Leakage-Free Random Forest Shows No Useful Cervical Cancer Prediction

In a 100-patient cervical cancer study, an exploratory Random Forest model showed no useful ability to identify patients whose circulating neutrophil gelatinase-associated lipocalin, or NGAL, declined after treatment. Using only baseline clinical and conventional laboratory information, the classifier achieved mean accuracy of 43.0% and mean balanced accuracy of 42.1% during stratified five-fold cross-validation. The outcome was not objective treatment response because the database contained no independent radiologic, pathological, or clinical response endpoint. Instead, the model attempted to classify an NGAL-decrease phenotype, defined as a post-treatment NGAL concentration lower than the pretreatment value.

The investigation included women treated for histologically confirmed cervical cancer at the Oncology Clinic of the University Clinical Center of Kosovo in Pristina between July 2023 and July 2024. Eligible patients were 25 to 75 years old, and the final analyzable cohort had a mean age of 55.9 years, with a range of 29 to 75. Thirty-eight patients were considered inoperable and 62 were considered operable. Histology identified 80 squamous or planocellular carcinomas, 18 adenocarcinomas, and two microinvasive carcinomas. Broad FIGO groups comprised 45 stage I cases, 23 stage II cases, 30 stage III cases, and two records with inadequate standardized information for reliable substage assignment. A separate control group included 30 healthy women, but the machine-learning analysis concerned the 100 patients with cervical cancer.

Treatment patterns varied across the cohort. Sixty-seven patients received concurrent chemoradiotherapy with brachytherapy, 13 received radiotherapy plus brachytherapy, 12 received brachytherapy alone, six received concurrent chemoradiotherapy without a documented brachytherapy entry, and two had no treatment recorded. For selected early-stage patients, brachytherapy was delivered in two 9 Gy fractions. Patients with non-metastatic locally advanced disease received intensity-modulated radiotherapy or volumetric-modulated arc therapy to 50.4 Gy in 28 fractions, with a 5.4 Gy boost in three fractions when warranted. Weekly concurrent cisplatin at 50 mg/m² was used when clinically suitable, although the database did not record the number of cycles completed by each patient.

Blood samples were collected three to five days before external-beam radiotherapy or chemoradiotherapy, or approximately three days before brachytherapy in early-stage cases, and again two to four weeks after treatment concluded. NGAL was measured in EDTA plasma with a human NGAL ELISA kit, while hematologic parameters were measured in whole blood and vitamin D was assessed with a separate laboratory system. The panel also included carcinoembryonic antigen, CA-125, ionized calcium, white blood cell count, red blood cell count, hemoglobin, hematocrit, platelet count, neutrophil-to-lymphocyte ratio, and monocyte-to-lymphocyte ratio. NGAL, also called lipocalin-2, is secreted principally by activated neutrophils and is also expressed by epithelial cells and other tissues. The paper describes it as an acute-phase protein involved in antibacterial immune processes.

To prevent target leakage, the researchers removed every NGAL-derived variable, including pretreatment NGAL, post-treatment NGAL, and the change between the two measurements. All other post-treatment variables were also excluded. Eligible predictors were age, histological grade, white blood cell count, red blood cell count, hemoglobin, hematocrit, platelet count, vitamin D, carcinoembryonic antigen, CA-125, ionized calcium, neutrophil-to-lymphocyte ratio, monocyte-to-lymphocyte ratio, operability, and broad FIGO stage. Missing numerical values were median-imputed, categorical variables were encoded, and class balancing was applied. The classifier used 1,000 trees, a minimum leaf size of two, and stratified five-fold cross-validation with a fixed random seed.

Cross-validated predictions correctly classified 43 of the 100 patients. Of the 47 patients without an NGAL decrease, 14 were correctly identified and 33 were misclassified; among the 53 with an NGAL decline, 29 were correctly identified and 24 were assigned to the wrong group. The no-decrease class had a precision of 0.37, recall of 0.30, and F1 score of 0.33. The NGAL-decrease class had a precision of 0.47, recall of 0.55, and F1 score of 0.50, while the macro-average F1 score was 0.42. Mean cross-validated accuracy was 43.0%, with a standard deviation of 7.6% and a fold range of 35.0% to 55.0%; mean balanced accuracy was 42.1%, with a standard deviation of 8.1%. These results did not provide useful discrimination or support clinical decision assistance.

White blood cell count ranked first in the model’s descriptive feature-importance list, followed by hematocrit, vitamin D, age, neutrophil-to-lymphocyte ratio, platelet count, hemoglobin, CA-125, ionized calcium, red blood cell count, monocyte-to-lymphocyte ratio, and carcinoembryonic antigen. These rankings reflect how the fitted trees divided the data and do not establish independent prognostic or predictive value. The same distinction applies to the biomarker trends: associations observed between NGAL and operability, FIGO stage, or histological grade are descriptive observations rather than evidence that NGAL independently predicts survival, response, or treatment benefit. Mechanistic ideas discussed by the investigators, including iron uptake by malignant cells, matrix metalloproteinase interactions, hypoxia-related signaling, and neutrophil-mediated inflammation, were not mechanisms demonstrated by this dataset.

Patients considered inoperable had a higher mean pretreatment NGAL concentration than those considered operable: 156.22 ng/mL with a standard deviation of 70.26, compared with 123.49 ng/mL with a standard deviation of 71.38. This exploratory comparison, assessed with Welch’s two-sample t-test, produced a p-value of 0.027. The authors also reported higher carcinoembryonic antigen and lower vitamin D among inoperable patients. These group differences do not demonstrate that the biomarkers independently predict prognosis, and the analysis did not establish a model with superior performance in either operability subgroup. Pre-treatment NGAL also differed across broad FIGO stage groups: mean concentrations were 103.9 ng/mL with a standard deviation of 38.5 in stage I, 168.9 ng/mL with a standard deviation of 71.6 in stage II, and 159.3 ng/mL with a standard deviation of 91.9 in stage III.

After treatment, mean NGAL values in broad stages I, II, and III were 98.4, 175.6, and 179.3 ng/mL, respectively. The lower stage I mean before treatment was followed by a small descriptive decrease, whereas the stage II and III means were higher after treatment. The investigators did not attribute these shifts to tumor reduction, nodal sterilization, therapy-induced inflammation, or residual illness because the study lacked the imaging, pathology, and clinical data needed to distinguish those explanations. Broad stage groups were used because nodal-location information was insufficient to assign every record consistently to the finer FIGO 2018 IIIC1 and IIIC2 categories. Histological grade showed a similar descriptive gradient: pretreatment NGAL averaged 93.6 ng/mL with a standard deviation of 40.5 for grade 1 tumors in nine patients, 125.1 ng/mL with a standard deviation of 55.2 for grade 2 tumors in 68, and 184.6 ng/mL with a standard deviation of 100.6 for grade 3 tumors in 23.

The study did not identify a validated NGAL cutoff for distinguishing early from late disease or defining treatment response. Earlier proposed thresholds were removed because the investigators had not performed a prespecified receiver-operating-characteristic analysis, reported sensitivity, specificity or confidence intervals, or validated a cutoff against an independent clinical endpoint. Several limitations constrain the findings. The cohort included only 100 patients from one institution, and strict exclusions for diabetes, autoimmune disorders, hypertension, hyperlipidemia, viral infections, and impaired renal or hepatic function limited the number of eligible cases and may limit generalizability. The modest sample also makes cross-validation estimates unstable. Most importantly, the target was defined from NGAL itself: the model classified whether NGAL decreased without an independent endpoint against which that classification could be tested. The authors therefore characterize the biomarker panel and machine-learning results as hypothesis-generating. Larger multicenter cohorts, complete FIGO 2018 staging, independently defined clinical outcomes, predetermined statistical plans, and external validation would be needed before NGAL or a combined biomarker model could be evaluated for clinical use.

Subject of Research: Exploratory evaluation of NGAL and conventional biomarkers in cervical cancer using a leakage-free Random Forest classifier

Article Title: NGAL and traditional biomarkers in cervical cancer: a comparative evaluation using random forest

Article References: Raci, B., Stojkovska, S., Hodolli, G., Klisarovska, V., Dimitrov, G., Jusufi, A. K., Sejdiu, H., Sermaxhaj, B., Rizaj, D., Kurtishi, I., & Kurshumliu, M. (2026). NGAL and traditional biomarkers in cervical cancer: a comparative evaluation using random forest. Frontiers in Oncology, 16, Article 1933456. https://doi.org/10.3389/fonc.2026.1933456

Image Credits: AI Generated

DOI: 10.3389/fonc.2026.1933456

Keywords: cervical cancer, NGAL, lipocalin-2, Random Forest, biomarkers, machine learning, cross-validation, gynecologic oncology, traditional, cervical, cancer, comparative