Predicting Fetal Distress from Cardiotocography: A Machine Learning Model in R
Data & risk · July 2021SYRACUSE MS| COURSE | Introduction to Data Science · IST 687 |
|---|---|
| PROGRAM | M.S. in Information Management, Syracuse University iSchool |
| COMPLETED | July 2021 |
| AUTHOR | Emporia Meng |
The brief asked for an end to end applied machine learning project in R: select a dataset, engineer its features, benchmark competing models and defend the strongest one. The dataset chosen carries delivery-room stakes, 2,126 fetal cardiotocograms from the UCI Cardiotocography dataset, each carrying an expert label for fetal state as normal, suspect or pathologic.
DATA
The recordings hold 23 attributes measured from fetal heart rate and uterine contraction traces, including baseline heart rate, accelerations, fetal movement, uterine contractions and decelerations. Single measurements explain very little on their own: linear regressions on individual attributes produced adjusted R-squared values from under 0.01 to 0.12.
RESULTS
The signal emerges once the attributes are combined. Swapping four deceleration attributes into the model lifted multiple regression to an adjusted R-squared of 0.26, and a support vector machine trained on 30% of the data and tested on the remaining 70% attained 91% accuracy on unseen recordings, the strongest of the three approaches.
CONCLUSION
The result is a working demonstration that machine learning can extract clinically meaningful classifications from physiological traces whose individual measurements reveal almost nothing, and the full analysis documents the modeling choices behind it.
A support vector model classified fetal state correctly 91% of the time on recordings it had never seen, drawing signal from 23 physiological measurements that reveal almost nothing on their own.
Full version available upon request.
The support vector model did better still, classifying fetal state correctly 91% of the time on unseen data.