Researchers develop new mobile app to help detect birth asphyxia

Birth asphyxia (BA) is a condition that occurs when newborn babies do not receive enough oxygen during delivery, and it’s one of the primary causes of neonatal death. Developing countries, particularly in sub-Saharan Africa, experience the highest under-five mortality rates. Researchers from Carnegie Mellon University are developing a new mobile application, called HumekaFL, to detect BA.

Related posts

Urgent policy actions needed to address real AI threats, scientist reveals

June 8, 2025

A New Law of Nature Attempts to Explain the Complexity of the Universe

June 8, 2025

The study is published on the arXiv preprint server.

Early detection of BA and timely intervention can facilitate full recovery for infants with mild or moderate asphyxia. Delayed detection results in prolonged oxygen deprivation, leading to permanent injuries that can affect organs such as the brain, heart, lungs, kidneys, and bowels. HumekaFL records newborn babies’ cries through a smartphone app and passes it through a machine learning model to detect BA.

This is not the first software designed to detect BA, but it seeks to solve some of the problems that prevent other applications from being widely adopted. One barrier is security and privacy concerns. Other prominent BA-detecting apps use centralized machine learning techniques that introduce privacy vulnerabilities because they require sensitive health data to be exported to a central server.

Unlike other apps, HumekaFL uses a special type of machine learning called federated learning (FL). This decentralized method prioritizes security and privacy by distributing the model training across multiple clients (clients are repositories like servers and computers).

For example, local data can be collected at a hospital. Babies born there are checked for BA using that hospital’s version of the HumekaFL model. This data can also be used to update the HumekaFL model at that hospital. Models from all clients are periodically aggregated and sent back so they can learn from each other and perform optimally. The key distinction is that individuals’ health care data never leaves the local client, only the models do. This makes the data less susceptible to a large-scale breach.

HumekaFL also addresses the lack of user-friendly BA detection methods that don’t require prior expertise or special equipment. There are established methods physicians can use to diagnose BA, but they can require extensive training and experience to be accurate. HumekaFL is a user-friendly and accurate detection method that runs on commodity hardware like smartphones. This could be helpful in under-resourced areas.

“We would like to deploy machine learning in a very easy-to-use way,” explained Carlee Joe-Wong, a professor of electrical and computer engineering who worked on HumekaFL. “We’re targeting under-resourced clinics or hospitals where there might not be enough trained personnel to be fully using all of the state-of-the-art techniques to monitor babies with birth asphyxia.”

Another obstacle is a lack of computing resources. Other BA detection apps use models that require large datasets and intensive processing power, but HumekaFL uses support vector machine algorithms along with smaller datasets to train their model. These algorithms differ from other machine learning methods because they are highly efficient at learning from small, high dimensional datasets.

HumekaFL’s development was led by students and faculty from Carnegie Mellon University Africa, the College of Engineering’s location in Kigali, Rwanda. Joe-Wong, who is based in Pittsburgh, said she’s enjoyed the global collaboration.

“I tried to take on more of an advisory role because the idea came from them. A lot of this architecture that handles privacy and resource constraint challenges was their idea,” said Joe-Wong. “They’ve demonstrated that they can take some of the machine learning that they’ve learned in their courses and apply it to a real problem that they’re trying to solve.”

For HumekaFL to be fully effective in Africa, researchers need to run more experiments that use African health data. Using this specific data will help prevent biases and improve model performance in the African communities HumekaFL seeks to help.

“You’re never going to know exactly what those nuances are or exactly what those specific patterns are if you don’t collect data from the population where you actually want to deploy this model,” said Joe-Wong. “We are looking at partnerships with African hospitals to collect more data from more representative populations to really validate all the details of how the model is going to work.”

HumekaFL was featured at the Association for Computing Machinery’s 2024 COMPASS event. The researchers include Joe-Wong; Assane Gueye, associate teaching professor at CMU-Africa and co-director of CyLab-Africa and the Upanzi Network; and CMU-Africa students Pamely Zantou, Blessed Guda, Bereket Retta, and Gladys Inabeza.