EUVIMEDEUVIMED entdecken
LeserbriefOpen Access · CC BY-NC 3.0

Methodological and Translational Considerations for Blood-Based Machine-Learning Prescreening of Esophageal Cancer [Letter]

Zhipeng Fan iD, Lei Li

International Journal of General Medicine · 2026-08 · Volume 19 · 1-2

DOI: 10.2147/ijgm.s640300 ↗

Kurz gefasst

Leserbrief in der Zeitschrift International Journal of General Medicine, 2026-08. Verfasst von Zhipeng Fan und 1 weiteren Autor:innen. Beteiligte Einrichtung: Department of Oncology, Wuhan Asia General Hospital Affiliated to Wuhan University of Science and Technology, Wuhan, Hubei, 430056, People's Republic…. Der Volltext ist frei zugänglich (Open Access, CC BY-NC 3.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

We read with great interest the study by Wang et al, which developed an interpretable random forest model for esophageal cancer prescreening from routine peripheral blood indices. 1 The use of low-cost laboratory data, comparison of seven algorithms, and SHAP-based interpretation address a clinically relevant problem, particularly where endoscopic capacity is limited.The reported validation AUC of 0.973 and Brier score of 0.059 are encouraging.While acknowledging these strengths, we believe four issues merit consideration before the model is advanced toward clinical implementation.First, the validation architecture requires additional clarification regarding the machine-learning development process.The authors reported that LASSO regression was performed within the training set before 10-fold cross-validation and hyperparameter tuning.However, if feature selection was conducted only once before cross-validation rather than repeated within each resampling fold, the internal performance estimation may remain optimistic.In addition, the holdout set appears to have been used both for comparing seven algorithms and for identifying the final model.The AUC differences from XGBoost and LightGBM were only 0.013 and 0.018, respectively, without paired uncertainty estimates.A nested pipeline with selection and tuning confined to inner folds, followed by one locked evaluation in an untouched dataset, would provide a less biased comparison. 2 Second, interpretation of predictive performance requires careful consideration of the study design.The reported case fraction of esophageal cancer was 43.6% in this case-control cohort, which is determined by sampling rather than reflecting population prevalence.Therefore, positive predictive value, negative predictive value, calibration performance, Brier score, and decisioncurve estimates from this dataset cannot be directly extrapolated to screening populations without appropriate prevalence adjustment, weighting, or recalibration.3 The fixed 0.50 threshold also missed 7 of 59 cancers in the validation cohort.That operating point may be too restrictive for a prescreening tool when the consequence of a missed early cancer is weighed against additional endoscopies.Future evaluations should report calibration intercept and slope, recalibrate predictions to the intended population, present precision-recall performance after appropriate prevalence adjustment, and select thresholds using explicit clinical harms, endoscopic capacity, and cost-effectiveness assumptions.Third, the biological interpretation of model predictors deserves further refinement.Although the authors appropriately discussed the limitations of case-control design and spectrum bias, additional attention should be given to whether laboratory abnormalities represent tumor-specific signals or systemic consequences of disease.Samples were obtained before treatment, but the study does not establish that they preceded symptoms or the diagnostic workup.The algorithm therefore distinguishes prevalent cancer from controls rather than predicting future cancer risk.Creatinine, the dominant feature, was lower in cases than controls, with medians of 51.64 and 61.13 µmol/L, and lower values increased predicted risk.Because serum creatinine is strongly influenced by muscle mass, this signal may partly capture dysphagiarelated nutritional decline or cancer-associated body-composition change rather than tumor-specific susceptibility.4,5 The

Autor:innen und Institutionen

  • Department of Oncology, Wuhan Asia General Hospital Affiliated to Wuhan University of Science and Technology, Wuhan, Hubei, 430056, People's Republic of China.

ORCID-Profile: Zhipeng Fan

Zitieren

APA

Fan, Z., & Li, L. (2026). Methodological and Translational Considerations for Blood-Based Machine-Learning Prescreening of Esophageal Cancer [Letter]. International Journal of General Medicine, Volume 19, 1-2. https://doi.org/10.2147/ijgm.s640300

Vancouver

Fan Z, Li L. Methodological and Translational Considerations for Blood-Based Machine-Learning Prescreening of Esophageal Cancer [Letter]. International Journal of General Medicine. 2026;Volume 19:1-2. doi:10.2147/ijgm.s640300

BibTeX
@article{Fan2026,
  author = {Fan, Zhipeng and Li, Lei},
  title = {Methodological and Translational Considerations for Blood-Based Machine-Learning Prescreening of Esophageal Cancer [Letter]},
  journal = {International Journal of General Medicine},
  year = {2026},
  volume = {Volume 19},
  pages = {1--2},
  doi = {10.2147/ijgm.s640300}
}
Kostenlos starten →