The capability to optimally regressively forecast pulmonary tuberculosis (TB) in Florida is demonstrated. We develop analytical approaches for accurately assessing exogenous forces of infection while accounting for individual, county-level characteristics that may affect susceptibility to disease and co-factors that may affect transmissibility. An empirical georeferenced epidemiological dataset of geographical, spatiotemporal (henceforth, geo-spatiotemporal), diagnostic, stratified, descriptive, TB-related, explanatory, discrete integer values was acquired from the Center of Disease Control (CDC). A non-homogeneous, gamma-distributed, mean, negative, binomial regression was constructed to treat Poisson overdispersion in the determinants in scipy.stats.poisson. A Newton's method was employed to find zeroes of a function \(g\) of the stratified CDC variables which was provided by: \(x_{n + 1} = x_{n} - \left[ {J_{g} \left( {x_{n} } \right)} \right]^{ - 1} g\left( {x_{n} } \right)\) in scipy.optimize.newton while \(\left[ {J_{g} \left( {x_{n} } \right)} \right]^{ - 1}\) evaluated the model outcome for \(\left( {x_{n} } \right)\) . Newton's method found zeroes of a function \(g\) in the stratified TB data as generated by: \(x_{n + 1} = x_{n} - \left[ {J_{g} \left( {x_{n} } \right)} \right]^{ - 1} g\left( {x_{n} } \right)\) where \(\left[ {J_{g} \left( {x_{n} } \right)} \right]^{ - 1}\) was the left inverse of the matrix \(J_{g} \left( {x_{n} } \right)g\) which was evaluated for \(x_{n}\) . Broyden's Method solved for F(x) = 0. We computed the whole Jacobian only at the first iteration and conducted rank-one update at the other iterations. The secant method was found to be a non-robust estimator. Subsequently, we introduced a novel non-linear Kalman filter that employed semi-parametrization gradients for optimization of the CDC data. The semi-parametric approximation was based on a joint Gaussian assumption, which in this experiment was equivalent to minimizing an approximation to the Kullback-Leible divergence. Since residual, zero, autocorrelation coefficients may have provoked non-Gaussianism in the model output due to violations of regression assumptions [ie., multicollinearity] in space, time, and geography, we employed Markovian, semi-parametric, eigen-Bayesianism and a conditional heteroscedastic [GARCH] paradigm in PyMC3 for robustification of random geographical chaos in the prognosticated, TB-related, eigen-spatial, filtered estimated determinants. In the non-frequentistic geo-spatiotemporal, residual analysis we specified a prior uncertainty about the model parameters. The forecasts revealed the temporal error variance in the residuals. We generated a diagnostic summary report which revealed levels of heteroscedasticity [i.e., unequal scattered distribution of residuals], and /or inflated standard errors in the regressors in eigenvector eigen-Bayesian eigen-geospace. The fit of the final model revealed that the adjusted estimator determinant ‘Personal Income < 20,000’ was statistically important to the county-level, capture point, sentinel site scalable geolocations (adjusted improvement \(\chi^{2}\) was—1.299). Diagnostically stratifiable. non-Gaussian, TB-related, temporality, multicollinear, and or zero autocorrelated model heteroscedastic biasness due to violations of regression assumptions may be detectable and treatable in eigen-Bayesian, eigenvector eigen-geospace. Our final model revealed that from 2025 to 2030, Duval, Orange, and Broward counties would require immediate intervention to prevent TB transmission. The model also revealed that from 2035 to 2050, Hillsborough and Palm Beach counties could become hyperendemic without the implementation of control strategies. All non-Gaussian semiparametric, Markovian, space time and geography compatibility tests in this experiment were performed in Python 3.8, however, the eigen-spatial filter algorithms are executable on MacOS 11.3 and Linux Ubuntu Server 20.04 LTS environments.