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

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

JOURNAL=Frontiers in Artificial Intelligence

VOLUME=Volume 4 - 2021

YEAR=2022

URL=https://www.frontiersin.org/journals/artificial-intelligence/articles/10.3389/frai.2021.554017

DOI=10.3389/frai.2021.554017

ISSN=2624-8212

ABSTRACT=In 2018, children under-five represented 67% of malaria deaths globally (WHO, 2020). Digital health technology represents an opportunity for improved access and quality healthcare in Low-to-Middle Income Countries (LMICs). Several electronic point-of-care (POC) clinical-decision support algorithms (CDSAs) have been developed based on the WHO’s Integrated Management of Childhood Illnesses (IMCI) protocol to guide frontline health workers (FHWs) through the clinical assessment, triage, and treatment of children under-5 years of age. THINKMD developed a progressive web-based, integrated POC clinical assessment, triage, treatment, and recommendation platform that incorporates IMCI protocols and other evidenced-based medicine. Here we present field-based feasibility and utilization studies in Kano, Nigeria focused on using the THINKMD mHealth platform with an embedded integrated malaria rapid diagnostic test (mRDT) to manually and via the development of machine learning (ML) algorithms improve the POC screening of febrile children in Nigeria. Using paired results with clinical health data and malaria risk assessment data from over 555 patients seen at five different health clinics, we trained a ML algorithm to identify malaria cases using symptom and location data. Among 480 children identified as at risk of malaria following IMCI guidelines, 66.7% had positive mRDT results and 33.3% had negative mRDT results.  The overall sensitivity and specificity of THINKMD mHealth platform (field-test) was 43% and 64%, respectively, and the malaria PPV was 71% compared with 67%. Confirmatory mRDT results and symptomology data then served as the basis for generating new algorithms both manually and through a supervised ML random forest approach, the later using 80% of our field-based data as the ML training set and 20% to test our new ML logic. New ML-based malaria clinical algorithms showed an increased sensitivity and specificity of 60% and 79%, respectively, with a PPV and NPV of 76% and 65%. Results confirm that incorporation of mRDT results into digital mHealth POC platforms can improve assessment of children with fever and can be used in the development of more accurate machine-learning-based frontline clinical assessment algorithms to improve identification of malaria/non malaria attributable febrile illnesses.