| Issue/Objective |
Essential medicine stock-outs remain a major constraint on effective service delivery in low-resource health systems, yet most supply chains still rely on retrospective reporting rather than prospective risk detection. In Ethiopia, preventable shortages disrupt PHC, undermine trust, and weaken accountability. This study developed and field-tested an AI-enabled early warning tool to predict stock-outs using routine logistics data and to assess whether risk alerts improved local supply response. |
| Methodology/Approach |
We used a retrospective-prospective design. First, we assembled 1.94 million stock transaction records from 312 public facilities across 6 regions, covering 38 tracer commodities from January 2020 to June 2025. Predictors included consumption volatility, supplier lead time, reporting completeness, order fill rate, seasonality, emergency orders, and facility case-mix. Gradient-boosted trees, random forest, and logistic regression were compared using temporal cross-validation. The best-performing model was then prospectively piloted for 16 weeks in 58 facilities, with weekly risk alerts sent to district pharmacy teams. Primary outcomes were 30-day stock-out prediction accuracy and stock-out days; secondary outcomes included emergency orders and reporting timeliness. |
| Results |
The gradient-boosted model achieved an AUROC of 0.87, sensitivity of 81%, specificity of 76%, and positive predictive value of 63% for 30-day stock-out risk. During the prospective pilot, alert facilities had 29% fewer stock-out days than matched controls (17.1 vs 24.2 days per facility-quarter), a 24% reduction in emergency orders, and an 11 percentage-point improvement in on-time reporting. The largest gains were observed for amoxicillin dispersible tablets, oxytocin, oral rehydration salts, and rapid diagnostic tests. |
| Discussion/Conclusion |
AI can add practical value to global health systems when it is built on routine data, linked to real operational decisions, and embedded in accountability workflows. Prediction alone is insufficient; impact comes when alerts trigger district-level action. In fragmented and resource-constrained settings, such tools can help translate evidence into timely implementation and more equitable PHC delivery. |
| Presenters and Affiliations |
Kassa Fentaw Economic PolIcy and Innovation Center for Health Systems |
