News thumbnail
Science / Thu, 19 Nov 2026 Nature

A Caputo fractional tuberculosis model with saturated incidence: stability analysis, optimal control, and machine learning forecasting

We formulate a Caputo fractional-order TB model with vaccination, a saturated (Holling type-II) incidence that combines person-to-person and environmental transmission, and an explicit isolation compartment. For a Khyber Pakhtunkhwa-scale parameterisation (\({\mathcal {R}}_0\approx 3.79\)), \({\mathcal {R}}_0\) is most sensitive to the transmission rate, the progression rate from latency and the treatment-initiation rate. Isolation at rate \(q=0.2\) \(\hbox {year}^{-1}\) lowers \({\mathcal {R}}_0\) from 4.72 to 3.79 and almost halves the active TB burden after 70 years. Combining all four controls minimises the objective functional and reduces active TB at 50 years by about \(92\%\); omitting awareness is the most costly omission. A biologically consistent model structure, rigorous threshold results and a clearly delimited data-driven layer together give a transparent tool for TB control planning.

Tuberculosis (TB) is still the leading cause of death from a single infectious agent, and high-burden settings need models that are biologically faithful, analytically tractable, and usable for forecasting and intervention planning. Most fractional-order TB studies address only one of these aspects. We formulate a Caputo fractional-order TB model with vaccination, a saturated (Holling type-II) incidence that combines person-to-person and environmental transmission, and an explicit isolation compartment. Every transition follows the natural history of TB: latently infected individuals can only progress to active disease or die; isolated individuals are active cases who either start treatment, leave isolation untreated, or die; and recovery occurs only through treatment completion or documented spontaneous self-cure, with relapse. We prove positivity and boundedness, derive the reproduction number \({\mathcal {R}}_0\) in closed form with the next-generation method, verify the Matignon stability condition with computed Jacobian eigenvalues, and prove global stability of the disease-free equilibrium with an explicitly constructed linear Lyapunov function. We also prove well-posedness of the Caputo system and the existence of an endemic equilibrium when \({\mathcal {R}}_0>1\), and we give an explicit condition for the bifurcation at \({\mathcal {R}}_0=1\) to be forward. We then pose an optimal-control problem with four controls (awareness, vaccination, active case finding with isolation, and environmental decontamination), derive the exact adjoint system and control characterisations from Pontryagin’s maximum principle, and couple the mechanistic model with a NARX neural network trained by the Levenberg–Marquardt algorithm. For a Khyber Pakhtunkhwa-scale parameterisation (\({\mathcal {R}}_0\approx 3.79\)), \({\mathcal {R}}_0\) is most sensitive to the transmission rate, the progression rate from latency and the treatment-initiation rate. Isolation at rate \(q=0.2\) \(\hbox {year}^{-1}\) lowers \({\mathcal {R}}_0\) from 4.72 to 3.79 and almost halves the active TB burden after 70 years. Memory effects (\(\alpha <1\)) markedly slow the epidemic. Combining all four controls minimises the objective functional and reduces active TB at 50 years by about \(92\%\); omitting awareness is the most costly omission. The NARX surrogate achieves 1.4–\(2.3\%\) one-step-ahead error on noisy synthetic data, whereas fully recursive multi-step forecasts drift by \(12\%\) on average over a 10-year horizon. A biologically consistent model structure, rigorous threshold results and a clearly delimited data-driven layer together give a transparent tool for TB control planning. The results identify early case detection, treatment initiation and contact reduction as the most effective levers.

© All Rights Reserved.