Most AI health projects start in well-resourced hospitals with clean data and reliable infrastructure. This one started with the opposite problem. Researchers from the Wharton School and Penn Engineering partnered directly with Sierra Leone’s government to build a machine learning tool designed for a system where data is incomplete, supply chains are fragile, and the cost of getting it wrong is a patient going without essential medicine.

What the tool actually does

The system is a decision-support tool, not an autonomous replacement for human judgment. It uses machine learning to forecast medicine demand at the facility level, while also correcting for gaps in historical data — a chronic problem in low-resource health systems where stock records are inconsistently kept or simply missing. By accounting for those gaps rather than ignoring them, the forecasts are more accurate than they would be with a standard approach that treats missing data as zero demand.

That distinction matters. Undercounting demand because of data gaps is one of the main reasons facilities run out of stock. The tool addresses the problem at its source.

Why this approach is different

A lot of digital health investment in low- and middle-income countries has produced tools that work in pilots and then stall. They’re too expensive to maintain, too complex for local staff to use, or built without genuine government buy-in. This project was built with Sierra Leone’s government as a partner, not an afterthought. That structural choice is probably the most important thing about it.

The low-cost design is also significant. Resource-constrained health ministries can’t absorb expensive software contracts. A tool that delivers real forecasting value without requiring significant ongoing spend has a much better chance of being used consistently and at scale.

The broader supply chain problem this targets

Medicine stockouts are a persistent and underreported crisis across sub-Saharan Africa. Facilities routinely run out of antibiotics, antimalarials, and maternal health drugs, not because the medicines don’t exist in-country, but because distribution is poorly matched to actual demand. Forecasting is the missing piece. When procurement and distribution decisions are based on rough estimates or last year’s numbers, the system stays inefficient regardless of how much funding flows through it.

Machine learning tools that can work with imperfect data are particularly well-suited to this problem because perfect data is simply not available. So building a system that performs despite data quality issues is a practical necessity, not a compromise.

What this could mean going forward

The immediate impact is better medicine availability for communities in Sierra Leone. But the model here — a low-cost, government-integrated, machine learning forecasting tool built specifically for low-resource conditions — is something other health ministries will be watching. If the results hold up in implementation, it offers a replicable approach to one of global health’s most persistent logistics failures.

• Forecasts medicine demand at the individual facility level
• Corrects for missing or incomplete historical stock data
• Built in direct partnership with Sierra Leone’s national government
• Designed to be low-cost and practical to maintain long-term
• Targets the gap between available medicines and their actual distribution

That’s not a small thing. Getting essential medicines to the people who need them is one of the most basic functions a health system has to perform. Tools that help do that more reliably, cheaply, and in real-world conditions deserve serious attention.