Article Text

Download PDFPDF

2022-RA-1457-ESGO Gynecological cancer detection using fourier-transformed infra-red spectroscopy in urine samples: potential and accuracy of machine learning processing
  1. Francesco Vigo,
  2. Alessandra Tozzi,
  3. Muriel Disler,
  4. Vasileios Kavvadias,
  5. Andre Fedier,
  6. Viola Heinzelmann-Schwarz and
  7. Tilemachos Kavvadias
  1. Gynecology and Gynecologic Oncology, University Hospital of Basel, Basel, Switzerland


Introduction/Background Making an early diagnosis of cancer is the challenge that modern medicine has been setting for several decades. In gynecology, no effective screening has yet been found and approved for endometrial and ovarian cancer, and, despite cervical cytology testing, cervical cancer remains a leading cause of morbidity and mortality among gynecological cancers worldwide. The emerging technique of liquid biopsy has been proposed as a method for detecting cancer in early stage using biofluids and their properties as biomarkers.

Methodology In this study, we tested the application of an artificial intelligence (AI) algorithm on infra-red spectra taken from urine samples from 84 female patients with gynecological cancer (28 breast, 32 endometrial, 24 ovarian and 10 cervical) and 200 non-tumor patients who were used as controls. The spectra were normalized, and outlier values were detected and removed using a DBSCAN algorithm. To overcome the possible problem of an unbalanced dataset, we used a SMOTE algorithm enhancing the generalization of the predictive model. The AI-model was trained and tested in classifying healthy urine samples vs different cancer types.

Results The spectra were divided into training- and testing-datasets with a ratio of 80/20 randomly + 10-fold cross validation and various classifiers were put under test: decision trees, discriminant analysis, support vector machines, logistic regression and random forest, with the latter giving the best results. In the classification report Precision-, Recall-, and F1-scores varied from 0.93 to 1.00, 0.88 to 1.00 and 0.94 to 0.99 respectively

Conclusion These results confirm the reports from previous, smaller studies and show that AI-models could be useful in differentiating biofluid samples, such as urine, between patients and healthy controls. Further research is needed in order to confirm the validity of the method and to assess its potential on clinical applications.

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.