Predictive model for Interruption in Treatment in Patient Treatment Response Dashboard (Nigeria)

Nigerian Medical Records System (NMRS) Collaborative Development Team, Nigeria (September 2023)

Countries: Nigeria

Intervention Description

The Nigerian Medical Records System (NMRS) Collaborative Development Team, consisting of various partners, implemented an updated patient dashboard, referred to as the Patient Treatment Response Dashboard, that is integrated in the NMRS. This updated dashboard includes a predictive model for Interruption in Treatment (IIT) that utilizes machine learning, logistic regression, Python, and NMRS data elements. Reasons identified for a having a predictive model for IIT include helping with clinical decision making and proactive patient intervention; maintaining high retention rates and adherence to therapies for clinical trial research outcomes; having continuous treatment and monitoring; decreasing costs and resources associated with missed appointments or treatments; and preventing the emergence of resistant infectious disease strains resulting from treatment interruptions.

Presenter: Gibril Gomez, PHIS3 and Nigerian Medical Records System (NMRS) Collaborative Development Team

DUC Meeting Recording

Intervention Details

Data Elements	Selected NMRS data elements from the National Data Repository (NDR) predicting the outcome of interest – Current Status: TB Status Last drug regimen ART Delay First HIV Diagnosis date – antiretroviral therapy (ART) start date Functional status Occupation Marital status Education level Gender Is surge site? Last recorded weight (in kilogram) Current viral load (in copies/mL) Current age (in years) Duration on treatment ART start date – current date (active) Patient Treatment Response Dashboard contains Demographic Clinical Enrollment in Programs Recent Visits Flags General Actions Graph Fingerprints
Factors To Scale	Adding the IIT component to dashboard Challenges: Using R for the predictive model on NRMS Extracting data to CSV Older technology resulting in slow processing Waiting a long time for facility staff to adopt the updated patient dashboard
Evidence	Background: There has been growth in HIV+ persons receiving antiviral therapy, with UNAIDS reporting that 27.5 million individuals were receiving HIV treatment in 2020 as opposed to 7.8 million in 2010. In Nigeria there are 2,435 health facilities and 1,905,786 patients actively on treatment Long term retention remains an obstacle to achieving the 95-95-95 targets by 2030. Initial few months of HIV therapy have the most IIT, loss to follow-up (LTFU), and other pain sites. Costs: Partners had to upgrade existing servers to have higher specifications and capacity Model building and integration did not have extra costs Successes: 6 partners and 22-23/ out of 36-37 states Utilizing Python for the predictive model Utilizing JavaScript Object Notation (JSON) for data extraction Automated processes Relevant patient information listed Chart with IIT number using 28-day threshold; pill surplus; doses remaining; regimens; color coded flags: green (normal), yellow (warning), red (immediate action needed) Impact: Clinicians can make instant patient management decisions Flags allow prompt intervention Commodity monitoring due to pill balance shown Drastic IIT reduction in facilities utilizing the Patient Treatment Response Dashboard seen in recent NDR reports
Technology Requirements / Interoperability	NMRS (data extraction, prediction, processing, rendering results on Patient Treatment Response Dashboard) Connections with all facilities and communities Machine learning Linear regression Python SQL JSON Deployed on the client browser -> client request -> OpenMRS web server (JSON) -> MySQL data base Data sets to be processed -> Python Daemon -> Python analysis script -> MySQL data base
Calculations / Algorithms	Logistic regression machine learning predictive model utilizing NMRS data elements to predict IIT
Implementation Considerations	Size and quality of datasets impact on the computing feasibility of some methods Quality concerns and the impact of missing and/or imbalanced data Trade-off of interpretability and accuracy Computational complexity when deploying the model real time Verifying the algorithm aligns with the problem type Choosing a model – looked at Logistics Regression, Random Forest, and XGBoost Flags and reminders Stakeholders: Federal Ministry of Health; Agencies (CDC, USAID, DOD, AHF); System/Business Analysts Scientists; Software Engineers; Developers; Testers; Database Administrators; Monitoring/Surveillance; Implementing Partners; Facility Staff
Governance Considerations	Measures taken due to patient data to ensure optimal security and followed global standards