Endometrial cancer (EC) is the most common gynecological cancer in high-income countries, with over 380,000 new cases per year globally. Its incidence is projected to rise 55% by 2030, driven by increasing rates of obesity, diabetes, and metabolic syndrome. Despite this growing threat, there is currently no validated screening test for endometrial cancer - a striking gap in women's healthcare.
The main symptom is abnormal uterine bleeding, but only 5-10% of women with this symptom actually have EC. Confirming a diagnosis requires either transvaginal ultrasound followed by endometrial biopsy or hysteroscopy - procedures that are invasive, expensive, and sometimes painful enough that women avoid or delay them. Up to 25% of biopsies in asymptomatic patients fail to provide enough tissue for diagnosis.
The survival stakes are high: at stage I, the 5-year survival rate exceeds 90%, but drops to just 20-26% at stage IV. A non-invasive, accurate, and affordable screening test could dramatically improve outcomes by catching EC earlier - this is the problem this study set out to solve.
Metabolomics is the large-scale study of small molecules (metabolites) in biological samples like blood or urine. Unlike genomics or proteomics, which look at genes or proteins, metabolomics captures the final output of all cellular processes - the actual chemical state of the body. This means it reflects not just genetic factors but also environmental exposures, diet, and disease state in real time.
Cancer cells are known to have dramatically altered metabolism. The most famous example is the Warburg effect, named after Nobel laureate Otto Warburg: cancer cells tend to prefer a less efficient but faster form of energy production (glycolysis) even when oxygen is available, allowing them to grow rapidly while reducing damaging reactive oxygen species (free radicals). These metabolic changes create distinctive molecular signatures in the blood that can potentially be used for diagnosis.
Earlier work by this research group identified a specific serum metabolic signature that could distinguish EC from healthy women and women with other uterine conditions. The current study aimed to validate this signature in a much larger and more clinically realistic population - women presenting for gynecological surgery, who represent the kinds of patients who would actually use such a screening test.
The study used two separate patient cohorts. The first (691 women) was split into training (90), testing (38), and validation (563) groups to build and refine classification models. Blood samples were processed using gas chromatography-mass spectrometry (GC-MS), a technique that separates and identifies hundreds of metabolites simultaneously, detecting up to 273 distinct chemical signals per sample.
Seven different machine learning algorithms were trained on the metabolic data: Naive Bayes, Generalized Linear Model, Fast Large Margin, Deep Learning, Decision Tree, Random Forest, and Partial Least Squares Discriminant Analysis (PLS-DA). Rather than relying on any single algorithm, the researchers combined all seven into an ensemble machine learning (EML) model. Each model casts a weighted vote (weighted by its accuracy and confidence), and the sum of votes determines whether a sample is classified as EC or not.
A second independent validation cohort of 871 women was then used as a true blind test - EC status was unknown to the model beforehand. This cohort included women with healthy tissue, 15 different non-cancer conditions (endometriosis, fibroids, ovarian cysts, etc.), five other cancers, and 126 with confirmed EC, creating a realistic and challenging test of the model's real-world performance.
GC-MS analysis detected 251 metabolite signals consistently present across samples. Statistical analysis identified 12 metabolites with significantly different levels between EC patients and healthy controls. Three metabolites were elevated in EC: glycerol, 3-hydroxybutyric acid, and stearic acid. Nine were reduced in EC patients: glycine, phenyl pyruvic acid, serine, valine, urea, oxyproline, phenylalanine, glyceraldehyde 3-phosphate, and gluconic acid.
The most diagnostically important metabolites included serine, glutamic acid, phenylalanine, and glyceraldehyde 3-phosphate - all of which were lower in EC patients. These reductions are biologically meaningful. Serine and glutamic acid are amino acids involved in cellular energy metabolism; their depletion suggests EC cells are diverting metabolic resources in ways that affect circulating amino acid levels. Glyceraldehyde 3-phosphate is a central glycolysis intermediate - its reduction aligns with the Warburg effect's rewiring of glucose metabolism in cancer.
Among individual models, PLS-DA achieved the best training accuracy at 96%. The Deep Learning and Random Forest models achieved 100% sensitivity (catching every EC case) but at the cost of lower specificity (more false positives). This is why combining all seven models was essential - the ensemble approach balances these tradeoffs to maximize overall accuracy.
When applied to the blind validation cohort of 871 women, the ensemble machine learning model achieved an error rate of less than 5% for EC identification. EC patients showed a distinctly high EML score (mean 363 out of ~500 possible), clearly separated from healthy controls (mean 40.6) and most other disease groups. The AUC (area under the ROC curve) was 0.974 on the first validation set - indicating excellent discrimination between EC and non-EC.
The model performed especially well against non-gynecological conditions: 0% error rate for breast cancer patients, meaning no breast cancer patients were incorrectly flagged as having EC. Most other cancer types and benign conditions also showed clearly lower EML scores than EC. Three groups did not show significantly different scores from EC: uterine sarcoma, vaginal cancer, and low-grade squamous intraepithelial lesion (L-SIL) - but these were small groups with known histological overlap with EC.
Importantly, patient age did not bias results. When the cohort was split into under-50 and over-50 age groups, EML scores within each diagnostic class were not significantly different, confirming that the metabolic signal reflects disease state rather than aging alone.
The metabolomics approach requires only a small blood sample and avoids uterine tissue sampling entirely. This matters clinically because the current gold standard - histological examination via hysteroscopy or dilation and curettage - carries risks of pain, bleeding, infection, and rare uterine perforation. Many women, particularly those who are elderly, obese, or simply anxious about the procedure, delay or decline invasive testing, contributing to later-stage diagnoses.
Compared to existing screening alternatives, the metabolomics model performs strongly. Transvaginal ultrasound (the most common initial investigation) uses endometrial thickness as a proxy for cancer risk, achieving only 80.5% sensitivity and 86.2% specificity in large-scale studies. The EML model substantially exceeds these benchmarks. A liquid biopsy approach using Pap smear DNA achieved 100% accuracy, but was tested on only 24 cases - too small to be clinically meaningful.
The authors envision this blood test as a first-line screening tool: women with elevated EML scores would then proceed to the more definitive (but invasive) tissue biopsy, while low-scoring women could be reassured without further workup. This triage approach would concentrate invasive procedures on higher-risk patients, reducing unnecessary procedures while improving early cancer detection across the broader population.
This study represents the largest validation of a metabolomics-based EC screening test to date, with 871 women tested across a clinically realistic range of conditions. The prospective enrollment at a single gynecological surgery center mimics real-world clinical conditions better than many prior studies that used only healthy controls and EC cases.
Key limitations include the single-center design (data came from one Italian hospital), which means geographic and demographic diversity is limited. The three patient groups with non-significantly different EML scores from EC (uterine sarcoma, vaginal cancer, L-SIL) need larger cohorts for definitive characterization. The technology (GC-MS) requires laboratory infrastructure and trained personnel, which could limit deployment in lower-resource settings.
The researchers call for further multicenter validation studies across different populations and healthcare settings before clinical implementation. If confirmed, this approach could be integrated into routine gynecological care - particularly for postmenopausal women with abnormal bleeding - offering a genuinely non-invasive, accurate, and patient-friendly first step in the EC diagnostic pathway.