Abstract: P-713

Supervised Machine Learning Techniques for the Classification of Pregnancy Outcome Based on Early Pregnancy Urine Analyzed by MALDI TOF Mass Spectrometry.

Presenter: Raminta Zmuidinate


Abstract


OBJECTIVE: To apply classification machine learning algorithms to MALDI-ToF MS generated mass spectral data obtained from early pregnancy maternal urine in order to differentiate between ongoing pregnancies and pregnancy losses.
DESIGN: A retrospective cohort study of mass spectral data of maternal pregnancy urine collected in assisted reproduction clinics between May 2014 and August 2016 view more

OBJECTIVE: To apply classification machine learning algorithms to MALDI-ToF MS generated mass spectral data obtained from early pregnancy maternal urine in order to differentiate between ongoing pregnancies and pregnancy losses.
DESIGN: A retrospective cohort study of mass spectral data of maternal pregnancy urine collected in assisted reproduction clinics between May 2014 and August 2016
MATERIALS AND METHODS: A total of 716 urine samples at an average of 8 weeks of gestation were analyzed (range 5.5-11 weeks). Data included 678 ongoing pregnancy samples and 38 pregnancy loss samples, including blighted ovum, ectopic pregnancy and spontaneous abortion. Mass spectral data was acquired by MALDI-ToF mass spectrometer (Shimadzu Axima CFRplus) in a positive linear mode. Peak pre-processing in a mass region of 2000 to 12000 m/z was done using an established workflow utilizing “MALDIquant” package in R. Variable pattern features of the peaks were manually extracted (n=39), summing the ion count in the m/z regions defined by the peak area. Data classification was performed using Python’s Scikit library. Instances of under-represented class were increased using random oversampling method, achieving a balance ratio of 50:50. Tree classification ensemble learning was applied to reduce dimensionality. Decision Tree classifier model was built on a training set (67%) to predict the ongoing pregnancy outcome. The obtained accuracy estimate was 50-fold cross-validated.
RESULTS: Seventeen specific spectral features were selected as discriminant by the tree ensemble classification algorithm. Classification model was evaluated on a test set and has achieved an accuracy in predicting ongoing pregnancy of 96.9%, sensitivity 100% and specificity 93.6%. The estimated variance of an accuracy was 0.13%.
CONCLUSIONS: The study has shown encouraging results, enabling the utilization of an easy to collect urine sample as early as 6 weeks of pregnancy to provide an accurate prediction of the pregnancy outcome. Identification of a healthy pregnancy would reassure the patient and provider, and the knowledge of a potential loss would enable early counselling.

show less

Share this PosterTalk

About PosterTalks

PosterTalks allows meeting attendees the ability to view these presentations, download or bookmark their favorite presentations, download PDF versions of the posters, ask questions, leave comments, and share presentations with their colleagues – all from the convenience of a smart phone.

Contact Us

Have a question? Click here to contact us. Need technical support? Click here to email support.

© 2018 PosterTalks and Connect BioMed. All other content and data, including data entered into this website are copyrighted by their respective owners.