Early detection of dengue outbreaks from routine primary health care records: evaluation by an extended version of the Mixed Model of Artificial Intelligence and Next-Generation (MMAING)

A recently proposed surveillance framework is used to detect arbovirus outbreaks. It mixes traditional compartmental models with artificial intelligence techniques applied to large volumes of unconventional data on syndromic events, environment, and social information, and its performance is evaluated in the early identification of dengue epidemic events in Brazil at municipal level. We evaluated an extended version of the MMAING model, referred to as MMAING-Arbo , developed to enable the early detection of dengue outbreaks based on primary health care (PHC) encounters associated with specific arboviral syndromes at the municipal level in Brazil. The analysis covered the period from January 1, 2023, to July 1, 2024, encompassing 2,767 municipalities (49.68% of all municipalities in the country) and identifying a total of 3,478 epidemic events of varying duration observed in the PHC data. The results obtained with MMAING-Arbo indicate that the proposed approach consistently issues early warning signals (EWS) for a high proportion of identified dengue outbreaks. The comparative evaluation with EARS demonstrated that MMAING-Arbo performs comparably to or outperforms the EARS variants (C1, C2, and C3) across different epidemiological contexts, with performances varying according to municipal population size. In particular, MMAING-Arbo showed a lower alarm burden per outbreak in larger municipalities, while the EARS-C2 variant achieved higher sensitivity and timeliness in smaller ones. The analysis of PHC data for multiple mosquito-borne diseases using MMAING-Arbo yields promising results and indicates it might be a valuable tool to issue EWS of emerging outbreaks in real-life conditions. These findings reinforce the importance that future surveillance systems include data-driven approaches that benefit from the integration of multiple models, strengthening sensitivity, specificity, and timeliness of EWS.