# Kenyan Neonatal Mortality Risk Predictor: Protocol for a User-Centered Design Evaluation

## TL;DR

This study evaluates a machine learning-based neonatal risk prediction model in Kenyan healthcare facilities to improve early identification of high-risk newborns and reduce neonatal mortality.

## Contribution

The study introduces a user-centered evaluation of an ML-based neonatal risk predictor in a Kenyan context, focusing on feasibility, applicability, and usability in real-world clinical settings.

## Key findings

- The neonatal risk predictor is expected to integrate smoothly into existing triage systems with minimal workflow disruption.
- Early identification of high-risk neonates within 48 hours of birth is anticipated to improve timely clinical decision-making.
- Positive user experience and perceived usefulness among healthcare workers are expected outcomes of the usability assessment.

## Abstract

Neonatal mortality remains a major public health challenge in low- and middle-income countries (LMICs), particularly in sub-Saharan Africa, where health systems often lack effective triage mechanisms to identify and prioritize high-risk neonates. Existing clinical tools frequently fail to support timely decision-making during the critical early postnatal period. A previous machine learning (ML)–based neonatal risk prediction model developed using multicountry LMIC datasets demonstrated high predictive accuracy for neonatal mortality in the Indian context, achieving an area under the curve above 0.80. The model incorporates 11 neonatal parameters assessed from delivery through day 2 of life, with birth weight identified as the strongest predictor.

This study aims to evaluate the contextual feasibility, applicability, and usability of the neonatal risk predictor variables embedded in the ML model in Kenyan health care facilities to inform potential adoption aligned with national health priorities and global targets, including Sustainable Development Goal 3.2.

A mixed methods feasibility study will be conducted through the real-world preimplementation of a neonatal risk assessment tool in 3 Kenyan health facilities that provide neonatal services. Qualitative and quantitative approaches will be combined to strengthen methodological rigor and enhance the credibility of the findings. The study will be implemented in 3 sequential phases. Phase 1 will involve key informant interviews to capture contextual insights on existing workflows, stakeholder perceptions, and implementation considerations. Phase 2 will comprise a 4-month intervention period during which a paper-based neonatal risk tool will be integrated into routine clinical workflows to assess feasibility, applicability, and usability. Phase 3 will involve postintervention surveys conducted longitudinally at 2 time points to evaluate outcomes, user experience, and implementation barriers.

The neonatal risk predictors are expected to demonstrate strong contextual feasibility, allowing for integration into existing triage systems with minimal workflow disruption. Their applicability is expected to enable early identification of high-risk neonates within the first 48 hours of life, thereby enabling timely clinical decision-making. Usability assessments are expected to indicate positive user experience, acceptability, and perceived usefulness among health care workers. Six participants (frontline neonatal health workers) had been recruited by the time of initial submission of the manuscript (August 2025). Data analysis is ongoing and is expected to be concluded in March 2026. Publication of study findings is expected by June 2026.

The study is expected to generate actionable evidence to support the translation of ML-based neonatal risk prediction into routine clinical practice in LMIC settings, thereby reducing neonatal mortality and advancing progress toward Sustainable Development Goal target 3.2.