P-2006. Development of a Machine Learning Model for the Accurate Diagnosis of Common Tropical Febrile Illnesses

TL;DR

This study develops an AI-based tool to improve the diagnosis of tropical febrile illnesses like dengue and malaria by using machine learning models trained on clinical data.

Contribution

The novel contribution is the development of an AI diagnostic tool using feature selection and stacking classifiers to improve accuracy in tropical febrile illness diagnosis.

Findings

A stacking classifier achieved 89% accuracy in diagnosing tropical febrile illnesses.

Random Forest showed the highest individual model accuracy at 87%.

Twenty non-collinear clinical features were selected for model development.

Abstract

Acute febrile illnesses (AFIs) such as dengue, malaria, scrub typhus, leptospirosis etc, prevalent in tropical regions, account for 17% of the global disease burden. Overlapping clinical features and limitations of current diagnostic methods—including false positives, sensitivity/specificity issues made the diagnosis complicated. With increasing digital integration in healthcare, artificial intelligence (AI) offers a promising solution for improving diagnostic accuracy and efficiency. Our study aimed to develop an AI-based tool to aid differential diagnosis of AFIs and support clinical decision-making.Fig 1:Features selected through Recursive Elimination and MulticollinearityBased on Recursive_Feature_Elimination (RFE), a set of high-ranking features was initially identified. These features were then subjected to multicollinearity assessment to eliminate redundant variables with strong linear relationships. The final set of 20 non-collinear, clinically relevant features was selected for model development and is presented in Fig_1Fig 2:Accuracy of the developed models

Features selected through Recursive Elimination and Multicollinearity

Based on Recursive_Feature_Elimination (RFE), a set of high-ranking features was initially identified. These features were then subjected to multicollinearity assessment to eliminate redundant variables with strong linear relationships. The final set of 20 non-collinear, clinically relevant features was selected for model development and is presented in Fig_1

Accuracy of the developed models

A retrospective cross-sectional study analyzed records of 800 patients (200 per disease). Clinical data were extracted and preprocessed (cleaning, scaling, imputation). Feature selection was conducted using Recursive_Feature_Elimination (RFE) and multicollinearity assessment. Models were developed using supervised machine learning algorithms— Random_Forest(RF), Naïve_Bayes(NB), Logistic_Regression(LR), Support_Vector_Machine(SVM), K-Nearest_Neighbors(KNN), and Decision_Tree(DT). A stacking classifier served as a meta-model. Performance was evaluated using accuracy, precision, recall, and F1-score.Fig 3:Confusion matrix for Stacking ClassifierTable 1:Other performance Metrics

Confusion matrix for Stacking Classifier

Other performance Metrics

Based on RFE, set of high-ranking features were identified and were then subjected to multicollinearity assessment. The final set of 20 non-collinear, clinically relevant features was selected for model development (presented in Fig_1). Among the developed predictive models, the RF demonstrated highest classification accuracy (87%), followed by LR (86.5%), SVM (85.5%), KNN (73%), DT (72.5%), and NB (70.5%), (shown in Fig_2). Additional performance metrics, including precision, recall, and F1-score, are summarized in Table 1. A stacking classifier, integrating the predictions of all base models, achieved an overall accuracy of 89% on training dataset. The confusion matrix for the stacking model is presented in Fig_3

By integrating clinical data with advanced feature selection techniques, AI-based tool can be the future diagnostic aiding tool for screening the tropical diseases.

All Authors: No reported disclosures

Figures