Machine learning facilitates access to medicines in resource-limited healthcare settings

Managing healthcare supply chains in low- and middle-income countries can mean navigating situations that are prone to extreme and unexpected disruptions. In Sierra Leone, for example, a variety of external forces can complicate the public health system, from an attempted military coup to an infectious disease outbreak to a large-scale power outage.

The consequences are serious. Despite the national government’s efforts to provide free medical care and essential supplies to pregnant women and children under the age of five, Sierra Leone has one of the highest maternal mortality rates in the world, with 717 deaths per 100,000 live births, explains Hamsa Bastani, a management researcher and statistician at the Wharton School.

The main factor is not necessarily a shortage of medicines, but an inability to get the right supplies to the right place at the right time, Bastani said. Some clinics have excess inventory, while others run out of inventory.

To address this discrepancy, Bastani, computer scientist Osbert Bastani, Ph.D., and candidate Angel Tsai Shuan Chun partnered with the government of Sierra Leone to build a low-cost decision support system that uses machine learning to predict demand and optimize how medicines are allocated.

After a pilot rollout in five districts, researchers found a 19% increase in consumption of allocated medical products in treatment areas, evidence of improved access. Their discovery is nature.

The tool predicts how much of each product individual facilities need and calculates the most efficient way to distribute limited domestic inventory, explained lead author Tsai-Hsuan Chung. It is “designed for environments where data is often sparse, noisy, and incomplete.”

The new system also addresses existing inequalities. In facilities serving poorer, more remote populations, where chronic out-of-stocks were common, the use of new tools led to a 32% spike in drug consumption.

Based on these results, the government expanded the system across the country. Currently, we are supporting allocation decisions for more than 70 essential items across the country, including medications to treat postpartum hemorrhage and eclamptic seizures, along with other essentials such as tetanus vaccines, gloves, and anti-malarial drugs, reaching an estimated 2 million women and children under 5 years of age. The system runs at a server cost of just $30 per month and requires no additional staff.

On-site work leads to real work

To build a tool that could handle Sierra Leone’s highly diverse logistics ecosystem, the researchers knew they couldn’t rely solely on remote data feeds and Zoom calls, so Tsai-Hsuan Chung headed to the capital, Freetown.

Local officials worried that AI tools coming from abroad could replace their jobs or make them liable if something goes wrong. ”

Tsai-Xuan Chung, first author

To secure local buy-in and gain trust, she spent weeks conducting individual training sessions and ensuring fair compensation for her time. She led the design of a web application that closely mirrored the agency’s existing spreadsheet workflows, reducing the friction of having employees learn complex and disparate software systems.

Hamsa Bastani added: “Importantly, the system primarily acts as a ‘decision support’ tool, where local authorities retain the final say and can override recommendations.”

Inside the AI ​​system

Clinics that are understaffed and under-resourced are least able to consistently report data, concentrating data gaps around the very places they are needed most. That leads to subtle distortions. When a model learns only from the cleanest data, it overlooks clinics with poor records but in urgent need, in favor of the best-documented clinics, those that already provide better service.

Teams use multi-task learning to avoid this bias. This allows the model to borrow shared patterns, such as seasonal demand, from data-rich locations and apply them to locations where records are sparse.

They combined this with a “backstop” built from external information such as census data showing human activity and Google Earth images showing vegetation around the clinic. This approach helped define catchment areas based on travel times between those areas and facilities. By combining these data with census data on the proportion of women and children living in the area, the algorithm was able to derive a baseline estimate of the amount of medication a clinic would need based purely on local demographics.

Although these estimates do not capture all local variation, they can provide a stable baseline associated with population and geography.

Looking to the future

Now that ownership of the allocation tool has been fully transferred to the Sierra Leone government, the research team is starting to look abroad. Tsai-Hsuan Chung is currently working on another project with Somaliland officials and plans to work with partners in Taiwan to adapt similar data-driven approaches to health systems in other regions.

Ultimately, the team hopes their work will serve as a definitive blueprint for the future, demonstrating that machine learning can powerfully improve healthcare delivery in resource-constrained environments at low cost.