Data-Efficient LSTM Modeling for Climate-based Dengue Early Warning in Lampung, Indonesia

Fauzi, Rifky (Institut Teknologi Sumatera)
Sinaga, Mia Syntia Br (Institut Teknologi Sumatera)
Rizka, Nela (Institut Teknologi Sumatera)
Noor, Dear Michiko Mutiara (Institut Teknologi Sumatera)
Pribadi, Aswan Anggun (Institut Teknologi Sumatera)
Edriani, Tiara Shofi (Institut Teknologi Sumatera)

Zero : Jurnal Sains, Matematika, dan Terapan
Vol 9, No 2 (2025): Zero: Jurnal Sains Matematika dan Terapan

We present a data-efficient recurrent framework for climate-informed dengue early warning in Lampung Province. Monthly incidence and climate records are transformed into supervised sequences with 2–3-month lags, consistent with the observed lead–lag structure. Three architectures i.e. single-layer LSTM, stacked LSTM, and Temporal-Attention LSTM (TA-LSTM) are tuned via a compact genetic search under a time-ordered split. Performance improves with longer history; the TA-LSTM (37 units) attains the best accuracy. Permutation feature importance reveals a clear hierarchy: relative humidity and maximum temperature dominate, autoregressive incidence contributes moderately, while rainfall, sunshine, and minimum temperature are secondary; average temperature is largely redundant. The findings indicate that adding meaningful historical context and selective temporal weighting yields robust early-warning capability from coarse, time-limited data, and that humidity–temperature dynamics, together with short-term incidence persistence, are the principal drivers in this provincial setting.