Endometrial cancer (EC) is the most common gynecologic malignancy worldwide, with an estimated 417,000 new cases and 97,000 deaths per year. Its incidence is rising, particularly in developed countries, largely due to increasing rates of obesity, physical inactivity, and aging populations.
Current diagnostic tools have significant limitations. The commonly used blood test marker CA-125 has a diagnostic AUC of only 0.610-0.684 - barely better than chance - making it unreliable for early detection. Ultrasound-based distinction is also challenging, and definitive diagnosis currently requires invasive tissue biopsy.
This study explores metabolomics as an alternative approach. Metabolomics is the large-scale study of small molecules called metabolites in blood or other biological samples. Because cancer cells dramatically reprogram their metabolism - consuming more glucose, altering lipid synthesis, and disrupting amino acid pathways - these metabolic changes can potentially serve as detectable early-warning signals in blood tests.
The study analyzed serum samples from 191 endometrial cancer cases and 204 healthy controls. Serum metabolic fingerprints were measured using PELDI-MS (particle-enhanced laser desorption/ionization mass spectrometry), a technique that identifies molecules by their mass-to-charge ratio (m/z). A total of 272 metabolite features were captured per sample.
Before applying machine learning, the researchers used LASSO regression (a statistical technique that penalizes complexity) to select the most informative metabolites and reduce dimensionality. This step is critical with high-dimensional biological data to prevent models from fitting to noise rather than true cancer signals.
Five machine learning algorithms were then trained and compared: LightGBM, XGBoost, Random Forest, AdaBoost, and BaggedCART. All are ensemble methods - they combine many smaller decision trees into a single powerful predictor. Models were validated using 100 iterations of a 4:1 train-test split to ensure reliable performance estimates.
Among all five models tested, LightGBM (Light Gradient Boosting Machine) delivered the strongest performance, achieving an accuracy of 0.888 and an AUC of 0.955. This means the model correctly classified 88.8% of patients and had a 95.5% probability of ranking a random cancer patient above a random healthy individual - a clinically meaningful result.
LightGBM also showed excellent probability calibration with the lowest Brier score (0.0881), meaning its confidence estimates are well-aligned with actual outcomes - a critical property for clinical decision support. XGBoost performed nearly as well (AUC 0.951), while Random Forest (AUC 0.950) was also competitive. AdaBoost and BaggedCART lagged behind.
Separately, statistical analysis of all 272 metabolites found 25 with significantly different levels between EC and non-EC groups (FDR-adjusted p-values below 0.05). The three metabolites with the largest effect sizes were m/z=219.125 (effect size 1.516 - very large), m/z=672.6961 (effect size 0.913), and m/z=203.1564 (effect size 0.839), indicating the strongest biological separation between groups.
SHAP (SHapley Additive exPlanations) is a technique borrowed from game theory that assigns each feature a score representing its contribution to any individual model prediction. It transforms the LightGBM "black box" into an interpretable tool that shows exactly which metabolites are driving each diagnosis.
The SHAP analysis identified m/z=219.125 as the single most influential biomarker - having the strongest effect on whether the model predicted cancer or not. The metabolites m/z=672.6961 and m/z=127.0769 were ranked second and third. Notably, m/z=127.0769 did not appear in the top statistical performers but ranked highly in SHAP, showing that machine learning can detect biomarkers that interact with other features in non-obvious ways.
By contrast, m/z=106.0371 had a p-value of 0.06 in standard statistics but minimal SHAP influence - illustrating that statistical significance and ML-based importance do not always align. Using both approaches together produces a more biologically relevant biomarker panel than either method alone.
The top biomarkers all have plausible biological connections to endometrial cancer. m/z=219.125 is thought to correspond to a lipid or lipid-derived compound involved in membrane formation for rapidly dividing cancer cells - consistent with the known upregulation of lipid synthesis in tumors. Its very large effect size (1.516) makes it a strong candidate for further validation.
m/z=672.6961 may correspond to glycosphingolipids or nucleotide derivatives linked to altered steroid hormone pathways and tumor development. Endometrial cancer is strongly driven by estrogen excess (from obesity and hormonal imbalance), and disruptions to steroid metabolism are expected to show up in the metabolome.
m/z=127.0769 is associated with amino acid derivatives of tryptophan and the kynurenine pathway - a biological process known to help cancer cells evade immune system attack. This pathway is an active area of cancer immunotherapy research, suggesting this metabolite may have both diagnostic and therapeutic relevance.
The study's most significant clinical implication is the potential for a blood-based, non-invasive screening test for endometrial cancer. Currently, definitive EC diagnosis requires a tissue biopsy, which is invasive, uncomfortable, and not suitable for routine screening. A serum metabolomics test could potentially flag high-risk patients before symptoms develop.
The researchers propose that m/z=219.125 in particular could serve as a blood marker for monitoring treatment response and cancer recurrence after initial therapy. This would enable more frequent, less invasive follow-up compared to repeat biopsies or imaging.
Limitations include the relatively modest sample size (395 participants), the absence of important confounders such as BMI and menopausal status in the data, and reliance on a single mass spectrometry platform. Because obesity itself dramatically alters the metabolome, some signals may reflect obesity rather than cancer specifically. Larger, matched cohort studies are needed before clinical translation.