Search

Poor access to diagnostic testing in resource limited settings restricts surveillance for emerging infections, such as dengue virus (DENV), to clinician suspicion, based on history and exam observations alone. We investigated the ability of machine learning to detect DENV based solely on data available at the clinic visit. We extracted symptom and physical exam data from 6,208 pediatric febrile illness visits to Kenyan public health clinics from 2014-2019 and created a dataset with 113 clinical features. Malaria testing was available at the clinic site. DENV testing was performed afterwards. We randomly sampled 70% of the dataset to develop DENV and malaria prediction models using boosted logistic regression, decision trees and random forests, support vector machines, naïve Bayes, and neural networks with 10-fold cross validation, tuned to maximize accuracy. 30% of the dataset was reserved to validate the models. 485 subjects (7.8%) had DENV, and 3,145 subjects (50.7%) had malaria. 220 (3.5%) subjects had co-infection with both DENV and malaria. In the validation dataset, clinician accuracy for diagnosis of malaria was high (82% accuracy, 85% sensitivity, 80% specificity). Accuracy of the models for predicting malaria diagnosis ranged from 53-69% (35-94% sensitivity, 11-80% specificity). In contrast, clinicians detected only 21 of 145 cases of DENV (80% accuracy, 14% sensitivity, 85% specificity). Of the six models, only logistic regression identified any DENV case (8 cases, 91% accuracy, 5.5% sensitivity, 98% specificity). Without diagnostic testing, interpretation of clinical findings by humans or machines cannot detect DENV at 8% prevalence. Access to point-of-care diagnostic tests must be prioritized to address global inequities in emerging infections surveillance.

From 1975-2009, the WHO guidelines classified symptomatic dengue virus infections as dengue fever, dengue hemorrhagic fever, and dengue shock syndrome. In 2009 the case definition was changed to a clinical classification after concern the original criteria was challenging to apply in resource-limited settings and not inclusive of a substantial proportion of severe dengue cases. Our goal was to examine how well the current WHO definition identified new dengue cases at our febrile surveillance sites in Kenya. Between 2014 and 2019 as part of a child cohort study of febrile illness in our four clinical study sites (Ukunda, Kisumu, Msambweni, Chulaimbo) we identified 369 dengue PCR positive symptomatic cases and characterized whether they met the 2009 revised WHO diagnostic criteria for dengue with and without warning signs and severe dengue. We found 62% of our PCR-confirmed dengue cases did not meet criteria per the guidelines. Our findings also correlate with our experience that dengue disease in children in Kenya is less severe as reported in other parts of the world. Although the 2009 clinical classification has recently been criticized for being overly inclusive and non-specific, our findings suggest the 2009 WHO dengue case definition may miss more than 50% of symptomatic infections in Kenya and may require further modification to include the African experience.