Article Text

Download PDFPDF

EP169/#411  Machine learning method for differential diagnosis and prognosis prediction for early-stage uterine sarcoma using preoperative blood biomarker and age
  1. Yuichi Shoburu1,
  2. Nozomu Yanaihara1,
  3. Junya Tabata1,
  4. Ryoga Nishimura2,3,
  5. Miwako Shimazaki1,
  6. Kazuhiko Oka1,
  7. Yutaro Kubonoya1,
  8. Ritsuko Kobayashi-Ogasawara1,
  9. Rie Honda1,
  10. Atsuko Yamada4,
  11. Motoaki Saito1,
  12. Kyosuke Yamada1,
  13. Hirokuni Takano1,
  14. Yasuhisa Terao4,
  15. Eiryo Kawakami3,5 and
  16. Aikou Okamoto1
  1. 1The Jikei University School of Medicine, Obstetrics and Gynecology, Tokyo, Japan
  2. 2Yamaguchi University, Faculty of Medicine and Health Sciences, Yamaguchi, Japan
  3. 3RIKEN, Advanced Data Science Project (adsp), Riken Information Randd and Strategy Headquarters, Saitama, Japan
  4. 4Juntendo University Faculty of Medicine, Department of Obstetrics and Gynecology, Tokyo, Japan
  5. 5Chiba University Graduate School of Medicine, Department of Artificial Intelligence Medicine, Chiba, Japan


Introduction Preoperative differential diagnosis of clinical stage I uterine sarcoma (US) is essential for surgical intervention. Many studies have been done using CT or MRI imaging for machine learning prediction models but not with blood biomarkers. We aimed to develop a new model for diagnosis and prognosis prediction in the US using preoperative blood biomarkers and patient age.

Methods Overall, 143 US patients and 210 benign uterine myoma (UM) patients were randomly assigned to the ‘training and test’ cohort. 78(55%) cases were on clinical stage I. 30 preoperative peripheral blood parameters and patient’s age was surveyed. The Random Forest (RF) classifier was used to construct an algorithm. The accuracy, the area under the receiver operating characteristic curve (AUC), and the variable importance were calculated in the test cohort. The Ethics Committee approved this study.

Results The accuracy and AUC values for segregating stage I US from UM were 87% and 0.89, respectively. Variable important parameters for this classifier included age, CRP, and Hematocrit. Additionally, they were 85% and 0.95 in leiomyosarcoma, and 92% and 0.81 in ESS, respectively. Furthermore, unsupervised clustering analysis based on RF showed significant differences in two clusters in clinical stage I US with a median progression-free survival of 47 (3–115) vs. 13 (1–93) months (P < 0.001).

Conclusion/Implications The RF approach using common blood biomarkers and patient age can differentiate its malignancy and prognosis of US patients before primary intervention. This predictive model may provide a clinically useful approach to preoperative diagnosis distinct from conventional imaging techniques.

Statistics from

Request Permissions

If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.