Auxological predictors of intrauterine fetal growth restriction (FGR)

The repository with working scripts, figures, and raw data used for testing auxological FGR prediction


Team Lead: Dmitrii Zhakota

Team Members:Ruslan Alagov, Yury Malovichko, Anton Shikov



Bioinformatics Hackathon'2023

Project objectives

1. Estimate the level of comparability between obstetric, neonatal, and pathological evidence for fetal growth dynamics
2. Compare an actual dataset of newborn autopsy records to the external evidence
3. Assess the adequacy of the existing approaches to FGR diagnostics and suggest alternative approaches

Data

Analyzed raw data are included in the data directory.

• data/raw/ folder contains raw data from existing datasets and experimental data used for statistical analysis;
• data/output/ folder contains the results of gamma-regression applied on the Intergrowth21 dataset regarding body length and weight for boys, girls, or combined weight data.
• data/technical/ directory contains tables summarizing the relevance of regression.

Scripts

The src/ directory includes all code used for comparing methods, imputing, and calculating feature importance.

• src/HackFGR.Rproj file determines basic options within R scripts;
• src/to_15.ipynb script builds regressions based on Intergrow data. Data are available for different characteristics (weight, height, weight-for-height ratio) and different genders (boys, girls, combined). In each dataset, there are averages for fetal ages 24-42 weeks. For each characteristic, a regression is built for each gender, allowing us to recreate the picture for up to 15 weeks. Then, graphs are built for each regression, and a comparison is made with the data on Kiserud;
• src/Data_preparation.R script is used for comparing differences between genders for each time point of gestation weeks as well as between methods for assessing the body weight of a fetus using ultrasound, pathology analysis, and neonatological observations. In the script, comparisons between genders are calculated using pair-wise t-tests for a certain gestation week. Both raw and FDR-adjusted (false discovery rate) p-values are presented. For comparing diagnostic methods, three approaches were applied, namely, pair-wise t-test (with or without correction for multiple comparisons), meta-analysis (umbrella analysis with uniting datasets belonging to the same diagnostic method and one-sample mean meta-analysis per study), and time-series Chow test;
• src/FGR_filter.Rmd script includes approaches to both cleaning the initial source dataset from outliers and calculating metrics used for diagnosing FGR based on body length and weight of certain organs;
• src/HackFGR_regression_classification.ipynb script describes feature selection for FGR prediction using xgboost;

Plots

The pics/ directory contains graphical representations of the obtained results.

• pics/Descriptive_stats directory contains pictures depicting mean estimates of body weight within different diagnosis methods, studies, and genders as well as comparisons between genders per time and organ within the experimental dataset with pathology data;
• pics/t_tests_comparisions directory includes results of pair-wise t-tests between certain time points when comparing diagnosis methods both with and without FDR correction;
• pics/Descriptive_stats directory contains pictures depicting mean estimates of body weight within different diagnosis methods, studies, and genders as well as comparisons between genders per time and organ within the experimental dataset with pathology data;
• pics/Umbrella_meta_analysis directory comprises a graphical representation of meta-analysis, namely, umbrella meta-analysis with studies grouped according to the diagnostic method and between all studied per se with one-sample mean meta-analysis regarding individual studies;
• pics/Time_series_Chow_test directory contains plots depicting differences between diagnostic methods using all gestation weeks jointly with the Chow test applied;
• pics/Data_imputing directory includes imputing results for under-represented time points for up to 15 weeks by using a gamma regression model;
• pics/Predicting_feature_importance directory comprises plots obtained when predicting feature importances for predicting FGR assessment using xgboost as well as revealing relationships between features.

Conclusions

1. Auxological observations from pathologists, neonatologists, and obstetricians correspond to each other and can be plausibly compared
2. Several assumed patterns and approaches do not seem to hold upon multiple sources’ comparison
1. The assumption of lower weight-by-gestation the pathological evidence is only partially supported
2. Despite a long-running tradition, gender does not affect general auxological parameters
3. Advanced auxological indices may serve as plausible FGR predictors (needs further testing)