Abstract 6P
Background
Cancer screening methods transformed cancer management in the last decades. Ideally, a screening tool should be simple, non-invasive, cost-effective, and reliable. These criteria are only partially met for gynecological cancers, for the available screening methods; thus, newer, non-invasive techniques are being investigated. This study assesses the potential of combining Attenuated Total Reflection Fourier-Transform Infrared Spectroscopy (ATR-FTIR) in urine samples with machine learning (ML) algorithms to differentiate tumor patients from healthy controls.
Methods
251 patients were recruited from the gynecological unit of the University Hospital Basel, 72 diagnosed with cancer (22 ovarian, 10 cervical, 25 endometrial, 24 breast, 1 vaginal) and 179 control with confirmed benign histology. After collection, samples were aliquoted and immediately frozen at -80C°. After thawing, 20 μl were deposited on ATR-FTIR crystal and directly measured. Each sample’s spectrum was created averaging 24 measurements, repeated 3 times, obtaining a total of 18072 spectra. Outliers were excluded using a Density-based spatial clustering algorithm. The data obtained were then processed using an “in-house” built algorithm to solve the binary classification of healthy VS cancerous patients.
Results
All data were preprocessed using standardization and constant removal to eliminate noise and went through a Least absolute shrinkage and selector operator (LASSO) feature selection process to select the most valuable parts of the spectra and reduce computational complexity. A set of classification algorithms was tested, including Support Vector Machines, Logistic Regression, Random Forest (RF), and Decision Trees. RF was the most performing reaching an accuracy of 90%. The final model was validated by using a 10-fold cross-validation technique.
Conclusions
This study presents promising results in using ATR-FTIR combined with ML Algorithms to detect urine samples from gynecological tumor patients VS healthy controls. Although we provide a significantly larger data volume than other published studies, further research in larger sample sizes is required to confirm the results and establish a possible screening method.
Legal entity responsible for the study
The authors.
Funding
Has not received any funding.
Disclosure
All authors have declared no conflicts of interest.