Inspection of Blackbox Models for Evaluating Vulnerability in Maternal, Newborn, and Child Health

William Ogallo, Skyler Speakman, Victor Akinwande, Kush R. Varshney, Aisha Walcott-Bryant, Charity Wayua, Komminist WeldemariamIJCAI 2020 · OpenReview

Abstract

Improving maternal, newborn, and child health (MNCH) outcomes is a critical target for global sustainable development. Our research is centered on building predictive models, evaluating their interpretability, and generating actionable insights about the markers (features) and triggers (events) associated with vulnerability in MNCH. In this work, we demonstrate how a tool for inspecting "black box" machine learning models can be used to generate actionable insights from models trained on demographic health survey data to predict neonatal mortality.

Introduction

Improving Maternal, Newborn, and Child Health (MNCH) is a cornerstone of global sustainable development goals. However, the factors influencing health outcomes—like neonatal mortality—are complex, involving a web of biological, socio-economic, and environmental variables. While modern machine learning (ML) algorithms are excellent at predicting outcomes from this complex data, they are often "black boxes." This lack of transparency makes it difficult for policymakers and healthcare experts to trust them or understand why a certain subpopulation is at risk.

This research addresses the "interpretability gap" by presenting a framework for using Prospector, an interactive visual analytics tool, to interpret ML models. By moving beyond simple predictions to explainable insights, we aim to help health stakeholders identify specific vulnerabilities and plan targeted interventions to reduce neonatal mortality.

Methodology

The developed pipeline transitions from raw survey data to interactive visualization through three primary stages:

  1. Modeling: An ensemble classifier using Gradient Boosted Trees was trained to identify women who had experienced the death of a neonate (a child younger than 28 days).
  2. Explainability Integration: The trained models were uploaded into Prospector, a visual analytics tool designed for model inspection.
  3. Interpretability Techniques: The tool utilizes feature importance rankings to identify key variables, Partial Dependence Plots (PDPs) to visualize risk changes relative to single features, and localized inspection to generate "counterfactual" data for individual profiles.
The research pipeline: from DHS survey data to interactive model inspection tools.

Data and Features

The study utilized Demographic and Health Surveys (DHS) data from 10 Sub-Saharan countries. These national surveys, conducted every 3 to 5 years, provide a rich source of health indicators. The researchers extracted 120 features in total, comprising 43 household-level features (e.g., wealth index, source of water) and 77 individual-level features (e.g., education, ethnicity). The unit of analysis was the mother, specifically those who reported giving birth at least once in the five years preceding the survey.

Key Insights and Results

The core finding of this work is that the value of ML in global health lies in explanation rather than just prediction. Interactive visualization allows MNCH experts to validate models against clinical knowledge, building trust.

The most important markers for neonatal mortality across various years and countries were the number of births in the last five years and the number of household members. The model confirmed a well-known positive correlation between the frequency of births and neonatal mortality risk, serving as a vital "sanity check."

Feature importance rankings and a Partial Dependence Plot showing the relationship between household size and predicted risk.

The "Household Size" Discovery

A potentially novel finding emerged regarding household size. The model indicated a negative correlation between the number of household members and neonatal mortality. Using PDPs, the authors showed that the risk of neonatal mortality drops sharply as household size increases from 1 to roughly 10 members, after which the benefit plateaus. In many of the studied regions, larger households may provide a stronger support network or shared resources that act as a protective factor for newborns.

Conclusion and Future Work

While these findings highlight significant markers, the authors emphasize that these are correlations, not necessarily causal relationships. Future work will focus on:

  • Individual Variability: Incorporating Individual Conditional Expectation (ICE) plots to visualize how predictions vary for specific subjects, rather than just looking at average effects.
  • Anomalous Pattern Detection: Developing techniques to automatically scan for subpopulations that deviate from general trends to provide granular insights for local health authorities.

By converting standard DHS survey data into an actionable pipeline, this research provides a repeatable method for health stakeholders to use advanced ML for targeted intervention planning.

Summary updated

Abstract

Improving maternal, newborn, and child health (MNCH) outcomes is a critical target for global sustainable development. Our research is centered on building predictive models, evaluating their interpretability, and generating actionable insights about the markers (features) and triggers (events) associated with vulnerability in MNCH. In this work, we demonstrate how a tool for inspecting "black box" machine learning models can be used to generate actionable insights from models trained on demographic health survey data to predict neonatal mortality.
BibTeX
@inproceedings{ogallo2020inspection,
  author = {William Ogallo and Skyler Speakman and Victor Akinwande and Kush R. Varshney and Aisha Walcott-Bryant and Charity Wayua and Komminist Weldemariam},
  title = {{Inspection of Blackbox Models for Evaluating Vulnerability in Maternal, Newborn, and Child Health}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2020},
  url = {https://openreview.net/forum?id=-aR0MkuDLE},
}
Figure