An AI That Hears Birth Asphyxia in a Baby's Cry

Birth asphyxia is the oxygen deprivation that can occur during or just after delivery, and it is one of the deadliest things that can happen to a newborn. The World Health Organization estimates that around 4 million babies experience it each year. About 1 million of them die, and a similar number survive with lasting neurological damage. The condition is largely treatable if caught quickly, yet in much of the world the tools to catch it simply do not exist. Ubenwa, a startup founded by Nigerian engineer Charles Onu, offers a startling alternative: a smartphone app that listens to a baby cry and decides within seconds whether that baby is starved of oxygen.

The idea rests on a clinical observation that dates back decades. The cry of a newborn is not random noise. It reflects the state of the brain and the nerves controlling the vocal tract, and babies suffering from asphyxia and related brain injury cry differently. Trained clinicians can sometimes hear the difference. Machines, it turns out, can learn to hear it too, and they can do it on a phone that costs a few hundred dollars.

From a Nigerian Lab to Mila in Montreal

Charles Onu began working on cry analysis as a student in Nigeria, collaborating with clinicians at a local hospital who recorded infant cries for research. He later joined Mila, the Quebec Artificial Intelligence Institute in Montreal, where the project matured into Ubenwa, a name taken from the Igbo word for cry. The company attracted attention and support from figures including Turing Award winner Yoshua Bengio, and it has been recognized through MIT Solve's global health challenges.

The founding insight was simple and ambitious at once. Birth asphyxia is normally diagnosed through clinical assessment, blood gas tests, and in well equipped hospitals, continuous monitoring of brain function. None of that is available in the clinics where most of the world's births happen. But a cry is available at every birth, and a microphone is built into every phone. If the diagnostic signal lives in the sound, then the barrier to screening collapses.

What the Studies Found

The evidence base is peer reviewed and multi country. In a prospective cohort study published in the early 2020s, researchers enrolled newborns of at least 36 weeks gestational age at five hospitals across Nigeria, Brazil and Canada. Using a smartphone running the Ubenwa app, they recorded between 30 and 180 seconds of crying from each infant within six hours of birth or at admission, and independently graded the severity of hypoxic ischemic encephalopathy, the brain injury that follows asphyxia, using standard Sarnat staging.

Earlier validation work, presented in the company's founding research, showed the system achieving sensitivity of 85 percent and specificity of 89 percent on held out test data. That means the model correctly flagged roughly 17 of every 20 asphyxiated babies while correctly clearing a similar proportion of healthy ones. The company has reported overall accuracy around 88 percent, and MIT Solve has cited performance approaching 90 percent sensitivity and specificity. Those numbers put the tool in the range of usefulness for triage in settings where the alternative is no diagnosis at all.

Why Seconds Matter

The clinical stakes of speed are hard to overstate. When a baby is deprived of oxygen around the time of birth, every minute of delay in recognition adds to the risk of permanent injury or death. Therapeutic cooling, a treatment that reduces brain damage in eligible infants, works best when started within the first hours of life. In well resourced hospitals, monitors and specialists catch trouble fast. In under resourced clinics, a newborn with asphyxia may be mistaken for simply sleepy or quiet until the damage is done.

Ubenwa compresses the recognition step into something a nurse or midwife can do with a phone. Record the cry for up to three minutes, receive a risk assessment, and escalate immediately to resuscitation or referral. The app is designed for the realities of frontline care: it does not require constant internet, it does not need a lab, and it does not need a physician to operate.

The Honest Limits

A cry based screening tool is not a diagnosis, and Ubenwa is careful to frame it that way. Sensitivity of 85 percent means some affected babies are still missed, and specificity of 89 percent means some healthy babies will be flagged and referred unnecessarily. The tool is a triage layer designed to catch signals the health system currently misses entirely, not a replacement for clinical judgment or confirmatory assessment. Scaling also requires more validation across populations, languages of the delivery room, and phone microphones, and the company has continued to publish its clinical results as it grows.

What makes the story compelling is not that the AI is perfect. It is that the problem is enormous, the existing tools are absent exactly where the deaths concentrate, and someone built a working answer out of the sound a baby already makes at birth.

A Lesson for Global Health AI

Ubenwa fits a pattern seen across the most effective health AI deployments in low resource settings: use data the setting already produces, run on hardware the setting already owns, and target the point of care rather than the specialist hospital. A cry costs nothing to capture. A phone is already in the pocket of most health workers on earth. The intelligence, once trained, is cheap to replicate.

Roughly one million newborn deaths a year are attributed to this single condition. No algorithm will erase that number alone. But a tool that turns thirty seconds of crying into an early warning gives frontline health workers something they have never had before: a way to hear the danger before it becomes a tragedy.