Development and Validation of Manually Modified and Supervised Machine Learning Clinical Assessment Algorithms for Malaria in Nigerian Children.

Megan Mclaughlin,Aisha Giwa,Samuel V Scarpino,Ezra Mount-Finette,Sabine Dittrich,Karell G Pellé,Nirmal Ravi,Adam Thompson,Fatima Adamu,Barry Finette,Barry Heath,Nada Haidar

doi:10.3389/frai.2021.554017

Abstract

It is currently estimated that 67% of malaria deaths occur in children under-five years (WHO, 2020). To improve the identification of children at clinical risk for malaria, the WHO developed community (iCCM) and clinic-based (IMCI) protocols for frontline health workers using paper-based forms or digital mobile health (mHealth) platforms. To investigate improving the accuracy of these point-of-care clinical risk assessment protocols for malaria in febrile children, we embedded a malaria rapid diagnostic test (mRDT) workflow into THINKMD’s (IMCI) mHealth clinical risk assessment platform. This allowed us to perform a comparative analysis of THINKMD-generated malaria risk assessments with mRDT truth data to guide modification of THINKMD algorithms, as well as develop new supervised machine learning (ML) malaria risk algorithms. We utilized paired clinical data and malaria risk assessments acquired from over 555 children presenting to five health clinics in Kano, Nigeria to train ML algorithms to identify malaria cases using symptom and location data, as well as confirmatory mRDT results. Supervised ML random forest algorithms were generated using 80% of our field-based data as the ML training set and 20% to test our new ML logic. New ML-based malaria algorithms showed an increased sensitivity and specificity of 60 and 79%, and PPV and NPV of 76 and 65%, respectively over THINKD initial IMCI-based algorithms. These results demonstrate that combining mRDT “truth” data with digital mHealth platform clinical assessments and clinical data can improve identification of children with malaria/non-malaria attributable febrile illnesses.

Highlights

It is currently estimated that every two minutes a child under the age of five years dies from malaria
We demonstrate the development and testing of new supervised machine learning (ML) malaria-risk algorithms, using field-based malaria-risk assessments, clinical data, and malaria rapid diagnostic test (mRDT) diagnostic data generated via the use of THINKMD’s Integrated Management of Childhood Illness (IMCI) compliant mobile health (mHealth) platform, that could improve the identification of malaria in children with febrile illnesses by FHWs
Integrated Management of Childhood Illness (IMCI) designation of malaria risk for children were based on Nigeria’s standard IMCI protocol for febrile illness assessment, any child presenting with a history of fever OR having a current fever (>37.5°C) based on a thermometer-based measurement would be determined to be at risk for malaria and receive an mRDT

Summary

Introduction

It is currently estimated that every two minutes a child under the age of five years dies from malaria. A multiple indicator cluster Nigeria survey revealed that only 63.4% of children with a history of fever sought care from a health facility or provider. Of those who presented for care, only 13.8% received a malaria diagnostic test, with only 36.8% receiving antimalarial treatment, 20.6% being artemisinin-based combinations therapies (ACTs). For children testing positive for malaria, only 9.1% were given ACT treatment, while 29.4% were given an antibiotic (Fund, 2017). These findings indicate that to improve mortality and morbidity of febrile children with malaria, there needs to be a significant increase in quality clinical risk assessment screening of children with fever linked with diagnostic testing and improved appropriate therapeutic intervention for children with positive malaria diagnostic tests

Methods

Results

Discussion

Conclusion