Decentralizing Dengue Diagnostics: Federated Knowledge Distillation Across Heterogeneous Modalities via Edge Transpilation

The escalating severity of Dengue fever outbreaks
in Bangladesh necessitates rapid, data-driven clinical diagnostic
tools. While recent machine learning architectures demonstrate
high predictive accuracy using clinical and meteorological data,
their reliance on centralized cloud infrastructure and heavy
Python backends introduces severe data privacy vulnerabilities and renders them impractical for low-bandwidth, rural
healthcare centers. This paper proposes a decentralized, privacypreserving mobile diagnostic framework utilizing Heterogeneous
Federated Transfer Learning (HFTL). To the best of our knowledge, this is the first federated framework for dengue detection to collaboratively synthesize three distinct data modalities:
haematological parameters, subjective symptomatic profiles, and
environmental indicators. Furthermore, we introduce a novel
edge-transpilation pipeline that mathematically converts trained
balanced Random Forest classifiers directly into zero-dependency,
native Dart code. This approach completely decouples diagnostic
inference from the cloud, enabling zero-latency, 100% offline
execution directly on mobile edge devices. Experimental evaluations confirm exceptional diagnostic performance, achieving
96.0% accuracy on structural symptom vectors and 76.4% on
haematological markers. By ensuring that sensitive patient health
data never leaves the local client device, this framework bridges
the critical gap between robust computational epidemiology
and secure, deployable digital health infrastructure in resourceconstrained environments.