A Seasonally Forced Compartmental Model for Dengue Transmission in Brazil: Real Data Calibration and Cross Country Transfer Testing, with Narrative Context on Other Neglected Tropical Diseases in the Americas

Background. Neglected tropical diseases (NTDs) impose a disproportionate burden on impoverished communities throughout the Americas. Prior compartmental frameworks in this space are rarely fitted to, and transfer tested on, real multi country surveillance time series with transparently reported goodness of fit; this paper focuses on dengue, for which such time series are available, and reports its findings without inflating fit quality or transfer success.Methods. A seasonally forced extension of a susceptible exposed infectious partially immune recovered vector coupled (SEIZR V) model was fitted to real weekly national dengue case counts for Brazil (2016 to 2019; n=209 weeks) drawn from the OpenDengue global database, which itself compiles and standardises Pan American Health Organization (PAHO) PLISA surveillance records. Parameters were estimated by maximum a posteriori (MAP) optimisation with a Laplace (local quadratic) approximation to posterior uncertainty, then cross checked by a local adaptive Metropolis Markov chain Monte Carlo run initialised at the MAP with the Laplace covariance as proposal, which reached acceptable convergence (Rhat less than or equal to 1.10 on all five fitted parameters) and recovered posterior means within 0.5 SE of the Laplace estimates, providing a genuine, if partial, cross validation of the approximation, given that full Hamiltonian No U Turn sampling via PyMC could not be executed in the available single core, offline computational environment. The fitted transmission parameters were then transferred, without re estimation, to three held out countries (Colombia, Bolivia, Peru; 2021 to 2022, n=104 weeks each) to test out of sample generalisation, re estimating only a country specific reporting scale parameter.Results. The fitted model captured the order of magnitude and part of the seasonal timing of Brazilian weekly dengue incidence, but the fit was modest (Pearson r=0.49; MAPE=145%; roughly a quarter of the variance explained), reflecting real inter annual epidemic variability that a single annual harmonic cannot capture; a two harmonic extension, tested directly, did not improve the fit (delta r squared equals negative 0.03), indicating that the residual variability is more plausibly serotype or immunity driven than a matter of seasonal forcing misspecification. Direct parameter transfer to Bolivia and Peru showed weak to moderate positive concordance (r=0.18 and 0.57, respectively), while transfer to Colombia failed outright (r=negative 0.46). Brazil's fitted seasonal phase does not generalise across all Latin American transmission settings, and this failure is reported as a substantive result rather than minimised. A global variance based (Sobol) sensitivity analysis of the basic reproduction number, which does not depend on the case count fit, identified the mosquito biting rate as the dominant driver of R0 variance (S1=0.49), consistent with prior vector borne disease literature.Conclusions. The central, load bearing finding of this paper is not a validation success but a demonstrated, quantified limit. Real data fitting and cross country transfer testing surface disease and setting specific generalisation failures that synthetic data exercises cannot reveal, and any model of this class intended for operational early warning use should be transfer tested, not assumed transferable, before deployment in a new country.