Classification of intrauterine growth restriction at 34-38 weeks gestation with machine learning models

Abstract

Objective: Intrauterine growth restriction (IUGR) is one of the most common causes of stillbirths. The objective of this study is to develop a machine learning model that will be able to accurately and consistently predict whether the estimated fetal weight (EFW) will be below the 10th percentile at 34+0-37 + 6 week's gestation stage, by using data collected at 20 + 0 to 23 + 6 weeks gestation.

Methods: Recruitment for the prospective Safe Passage Study (SPS) was done over 7.5 years (2007-2015). An essential part of the fetal assessment was the non-invasive transabdominal recording of the maternal and fetal electrocardiograms as well as the performance of an ultrasound examination for Doppler flow velocity waveforms and fetal biometry at 20 + 0 to 23 + 6 and 34 + 0 to 37 + 6 week's gestation. Several predictive models were constructed, using supervised learning techniques, and evaluated using the Stochastic Gradient Descent, k-Nearest Neighbours, Logistic Regression and Random Forest methods.

Results: The final model performed exceptionally well across all evaluation metrics, particularly so for the Stochastic Gradient Descent method: achieving a 93% average for Classification Accuracy, Recall, Precision and F1-Score when random sampling is used and 91% for cross-validation (both methods using a 95% confidence interval). Furthermore, the model identifies the Umbilical Artery Pulsality Index to be the strongest identifier for the prediction of IUGR - matching the literature. Three of the four evaluation methods used achieved above 90% for both True Negative and True Positive results. The ROC Analysis showed a very strong True Positive rate (y-axis) for both target attribute outcomes - AUC value of 0.771.

Conclusions: The model performs exceptionally well in all evaluation metrics, showing robustness and flexibility as a predictive model for the binary target attribute of IUGR. This accuracy is likely due to the value added by the pre-processed features regarding the fetal gained beats and accelerations, something otherwise absent from previous multi-disciplinary studies. The success of the proposed predictive model allows the pursuit of further birth-related anomalies, providing a foundation for more complex models and lesser-researched subject matter. The data available for this model was a vital part of its success but might also become a limiting factor for further analyses. Further development of similar models could result in better classification performance even with little data available.

Keywords: Classification; Fetal heart rate accelerations; IUGR; Machine learning; Umbilical artery Doppler.

Fig. 1.

One of the scatter plot facet grids. This one compares the variables: GBoverTime(per hour), fGBoverTime(per hour) and F3_MCA_PI. Particular focus should be on the effect the ‘0′ values have on the data.

Fig. 2.

A screenshot of the Tree Model produced by Orange’s ‘Tree’ function.

Fig. 3.

Graphical representation of the preparation process for each model.

Fig. 4.

Visual results of the Hierarchical Clustering on the data. Average ‘Linkage’ used. Each colour represents a separate cluster of values.

Fig. 5.

Comparative bar graph showing the AUC values for the different models for each evaluation method. These results were obtained using Random Sampling with Stratification.

Fig. 6.

Comparative bar graph showing the AUC values for the different models for each evaluation method. These results were obtained using Cross Validation with Stratification.

Fig. 7.

The resulting ROC curve for the final model. The curve shown is for Target Class ‘0’.

Fig. 8.

The resulting ROC curve for the final model. The curve shown is for Target Class ‘1’.