nearly perfect at ranking cancer patients above healthy individuals. In an independent validation cohort (89 participants collected at a different time period), the AUC was 0.994. The model maintained a specificity of 95.5% (low false positive rate) and sensitivity of 97.8-98.5% (high true detection rate) across both cohorts.
Crucially, the test performed well even at stage I disease, when tumors are confined to the uterus and most treatable. Stage I sensitivity was 96.4%, and sensitivity reached 100% at later stages. This is especially significant because existing methods often struggle with early-stage detection. The model also correctly classified 91.5% of patients with uterine fibroids (non-cancerous growths) as not having cancer, showing it can distinguish cancer from a common benign condition that can mimic endometrial thickening.
The researchers also confirmed that the fragmentomics patterns were biologically meaningful, not just statistical artifacts. Fragment size abnormalities were concentrated in chromosomes 1, 6, 10, 11, and 17 - precisely the chromosomes most frequently mutated in endometrial cancer according to The Cancer Genome Atlas (TCGA) database. Pathway analysis of the nucleosome footprint features pointed to cancer-related processes including transcriptional dysregulation and immune system changes, consistent with the known biology of endometrial cancer.
For any blood test to be practically useful in clinical settings, it must be reliable under the varied conditions of real-world sample handling. The researchers systematically tested how transportation delays, storage temperature, freezing duration, and patient physiological state (before and after meals, before and after exercise) affected results.
Results were remarkably stable. Samples transported over 24, 48, or 72 hours - whether kept at room temperature or on ice - gave the same classification as samples processed within 2 hours. Samples frozen for up to one month also produced consistent results. The only exception was samples frozen for six months, which showed artificially elevated risk scores, suggesting a practical limit on storage time before testing. Repeated blood draws from the same individuals gave identical results, confirming reproducibility.
The model also maintained performance at lower sequencing depths below the target 5x coverage, which is important because sequencing depth can vary in practice. This robustness means the test is not dependent on laboratory perfection to function correctly, making it more suitable for broader clinical deployment across different healthcare settings.
To estimate the real-world impact of this test, the researchers used a validated mathematical modeling approach. Under current standard care in China, only about 21% of endometrial cancer patients are diagnosed at stage I. With the cfDNA test applied to at-risk women, that figure could potentially rise to 99% stage I detection. By catching more cancers early - when treatment is most effective and least invasive - this shift in stage distribution could raise the five-year survival rate from 84% to 95%.
These are projected figures based on modeling rather than direct clinical outcomes, and the researchers are appropriately cautious about them. The study sample of 222 participants is relatively small, and the findings need validation in larger, more diverse, multi-site cohorts before the test can be considered for widespread clinical use. The current analysis was also conducted within a single institution in China, so results may not generalize immediately to other populations or healthcare systems.
Nonetheless, the study represents a meaningful proof-of-concept. A low-cost, non-invasive blood test that can detect endometrial cancer before symptoms appear - with accuracy superior to both ultrasound and circulating tumor DNA methods - would represent a significant advance in women's cancer screening. Prospective clinical trials are the clear next step to establish whether the theoretical survival benefit translates into real patient outcomes.
This study demonstrates that cfDNA fragmentomics - the analysis of how tumor DNA is broken up and distributed in the bloodstream - can serve as a powerful signal for detecting endometrial cancer. The three-feature ensemble approach (combining copy number variation, fragment size distribution, and nucleosome footprints) substantially outperformed any single feature alone, and the resulting classifier achieved near-perfect performance in an independent validation cohort.
The work adds to a growing body of research applying cfDNA fragmentomics to other cancers, including lung, liver, and colorectal cancers. Endometrial cancer has been relatively underrepresented in this literature, making this contribution particularly valuable. The biological validity of the identified signals - anchored in known cancer-associated chromosomal abnormalities and pathways - strengthens confidence that the model is capturing genuine disease biology rather than spurious patterns.
The path to clinical implementation will require larger multi-center trials, cost-effectiveness analysis, and integration with existing screening workflows. However, the combination of high accuracy, robustness under practical conditions, and potential to dramatically shift detection to earlier stages makes cfDNA fragmentomics a compelling candidate for future endometrial cancer screening programs.