[Paper Review] Ubenwa: Cry-based Diagnosis of Birth Asphyxia

Charles C. Onu, Innocent Udeogu+6 more|arXiv (Cornell University)|Nov 17, 2017

Infant Health and Development5 references20 citations

TL;DR

Ubenwa proposes a smartphone-based machine learning system that diagnoses birth asphyxia in newborns using cry analysis, leveraging Mel Frequency Cepstral Coefficients (MFCCs) and Support Vector Machines (SVMs) to classify cries as normal or asphyxiating. The system achieved 85% sensitivity and 89% specificity on a test set of 1389 cry samples, offering a low-cost, non-invasive, and rapid diagnostic solution for resource-limited settings.

ABSTRACT

Every year, 3 million newborns die within the first month of life. Birth asphyxia and other breathing-related conditions are a leading cause of mortality during the neonatal phase. Current diagnostic methods are too sophisticated in terms of equipment, required expertise, and general logistics. Consequently, early detection of asphyxia in newborns is very difficult in many parts of the world, especially in resource-poor settings. We are developing a machine learning system, dubbed Ubenwa, which enables diagnosis of asphyxia through automated analysis of the infant cry. Deployed via smartphone and wearable technology, Ubenwa will drastically reduce the time, cost and skill required to make accurate and potentially life-saving diagnoses.

Motivation & Objective

• To develop a low-cost, accessible diagnostic tool for birth asphyxia in resource-poor settings where traditional blood gas analysis is impractical.
• To leverage infant cry patterns as a non-invasive biomarker for detecting neonatal asphyxia.
• To design a mobile-deployable machine learning system that requires minimal user expertise and operates in under 20 seconds.
• To validate the system in real-world clinical environments through data collection at hospitals in Canada and Nigeria.
• To optimize the algorithm for robustness to environmental noise, minimal memory usage, and short recording lengths.

Proposed method

• The system extracts Mel Frequency Cepstral Coefficients (MFCCs) from segmented infant cry audio to represent spectral features that mimic human auditory perception.
• A Support Vector Machine (SVM) classifier is trained on MFCC features to distinguish between normal and asphyxiating cries.
• Each cry sample is divided into time segments; classification is determined by majority voting across segments to improve robustness.
• The trained model is integrated into a mobile application named Ubenwa, enabling real-time diagnosis on smartphones.
• The approach prioritizes model portability and efficiency for deployment on low-resource mobile devices.
• Future work includes transitioning to deep learning models such as LSTMs as more data becomes available and mobile deployment frameworks mature.

Experimental results

Research questions

• RQ1 Can infant cry patterns reliably indicate the presence of birth asphyxia in newborns?
• RQ2 Can a machine learning model based on MFCCs and SVMs achieve clinically useful sensitivity and specificity for asphyxia detection?
• RQ3 Can a smartphone-based system outperform traditional diagnostic methods in terms of accessibility, cost, and speed in low-resource settings?
• RQ4 How can the system be optimized for real-world deployment, including noise robustness and minimal computational requirements?
• RQ5 What is the minimum cry recording length required for accurate diagnosis using this approach?

Key findings

• The MFCC-SVM model achieved 85% sensitivity in detecting asphyxiating infants on a test set of 1389 cry samples.
• The model demonstrated 89% specificity in correctly identifying normal infants in the same test set.
• The system can deliver results in under 20 seconds, enabling rapid clinical decision-making.
• The mobile app prototype, named Ubenwa, enables non-clinical users such as community health workers and parents to perform diagnosis with minimal training.
• The approach is non-invasive, requiring only a cry recording instead of blood sampling, significantly reducing risk and complexity.
• Data collection is underway at two hospitals—UPTH in Nigeria and MUHC in Canada—to validate the system in real-world settings and expand the dataset for future model improvements.