DengueAI Comparative Time-Series and Deep-Learning Modeling for Early Prediction of Dengue Outbreaks in Rural Peru Using Open Data

Quispe M.F.M.; Gonzales L.A. · Proceedings of the 2025 IEEE 32nd International Conference on Electronics, Electrical Engineering and Computing, INTERCON 2025 (2025) · DOI: 10.1109/intercon67304.2025.11244382

Resumen

Timely forecasting of dengue outbreaks is crucial for directing limited public-health resources in rural, low-income regions. We present DengueAI, an open-source framework that benchmarks classical time-series models against deep-learning architectures using > 25 years of nationwide open data from Peru. Weekly dengue notifications (1999-2024) were enriched with meteorological covariates and analysed with four families of predictors: ARIMA, Prophet, Long Short-Term Memory (LSTM) and Bidirectional LSTM (BiLSTM). In short-range horizons (1-4 weeks) the best ARIMA configuration achieved R2=0.9707 and MAE = 88.7 cases, outperforming neural networks by 5-8 %, whereas LSTM models captured medium-term, non-linear trends when multiple climate variables were included. All models are deployed in a lightweight Streamlit dashboard that generates forecasts, heat maps and downloadable artefacts for local health offices. The study confirms that well-tuned statistical baselines can equal or surpass deep networks in data-scarce settings, while hybrid strategies offer a practical path toward early-warning systems in underserved areas.