Source-linked AI summary
Ubenwa: Cry-based Diagnosis of Birth Asphyxia
Charles C Onu, Innocent Udeogu, Eyenimi Ndiomu, Urbain Kengni, Doina Precup, Guilherme M Sant'anna, Edward Alikor, Peace Opara
TL;DR
Birth asphyxia is difficult to detect early in resource-poor settings because current diagnosis requires costly equipment and expertise. Ubenwa analyzes infant cries with MFCC features and an SVM, achieving reported sensitivity of 85% and specificity of 89% while supporting a rapid mobile diagnostic prototype. The approach is intended to make screening more accessible, although the work’s validation and deployment plans remain institutionally constrained.
Problem
Early birth-asphyxia detection is limited in resource-poor settings because existing diagnosis requires costly equipment, logistics, and expertise.
Source: §1 · p. 1Method
Ubenwa segments infant cries, extracts MFCC features, classifies segments with an SVM, and aggregates them into a sample-level diagnosis.
Source: §2 · p. 2, §2.1 · p. 2Results
Sensitivity was 85% and specificity was 89% on a held-out test set.
Source: §3 · p. 2Takeaways & Limitations
A mobile cry-based prototype offers a non-invasive, low-cost, low-skill diagnostic process that returns results in under 20 seconds.
Source: §3.1 · p. 3Validation and deployment remained constrained by pending or ongoing institutional ethics-board approvals at MUHC and UPTH.
Source: §4.1 · p. 4
Abstract
from arXiv · showhide
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.
1 Problem Statement
Birth asphyxia is a major cause of newborn mortality and lifelong disability, but early detection remains difficult where diagnostic equipment, expertise, and logistics are limited.
- Nearly 3 million babies died within 28 days in 2016, while birth asphyxia ranked among the top three global causes of newborn mortality.Birth asphyxia also causes severe lifelong disabilities in over 1 million infants annually.Source: §1 · p. 1
- Early detection is limited in resource-poor settings because existing diagnostic systems are costly and logistically demanding.Asphyxia is often recognized only after visible symptoms emerge, when severe neurological damage may already have occurred.Source: §1 · p. 1
- Confirmatory diagnosis combines arterial blood-gas analysis with the APGAR score, but blood-gas analysis is unavailable in many developing regions.The resulting diagnostic gap contributes to high asphyxia-related morbidity and mortality in low- and middle-income countries.Source: §1 · p. 1
- The paper motivates cheap, easy-to-use methods that enable early diagnosis by community health workers, midwives, nurses, or parents.Such methods could make life-saving diagnosis possible in settings where current systems are inaccessible.Source: §1 · p. 1
2 Approach
Ubenwa diagnoses birth asphyxia from infant cries using MFCC audio features and an SVM, with segment-level predictions aggregated into a sample-level classification.
2 Approach
- The study uses infant cry recordings as a source of physiological information and develops a prototype for detecting birth asphyxia in resource-poor settings.The approach is evaluated retrospectively and targets accessible diagnosis.Source: §2 · p. 2
2.1 Mel Frequency Cepstral Coefficients (MFCC) with Support Vector Machines (SVM)
- MFCC features are combined with an SVM classifier, leveraging speech-recognition methods suited to limited examples and high-dimensional data.MFCCs represent audio signals in a way that closely mimics human auditory processing, while SVMs learn complex nonlinear decision boundaries.Source: §2.1 · p. 2
- Each cry sample is divided into time segments, preprocessed, converted to MFCCs, and classified before majority voting assigns the sample-level label.The final classification is normal or asphyxia according to the majority of segment predictions.Source: §2.1 · p. 2
3 Results
On a retrospective cry dataset, Ubenwa detected asphyxia with reported sensitivity of 85% and specificity of 89%, then packaged the model in a mobile prototype designed for rapid, accessible use.
3 Results
- The dataset contained cries from 69 normal, asphyxiating, and deaf infants, synthesized into 1389 normal and asphyxiating samples for analysis.Only normal and asphyxiating newborn cries were relevant to this study.Source: §3 · p. 2
- Sensitivity was 85% and specificity was 89% on a held-out test set comprising 20% of the data.The remaining 80% was used to train and validate the algorithms.Source: §3 · p. 2
3.1 Prototype
- Ubenwa was implemented as a mobile application incorporating the current diagnostic model and informed by feedback from clinicians and caregivers.The prototype supports the paper’s goal of making diagnosis more accessible beyond laboratory machine learning.Source: §3.1 · p. 3
- The prototype is non-invasive, low-cost, requires little or no skill to operate, and delivers results in under 20 seconds compared with blood-gas analysis.It uses cry instead of blood and costs only as much as a phone.Source: §3.1 · p. 3
- Birth attendants, parents, and other caregivers can use Ubenwa to detect asphyxia quickly and refer newborns for potentially life-saving treatment.The proposed deployment emphasizes rapid action after detection.Source: §3.1 · p. 3
3.2 MFCC/SVM vs Deep learning
- MFCC/SVM was selected for the pilot because it can be trained with few examples and ported to mobile devices, while deep models are reserved for future larger-data development.The authors state that future deep neural networks could potentially improve performance as data and mobile deployment frameworks mature.Source: §3.2 · p. 3
4 Next Steps
Ubenwa’s next steps focus on collecting more cry data, validating the mobile system in hospitals, and optimizing the algorithm for practical deployment.
4.1 Data acquisition and product validation plans
- One year of data collection at hospitals in Canada and Nigeria will expand pathological cry samples and support validation on new subjects.The planned sites are UPTH in Port Harcourt and MUHC in Montreal.Source: §4 · p. 3, §4.1 · pp. 3–4
- The planned hospital acquisition will validate the current system in a real-world scenario, with ethics approval obtained at UPTH and pending at MUHC.Source: §4.1 · pp. 3–4
- Figure 3 shows actual screenshots of the Ubenwa mobile app used in the deployment plans.Source: Figure 3 · p. 3
4.2 Optimisations
- Optimization will target environmental-noise robustness, shorter recordings, and lower memory and computation requirements on mobile devices.Source: §4.2 · p. 4