Early detection of six-week post-discharge mortality and readmission among postpartum women and newborns in Uganda: development and temporal validation of multivariable prediction models

In many parts of the world, the most dangerous time for a mother and her newborn is not during the birth itself, but in the weeks that follow. While hospitals have become better at delivering babies safely, the care that happens after a family leaves the building often falls apart. Health systems are frequently overwhelmed, and families may leave the hospital before they have received a full check-up, only to return days or weeks later when a simple infection has turned into a life-threatening emergency. The challenge for doctors is knowing which families are safe to go home with and which ones need extra attention, extra visits, or immediate help. Without a way to spot these risks early, health workers must treat every family the same, a strategy that wastes scarce resources on low-risk cases while missing the high-risk ones who need help the most.

To solve this, a team of researchers in Uganda set out to build a practical tool that could act like a triage system at the moment of discharge. They wanted to see if they could use the basic information already available in a hospital chart—things like a mother's heart rate, the type of delivery, or whether the baby needed help breathing—to predict who would get sick again within six weeks. By analyzing data from thousands of mothers and babies, they developed simple computer models that could sort families into risk categories. The goal was not to replace doctors, but to give them a clear, evidence-based signal to help them decide who needs a follow-up visit and who can go home with peace of mind.

The study took place at two large regional hospitals in Uganda, one in the southwest and one in the east, serving a combined population of nearly one million people. Over a period of more than a year, the research team enrolled thousands of women who had just given birth. They gathered a wide range of information, from the mother's blood pressure and temperature to the baby's weight and oxygen levels. They also recorded details about the pregnancy, such as whether the mother had been diagnosed with malaria or HIV, and how long the family had to travel to reach the hospital. Once the mothers and babies were discharged, the team followed up with them six weeks later, either by phone or in person, to see what had happened. They tracked whether the mothers had been readmitted to the hospital for infections related to childbirth, and whether the babies had died or been readmitted for serious illness.

The researchers found that while most families did well, a significant number faced serious complications. Among the nearly seven thousand mothers who completed the study, none died, but more than three percent had to return to the hospital, with over half of those returns caused by infections like surgical site infections or sepsis. For the babies, the situation was slightly more precarious; about 3.8 percent of the newborns were readmitted, and a small but tragic fraction died within that six-week window. The team then used this data to train their prediction models. They tested thousands of different combinations of variables to find the smallest set of facts that could still accurately identify the families at risk. They discovered that they did not need a complex battery of tests; a handful of routine observations were enough to make a reliable prediction.

For the mothers, the final model relied on eight key pieces of information. These included how the baby was born, the mother's heart rate at the time of discharge, whether she had a urinary tract infection during pregnancy, and whether the baby needed resuscitation at birth. Other factors, such as whether the mother had HIV or malaria during her pregnancy, and whether she was leaving the hospital against medical advice, also played a role. When the researchers tested this model, it proved quite effective at spotting the women who would get sick again. It achieved an AUC of 0.8 for identifying the women who would eventually be readmitted for an infection, while achieving an AUC of 0.7 for correctly clearing those who would remain healthy. This balance meant the tool could flag the right people for extra care without overwhelming the system with false alarms.

The model for the newborns was slightly different because it had to predict two things: death and readmission. The team found that the best predictors for a baby's risk included the mother's HIV status, whether the mother had poor milk production, the baby's weight, and how long it took the family to travel to the hospital. The model also considered whether the mother had any conditions diagnosed during pregnancy and whether antibiotics were used during the admission. While the newborn model was good at identifying risk, it was less precise than the maternal model, achieving an AUC of 0.8 for identifying the babies who would die or be readmitted, but with a higher rate of false alarms. This is partly because the number of deaths was very low, making it statistically harder to build a perfect predictor for that specific outcome.

To ensure their findings were not just a lucky fluke specific to one group of people, the researchers tested their maternal model on a different set of data from the same hospital, collected a few years earlier. The model performed just as well on this older data as it did on the new data, suggesting that the rules it learned were stable and reliable over time. This gave the team confidence that the tool could work in the real world, even as patient populations shifted slightly. The study also highlighted that the most common reasons for mothers returning to the hospital were infections directly related to the delivery, such as infections of the surgical wound or the uterus, rather than unrelated illnesses like malaria. This confirmed that focusing on these specific, preventable complications was the right strategy for a discharge tool.

The researchers were careful to note the limits of their work. The models were built and tested only in Uganda, and while the hospitals served a large and diverse population, the results might not apply exactly the same way in other countries or in smaller rural clinics. They also pointed out that their tool was designed specifically to catch obstetric infections and did not account for other reasons a mother might get sick, such as mental health issues or injuries. Furthermore, because the number of newborn deaths was so small, the model for predicting infant mortality was less certain than the one for readmissions. Despite these limitations, the study demonstrated that it is possible to create a simple, risk-based system using data that is already being collected in routine care.

The ultimate value of this work lies in its potential to change how postnatal care is delivered. Instead of a blanket approach where every family gets the same advice, health workers could use these models to identify the specific dyads that need the most support. A mother flagged as high-risk could be prioritized for a home visit, given more detailed counseling on warning signs, or referred to a community health worker immediately. This targeted approach respects the reality of overburdened health systems by directing limited resources to the people who need them most. By turning routine clinical observations into a clear signal for action, the study offers a path toward preventing the preventable, ensuring that the safety of a mother and her baby extends well beyond the hospital doors.