Machine Learning to Predict Interruption in Treatment (Mozambique)

Data.FI, Mozambique (September 2021)

Countries: Mozambique

Intervention Description

Data for Implementation team shared their work with predictive model as part of a software solution connected to OpenMRS, the EMR used at ECHO-supported facilities. They will be creating a software plugin to generate patient risk scores through the EMR.

Presenter: Yoni Friedman, Data.FI

DUC Meeting Recording

Intervention Details:

Data Elements	Demographics Medical history Publicly available data source
Evidence	Model showed strong predictive power achieving 0.65 AUC-PR compared to an underlying IIT rate of 23%. Learning activities are planned to track several ways in which this project can impact outcomes such as increasing intensity of interventions at high risk patients.
Technology Requirements / Interoperability	Machine learning Predictive model connected to OpenMRS.
Calculations / Algorithms	Precision and recall is used to evaluate model performance.
Factors to Scale	ECHO opted to generate risk prediction in raw form. A challenge will be to properly socialize the intrepretation of prediction. Factors to reach scale – Scalling the intervention include installing the software at additional facilities and should not be technically difficult. Preparing for updates to OpenMRS will be part of scaling plan.
Implementation Considerations	Contextual factors help to boost predictive accuracy
Governance Considerations