Charles Ono of Ubenwa Health describes using AI to detect neonatal asphyxia from infant cries. The company built a smartphone app and deep learning system trained on thousands of clinically annotated cry recordings from Brazil, Canada, and Nigeria to catch a condition that kills hundreds of thousands of newborns yearly.
Questions this video answers
Over 1 million babies die each year.
About 1.2 million babies develop such conditions each year.
Brazil, Canada, and Nigeria.
A convolutional neural network (CNN).
Chapters
Hi, my name is Charles Ono and I'm founder and head of AI at Ubenwa Health. Obeyua's mission is to bring about a world in which every baby has access to rapid medical diagnostics, regardless of where they are born in the world. So the first issue we are tackling is that of neonatal asphyxia, which is one of the top causes of sudden death and disability in infants. And asphyxia basically results when the newborn is unable to sustain spontaneous breathing. And this could be due to a variety of underlying causes. What happens is that it's a really critical situation where the brain cells are not getting enough oxygen, and if that continues for an extended period, it could result in irreversible brain damage. Today, medical diagnosis for asphyxia requires expensive equipment. Specialized personnel, typically involving blood analysis and the neurological exam and conducting such tests on a routine basis possess a big limitation for medical centers in developing countries. And on the other hand, even in developed countries, if an infant suffers asphyxia outside of a well resourced hospital, for instance, at home, it could still be fatal. And so for these reasons, the global casualty due to neonatal asphyxia is quite high. Every year, over 1 million babies die and another 1.2 million develop conditions like hearing loss, paralysis, learning difficulties and others. I personally first encountered asphyxia during my work with NGOs in my home country, Nigeria. And I can say that it was bad and led me to clinical data about the physiology of cryoproduction, the connection between the voice and the brain, and how conditions like asphyxia effectively alter the patterns of the infant's cry sounds. And so today my team at Ubinua is focused on helping newborns to survive and thrive by paying attention to the messages that their cry sounds carry. So what exactly is our solution? At Ubenwa, we have developed a smartphone app for parents and caregivers to rapidly identify or detect the early signs of neurological anomalies using a simple recording of the baby's cry. And the idea is that if babies can be identified early, then they can be treated on time. And when you compare this to the current methods of screening, the app offers at least three key benefits. One, it is fast facilitating early detection. Two, it is cost effective, meaning global access. More persons can use it regardless of their economic status. And three, it's non invasive, so causes no harm to the patient. And the test can be repeated as needed to develop puebnois, AI and machine learning has played a very important role. This is because the changes in cryoicaracteristics caused by neonatal asphyxia are very complex. And so even though we know some of these characteristics, like the high pitch or fundamental frequency, there are several others which will be non trivial to extract by hand engineering, but which are important to really developing an accurate system. And so what machine learning, and specifically deep learning, is allowing us to do is to build a neural network that automatically learns these features from cryo recordings and then internally figures out how they should be combined mathematically to make a classification. And in practice, this typically involves training a convolutional neural network, a CNN on spectrogram representations of the impute audio signals. The data used to train the CNN is probably the most important component of this setup. It has to be large and diverse because cry patterns vary from individual to individual. And we want to learn features that are truly global. And also the data has to be strongly annotated, else the resulting model may not have clinical utility. So to this end, we partnered with neonatologists in different hospitals across three countries, Brazil, Canada and Nigeria. And with them we have now developed the largest database of clinically annotated infant cry sounds, which has now played a vital role in the development and the training of our machine learning models. You have learned about why we are developing Ubenoa. Ubenua is actually an expression in my native language, Igbo, which means the cry of a baby. We're not just developing a single mobile application, but actually a flexible software platform that can be deployed onto a range of devices to really leverage the cry of a baby as a vital sign. We hope that in no distant time we'll be able to save and improve the lives of millions of babies all over the world.
