Prediction of neonatal mortality in Sub-Saharan African countries using data-level linkage of multiple surveys

Girmaw Abebe Tadesse, Celia Cintas, Skyler Speakman, Komminist WeldemariamarXiv 2020 · arXiv

Abstract

In many developing nations, health data is often fragmented across disjoint surveys with no common participants or features. This study introduces a "data-level linkage" framework that projects these disparate datasets into a shared latent space to enable cross-survey learning. Applying this method to neonatal mortality prediction in Ethiopia, the researchers achieved a significant performance boost, increasing the Area Under the Receiver Operating Characteristic (AUROC) curve from a baseline of 45.9% to a peak of 94.1% by linking local data with surveys from other Sub-Saharan African countries.

Introduction

Critical health data in developing regions, such as records on child mortality or family planning, is frequently collected in "silos." Organizations like the DHS Program and PMA2020 conduct surveys at different times and locations using varying questionnaires. This fragmentation creates "small data" challenges, particularly for rare events like neonatal death, which may occur in only 2-4% of samples. Standard data augmentation techniques often fail with tabular survey data. This paper addresses the challenge of linking these disjoint datasets to improve predictive power without requiring overlapping participants.

Methodology

The proposed framework establishes a bridge between two unrelated datasets ($D_1$ and $D_2$) through a multi-step pipeline:

  1. Dimension Reduction: Disparate features are projected into a shared, equal-dimensional latent space ($R$). The authors evaluated Feature Importance, Principal Component Analysis (PCA), and Autoencoders (AE) for this task.
  2. Linkage Matrix Computation: Euclidean distances are calculated between all samples in $D_1$ and $D_2$ within the shared space.
  3. Cross-Domain Neighbor Selection: For each sample in the target dataset, the model identifies its closest neighbors in the source dataset.
  4. Feature Augmentation: The feature values of these neighbors are aggregated (e.g., using the median) and concatenated onto the original sample, enriching it with borrowed context.
The cross-domain linkage pipeline for integrating disjoint datasets.

Experimental Setup

The study utilized data from two major sources:

  • DHS (Demographic and Health Surveys): Data from Burkina Faso (BF), Ethiopia (ET), Ghana (GH), Kenya (KE), and Nigeria (NG) spanning 2004–2017.
  • PMA (Performance Monitoring for Action): PMA2020 data collected in Ethiopia in 2016.

The primary objective was to leverage the larger, multi-country DHS datasets to enhance predictions on the smaller Ethiopia PMA dataset.

Results and Discussion

The use of Autoencoders for latent alignment proved most effective, as they capture non-linear commonalities in biological and socioeconomic factors that linear methods like PCA miss.

AUROC performance comparison across different linking methods and source datasets.

As shown in the results table, the baseline AUROC for predicting neonatal death using only the Ethiopia PMA dataset was 45.9%—worse than random chance due to extreme class imbalance. Implementing the Autoencoder-based linkage caused performance to skyrocket. Notably, linking Ethiopian PMA data with DHS data from Ghana yielded the highest AUROC of 94.1%, outperforming the linkage with Ethiopia's own DHS data (90.0%). This suggests that high-quality regional data can sometimes provide more predictive variance than local data.

Conclusion

The data-level linkage framework demonstrates that "neighbor-borrowing" from external surveys can effectively mitigate data sparsity and imbalance in health informatics. While the results are promising, future work should address the computational complexity of pairwise distance calculations for larger datasets and explore the interpretability of the specific features driving these improved predictions.

Summary updated

BibTeX
@misc{tadesse2020prediction,
  author = {Girmaw Abebe Tadesse and Celia Cintas and Skyler Speakman and Komminist Weldemariam},
  title = {{Prediction of neonatal mortality in Sub-Saharan African countries using data-level linkage of multiple surveys}},
  year = {2020},
  eprint = {2011.12707},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2011.12707},
}
Figure