Harnessing infant cry for swift, cost-effective diagnosis of Perinatal Asphyxia in low-resource settings

Abstract

Perinatal Asphyxia is one of the top three causes of infant mortality in developing countries, resulting to the death of about 1.2 million newborns every year. At its early stages, the presence of asphyxia cannot be conclusively determined visually or via physical examination, but by medical diagnosis. In resource-poor settings, where skilled attendance at birth is a luxury, most cases only get detected when the damaging consequences begin to manifest or worse still, after death of the affected infant. In this project, we explored the approach of machine learning in developing a low-cost diagnostic solution. We designed a support vector machine-based pattern recognition system that models patterns in the cries of known asphyxiating infants (and normal infants) and then uses the developed model for classification of 'new' infants as having asphyxia or not. Our prototype has been tested in a laboratory setting to give prediction accuracy of up to 88.85%. If higher accuracies can be obtained, this research may be a key contributor to the 4th Millennium Development Goal (MDG) of reducing mortality in under-five children.

Ubenwa: Cry-based Diagnosis of Birth Asphyxia

“…80% of the data was used to train and validate the algorithms while 20% was kept aside as a test set. Results on this test set showed sensitivity (accuracy in detecting asphyxiating infants) and specificity (accuracy in detecting normal infants) of 85% and 89%, respectively [11].…”

“…SVMs are powerful classifiers that can learn complex, non-linear decision boundaries. Compared to other non-linear classifiers like neural networks, SVMs are designed to work effectively with limited examples and high-dimensional data [11], as is the case in our problem.…”

Neural Transfer Learning for Cry-Based Diagnosis of Perinatal Asphyxia

“…They achieved a precision and recall of 72.7% and 68%. Building on this work, Onu et al [5] improved the precision and recall to arXiv:1906.10199v2 [cs.LG] 2 Jul 2019 73.4% and 85.3%, respectively, using support vector machines (SVM). It is worth noting that both works represent an overestimate of performance as authors split train/test set by examples, not by subjects.…”

“…We implement and compare the performance of our transfer models with 2 baselines. One is a model based on a radial basis function Support Vector Machine (SVM), similar to [5]. The other is a res8 model whose initial weights are drawn randomly from a uniform Glorot distribution [14] i.e., according to U (−k, k) where k = √ 6 n i +no , and ni and no are number of Figure 2: Audio length analysis highlighting the impact of using shorter amounts of input audio on UAR performance.…”

“…The Baby Chillanto Infant Cry database [4], based on 69 infants, remains the only known available database for this task. Previous work using this data has mainly focused on classical machine learning methods or very limited capacity feed-forward neural networks [4,5].…”