Urban Spatiotemporal Intelligence for Predicting Recorded ECU911 Incident Volume and Local Hotspot Dynamics in Intermediate Territories

Administrative emergency records can support spatiotemporal analysis in territories that are rarely represented in urban-computing research, but recorded volume is not equivalent to population risk or underlying incidence. This study develops an integrated framework for characterizing and predicting ECU911 recorded incident volume in Loja and Zamora Chinchipe, Ecuador, from January 2015 to 30 June 2026. The analysis used 925,000 administrative rows; 905,038 fell within the prespecified study period, and 899,232 were retained for spatial analysis after date and coordinate control. Annual spatial analysis used a 1 km grid with 3414 active cells; k = 10 nearest neighbours, including self-neighbours; row-standardized weights; conditional permutations; and Benjamini–Hochberg adjustment. The main analysis used 999 permutations; additional sensitivity analyses used 199 permutations. Global Moran’s I ranged from approximately 0.461 to 0.524 (permutation p = 0.001). Local Moran’s I produced no significant clusters after adjustment, whereas Gi* identified a small number of local hotspots. Across the complete years 2015–2025, no cell was identified as a hotspot in at least 50% of years; temporal classes comprised 18 occasional, 7 consecutive, 4 sporadic, 1 new, and 3384 cells, with no pattern detected. Leakage-free random forest regression achieved MAE = 0.552, RMSE = 2.230, RMSLE = 0.285, and R2 = 0.982 in the temporal test, but for pooled spatial block cross-validation, these rates declined to MAE = 0.797, RMSE = 10.423, and R2 = 0.661. A high-count classifier using the upper quintile of positive monthly cell counts achieved F1 = 0.883, ROC–AUC = 0.998, and PR–AUC = 0.966 in the temporal test. The contribution lies in the dataset, the understudied territory, and the methodological integration rather than in algorithmic novelty.