A low-cost artificial intelligence tool built by researchers at the Wharton School and Penn Engineering is helping to get essential medicines to the communities in Sierra Leone that need them most, by forecasting demand and correcting for missing data.
The researchers partnered with Sierra Leone’s government to build a decision-support system that uses machine learning to forecast patient demand and optimise the allocation of essential medical supplies. The algorithm works out the most efficient way to distribute limited national stock, so that lifesaving medicines reach the clinics where they are needed. After a successful pilot yielded a 19% increase in the consumption of allocated medicines, the government scaled the system nationwide, where it now supports more than 70 products on just $30 a month in server costs.
Managing a medical supply chain in low- and middle-income countries can mean navigating a landscape prone to extreme and unexpected disruption. In Sierra Leone, external forces ranging from an attempted military coup and an infectious disease outbreak to a widespread electricity outage can complicate public health logistics.
The consequences are severe. Despite a national government initiative dedicated to providing free medical care and essential supplies to pregnant women and children under five, Sierra Leone has one of the highest maternal mortality rates in the world, at 717 deaths per 100,000 live births, explains Hamsa Bastani, an operations researcher and statistician at the Wharton School.
A major driver is not always a lack of medicine but a failure to get the right supplies to the right place at the right time, says Bastani. Some clinics end up overstocked while others run dry.
To address that mismatch, Bastani, computer scientist Osbert Bastani and PhD candidate Angel Tsai-Hsuan Chung partnered with Sierra Leone’s government to build a low-cost decision-support system that uses machine learning to forecast demand and optimise how medicines are allocated. Following a pilot rollout in five districts, the researchers found a 19% increase in consumption of allocated medical products in treated areas, a proxy for improved access. Their findings are published in Nature.
The tool predicts how much of each product individual facilities are likely to need, then computes the most efficient way to distribute the limited national stock, explains first author Tsai-Hsuan Chung. It is, she says, “designed for a setting where data are sparse, noisy, and often incomplete.”
The new system also addresses previous inequities. Facilities serving poorer, more remote populations that frequently experienced chronic stockouts saw a 32% surge in medicine consumption with the new tool.
Based on these results, the government scaled the system nationwide. Today it supports allocation decisions for more than 70 essential products, including medicines to help with postpartum haemorrhaging and treat the seizures of eclampsia, alongside other essentials such as tetanus vaccines, gloves and antimalarial medicines, reaching an estimated two million women and children under five. The system runs on only $30 per month in server costs and requires no additional workforce.
To build a tool capable of handling Sierra Leone’s highly varied logistical ecosystem, the researchers knew they could not rely solely on remote data feeds or Zoom calls, so Tsai-Hsuan Chung travelled to the capital city of Freetown.
“Local officials were worried that an AI tool arriving from abroad might replace their jobs or leave them responsible if something went wrong,” Tsai-Hsuan Chung says.
To secure local buy-in and gain trust, she spent weeks conducting personalised training sessions and ensuring fair compensation for their time. She led the design of a web application that closely mirrored the agency’s preexisting spreadsheet workflows, reducing the friction of forcing workers to learn a complex, alien software system.
“Crucially,” adds Hamsa Bastani, “the system chiefly functions as a ‘decision-support’ tool wherein local officials always retain final say and can override recommendations.”
Understaffed and under-resourced clinics are the least able to consistently report data, leaving gaps clustered around the very places where need is greatest. That leads to a subtle distortion. If a model learns only from the cleanest data, it will favour the best-documented clinics, the ones already better served, while overlooking those where the record is thin but the need is acute.
The team circumvents this bias using multitask learning, which allows the model to borrow shared patterns, such as seasonal demand, from places with richer data and apply them where records are sparse.
They paired that with a backstop built from external information, including census data and Google Earth images of the vegetation around the clinics, which indicate human activity. This approach helped define catchments on the basis of travel time between those areas and the facilities. When those data were combined with census data on the proportion of women and children living within zones, the algorithm could tease out a baseline estimate for how much medicine a clinic needed based purely on the local demographics.
These estimates do not capture every local fluctuation, but they provide a stable baseline tied to population and geography.
With ownership of the allocation tool now fully transferred to Sierra Leone’s government, the research team is turning outward. Tsai-Hsuan Chung is currently working on another project with officials from Somaliland, collaborating with Taiwanese partners to adapt similar data-driven approaches to other regional health systems.
Ultimately, the team hopes its work serves as a definitive blueprint for the future, demonstrating that machine learning can powerfully improve healthcare delivery in resource-constrained environments at low cost.
