Lymphovascular space invasion (LVSI) occurs when tumor cells penetrate the walls of lymphatic vessels or small blood vessels within and around a tumor. Once inside these vessels, cancer cells can travel through the lymphatic system and bloodstream to reach distant organs. In endometrial cancer, LVSI is one of the most important indicators that a tumor has the potential to spread beyond the uterus.
LVSI is officially recognized by the European Society for Medical Oncology (ESMO) as a key factor in risk stratification for early-stage endometrial cancer. Patients whose tumors show LVSI have significantly higher rates of lymph node metastasis, which dramatically worsens prognosis. The 5-year survival of early-stage EC generally exceeds 80%, but drops sharply when lymph node involvement is present.
The critical practical question is whether to perform lymphadenectomy - surgical removal of pelvic and para-aortic lymph nodes - during hysterectomy. This procedure helps stage the disease and may prolong survival in high-risk patients, but carries real harms including surgical complications and lymphedema. For low-risk patients, the added benefit does not justify these risks.
The fundamental problem is timing: LVSI is currently diagnosed only after surgery, through examination of the removed tissue under a microscope. There is no reliable preoperative way to determine whether LVSI is present. This means surgeons must decide whether to perform lymphadenectomy before knowing whether the patient actually needs it. A validated preoperative prediction model could fill this critical gap.
The study enrolled 312 endometrial cancer patients who received surgical treatment at Xuzhou Medical University Affiliated Hospital of Lianyungang between May 2017 and March 2022. All diagnoses were confirmed by postoperative pathology. Patients were split 70/30 into a training group (219 patients) and a validation group (93 patients) to build and then independently test the prediction model.
Of the 312 patients, 77 (25%) had LVSI on final pathology (56 in training, 21 in validation). This reflects real-world LVSI prevalence in EC and meant the model needed to reliably identify a minority of patients with distinctly worse prognosis among a larger group who did not have LVSI.
Two complementary machine learning approaches were used. Multivariable logistic regression is a traditional statistical method that directly identifies independent risk factors by measuring how much each variable contributes to LVSI risk while controlling for other factors. It is interpretable and widely used in clinical prediction modeling.
LASSO (Least Absolute Shrinkage and Selection Operator) regression is a more advanced technique that automatically selects the most relevant features from a larger pool by penalizing variables that add little predictive value. LASSO can handle many potential predictors simultaneously and is less prone to overfitting than standard logistic regression, making it particularly useful when dealing with many clinical and laboratory variables.
Multivariable logistic regression identified six independent risk factors for LVSI in endometrial cancer. Two were anatomical tumor characteristics: deep myometrial infiltration (tumor penetrating more than half the thickness of the uterine muscle wall, OR = 17.9) and cervical stromal invasion (tumor extending into the cervical connective tissue, OR = 8.0). Both indicate locally advanced disease with higher likelihood of lymphatic spread.
Four blood-based laboratory markers were also independently predictive. Higher lymphocyte count (LYM) was associated with LVSI (OR = 0.16, meaning higher LYM was protective - LVSI patients actually had higher LYM, reflecting immune system activation in response to more aggressive tumors). Higher monocyte count (MONO) was similarly associated with LVSI risk.
Higher albumin (ALB) was paradoxically protective against LVSI (OR = 0.18), which is consistent with established literature showing that better nutritional status and systemic health correlate with less aggressive tumor behavior. Higher fibrinogen (FIB) strongly predicted LVSI (OR = 11.9). Fibrinogen is a blood clotting protein; elevated levels are associated with cancer-related inflammation and thrombosis, which directly facilitate tumor cell survival during vascular invasion.
The biological rationale for these blood markers is clear: LVSI involves tumor cells entering blood vessels and lymphatics, binding to platelets and white blood cells to form tumor thrombi. Changes in coagulation proteins (fibrinogen), immune cell composition (lymphocytes, monocytes), and nutritional proteins (albumin) all reflect the systemic conditions that either facilitate or resist this process.
While logistic regression identified 6 independent variables, LASSO regression identified 19 characteristic factors for LVSI prediction. These included all six logistic regression factors plus additional variables: patient age, menstrual and reproductive history (age at first period, menarche, gravidity, parity), history of diabetes, tumor diameter, FIGO staging, tumor grade, adnexal metastasis, neutrophil count (NEUT), and platelet-to-lymphocyte ratio (PLR).
The reason LASSO identified more factors is that it uses coefficient regularization - a mathematical technique that systematically reduces the contribution of weakly predictive variables toward zero rather than excluding them entirely. This allows LASSO to capture subtle combined effects of variables that individually appear insignificant but collectively improve model accuracy.
The LASSO model outperformed logistic regression in both the training group and the independent validation group, as confirmed by the Delong test (a statistical method for comparing AUCs). Importantly, the training and validation groups were well-matched in baseline characteristics (no significant demographic differences, P greater than 0.05), confirming the validation group was a fair test of generalizability.
The logistic regression model achieved AUC values of 0.946 (training) and 0.850 (validation), already indicating strong predictive performance. The LASSO model achieved significantly higher AUC values than logistic regression at both stages, demonstrating that capturing the multifactorial complexity of LVSI requires a richer feature set than logistic regression alone can effectively use.
The strong predictive power of myometrial infiltration depth makes biological sense. As tumors invade deeper into the uterine muscle wall, they encounter richer lymphatic and vascular networks. Deep invasion means tumor cells have physically reached the tissue layers where lymphatic vessels are most abundant, making LVSI not just possible but highly probable.
Fibrinogen's strong association with LVSI (OR = 11.9) reflects a direct pathophysiological link. Cancer cells that enter blood vessels or lymphatics can bind to fibrinogen and platelets to form protective clusters called tumor thrombi. These clusters shield individual tumor cells from immune attack during circulation and help them adhere to vessel walls at metastatic sites. High fibrinogen thus indicates a coagulation environment that actively facilitates cancer spread.
The nutritional and immune blood markers - albumin, lymphocyte count, monocyte count - reflect the systemic context in which tumor invasion occurs. Low albumin indicates a state of nutritional depletion often associated with advanced or aggressive cancer. Altered immune cell ratios suggest that the tumor microenvironment has begun to influence systemic immune composition, a hallmark of tumors with metastatic potential.
The addition of reproductive and obstetric history (gravidity, parity, menarche age) by LASSO reveals that hormonal exposure history contributes to LVSI risk, consistent with endometrial cancer's relationship with cumulative estrogen exposure over a lifetime. These factors would be invisible to a model focused only on current tumor characteristics, demonstrating the value of comprehensive patient history in preoperative risk prediction.
The primary clinical application of this model is guiding the decision to perform lymphadenectomy at the time of initial hysterectomy. Currently, many patients undergo lymph node removal without definitive knowledge of whether it is needed. A validated preoperative risk score combining myometrial depth (measured by MRI), cervical involvement, and blood tests (LYM, MONO, ALB, FIB) could identify high-risk patients before the operating room.
For patients identified as high LVSI risk, surgeons could plan a more comprehensive staging surgery from the outset, including systematic pelvic and para-aortic lymphadenectomy. For patients identified as low risk, lymphadenectomy could be safely omitted, avoiding surgical complications and lymphedema. This represents a clinically meaningful improvement over current practice where these decisions rely on subjective assessment of available preoperative data.
Because this model uses data already collected in routine preoperative workups - standard blood tests and pelvic MRI - its implementation would not require additional invasive procedures. The blood markers (LYM, MONO, ALB, FIB) are included in a routine complete blood count with metabolic panel, making the model practical and low-cost to apply.
The study's single-center retrospective design is a recognized limitation. Selection bias and the relatively small sample size may affect generalizability to different patient populations or institutions with different patient demographics. Future prospective, multi-center studies with larger sample sizes are needed to validate model performance before routine clinical adoption and to potentially incorporate additional molecular markers for further refinement.