Proof-of-Concept Machine Learning Framework for Arboviral Disease Classification Using Literature-Derived Synthetic...

Area of Science:

• Medical Informatics
• Computational Biology
• Epidemiology

Background:

• Arboviral diseases (Dengue, Zika, Chikungunya) share vectors, geography, and symptoms, complicating diagnosis.
• Data scarcity poses a significant challenge for developing machine learning diagnostic tools for co-circulating arboviruses.

Purpose of the Study:

• To demonstrate the proof of concept for using synthetic data to establish computational feasibility for arboviral disease diagnosis.
• To guide future real-world validation efforts for machine learning diagnostic tools.

Main Methods:

• Assembled a synthetic dataset of 28,000 records (7,000 each for Dengue, Zika, Chikungunya, plus Influenza as control).
• Created a binary matrix of 67 symptoms for statistical analysis (Odds Ratios, Chi-Square).
• Trained and evaluated machine learning algorithms (MLP, NN, QSVM, BT) using performance metrics (accuracy, precision, sensitivity, specificity, F1-score, AUC-ROC, Cohen's kappa).

Main Results:

• The synthetic dataset's clinical relevance was validated against PAHO guidelines and existing arboviral databases.
• The Narrow Neural Network (NN) model achieved high performance: 0.92 accuracy, >0.98 AUC, >0.85 precision/sensitivity/specificity, and 0.89 Cohen's Kappa.
• The NN model reliably distinguished Dengue from Influenza, with slightly lower performance for Zika and Chikungunya differentiation.

Conclusions:

• Machine learning and deep learning models leveraging symptom features can accelerate early diagnosis of arboviral diseases.
• These models can serve as valuable support tools in resource-limited regions.
• The developed models do not replace, but augment, clinical medical expertise.