Multi-Cohort Machine Learning for Dengue Detection Using Routine Complete Blood Count Parameters

Abstract

Diagnosis of dengue is hard in resource constrained environments and It has suggested CBC-based machine
learning models as available alternatives. but are seldom externally tested. In this paper, we present a detailed
A comparison of eight ML models to detect dengue with a large cohort. n= 1,523; 18 CBC features) with
external validation on an independent. Cohort COVID-19 febrile illness (n=103). It was fair that there should be
internal discrimination. (best AUC = 0.698; 95% CI= 0.622−0.768), and external validation revealed catastrophic
generalization failure: tree-based ensembles and SVM were able to get. specificity = 0.000 on COVID-19 patients,
and it was only Naive Bayes that was able to reach. perfect specificity= 1.000. A three-class differentiation diagnosis
system. enhanced its performance to 78.2% accuracy and 0.786 macro − F1. Fairness auditing exhibited gross
age disparities (specificity= 0.171 − 0.568), and a three-feature CBC panel (Platelet Count, WBC, RDW-CV) had
AUC 0.595. These evidence demonstrates that clinical can only be achieved through internal measures. validity
and external validation, calibration, and fairness auditing should. be compulsory in dengue ML studies