Managing medical supply chains in low- and middle-income countries can be particularly difficult when health systems face sudden and unpredictable disruptions. In Sierra Leone, political instability, infectious disease outbreaks and widespread power failures can all interfere with the delivery of essential medicines. The consequences can be severe. Despite a national initiative providing free medical care and essential supplies to pregnant women and children under five, the country continues to have one of the world’s highest maternal mortality rates. Hamsa Bastani, an operations researcher and statistician at the Wharton School, says the problem is not always a shortage of medicine. Often, supplies fail to reach the facilities that need them most, leaving some clinics overstocked while others experience chronic shortages.
To tackle this imbalance, computer scientist Osbert Bastani and PhD candidate Angel Tsai-Hsuan Chung partnered with Sierra Leone’s government to develop an affordable decision-support system powered by machine learning. The tool forecasts demand for medical products at individual health facilities and recommends how limited national supplies should be distributed. Following a pilot programme in five districts, researchers recorded a 19% increase in the consumption of allocated medical products in areas using the system, indicating improved access to essential supplies. The findings, published in Nature, also showed particularly strong benefits for poorer and more remote communities, where medicine consumption rose by 32%.
Encouraged by the results, the government expanded the system nationwide. It now helps guide allocation decisions for more than 70 essential health products, including medicines used to manage postpartum haemorrhage and eclampsia, as well as tetanus vaccines, gloves and antimalarial treatments. The programme is estimated to reach around two million women and children under five across Sierra Leone. Remarkably, the technology operates at a server cost of only about $30 a month and requires no additional workforce. The system demonstrates how relatively inexpensive artificial intelligence can strengthen health care logistics without demanding major investments in infrastructure or staffing.
Developing an effective tool required more than analysing data remotely. Chung travelled to Freetown to understand the country’s complex logistics network and work directly with local officials. Some employees initially worried that an overseas AI system might replace their jobs or leave them responsible for mistakes. To build trust, Chung spent weeks providing personalised training and ensuring officials were fairly compensated for their time. She also helped design a web application that closely resembled the spreadsheet-based processes staff already used. Rather than replacing human decision-makers, the technology was deliberately built as a support tool. Local officials retain final authority and can override its recommendations whenever necessary.
One of the biggest technical challenges was incomplete and uneven data. Understaffed and under-resourced clinics are often the least able to submit consistent records, meaning the greatest data gaps tend to occur in communities with the greatest needs. A conventional model trained on the cleanest information could therefore favour well-documented facilities that are already better served. The researchers addressed this problem through multitask learning, allowing the system to identify shared patterns, such as seasonal changes in demand, from facilities with stronger data and apply those insights to clinics with limited records.
The team also created a backup approach using census information and satellite imagery from Google Earth to estimate local health care needs. Vegetation and other geographic indicators helped identify areas of human activity, while travel times were used to estimate which populations were likely to rely on particular clinics. Combined with demographic information about women and children, these data provided a stable baseline for the demand for medicine. With ownership of the allocation system now transferred to Sierra Leone’s government, the researchers are exploring similar approaches elsewhere, including Somaliland. Their work suggests that affordable, carefully designed machine learning can improve health care delivery in resource-constrained settings while keeping local officials firmly in control.
More information: Angel Tsai-Hsuan Chung et al, Improving access to essential medicines via decision-aware machine learning, Nature. DOI: 10.1038/s41586-026-10433-7
Journal information: Nature Provided by University of Pennsylvania
