pith. sign in

arxiv: 2606.02650 · v1 · pith:KTCZRB42new · submitted 2026-05-31 · 🧬 q-bio.QM

Using Machine Learning to Enhance Hyperparameter Optimization in Pandemic Modeling: Case study of COVID-19 Dynamics in Ghana

Thomas Izgin , Andreas Meister , Isaac Azure This is my paper

Pith reviewed 2026-06-28 15:40 UTC · model grok-4.3

classification 🧬 q-bio.QM
keywords COVID-19 modelingcompartmental modelsMPRK methodshyperparameter optimizationWENO reconstructionpandemic dynamicsGhana case studytime-dependent parameters
0
0 comments X

The pith

Reformulating COVID-19 models around common transmissions and solving them with MPRK methods produces hyperparameters that enable 5-day predictions within 10% error on Ghana data.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper examines five distinct COVID-19 models developed under different conditions and reformulates each one to isolate the shared transmission terms between compartments while keeping the original structure intact. Modified Patankar-Runge-Kutta methods are applied to the resulting nonlinear ODE systems to generate unconditionally positive numerical solutions that also preserve the conservative properties of the equations. These solutions are placed directly inside a cost function to refine the non-autonomous hyperparameters, first as piecewise constants fitted to data and then via WENO reconstruction to recover smooth time-dependent coefficients. The approach is demonstrated on a Ghana-specific model, where the resulting parameters support forward predictions five days ahead with error below 10 percent. Such short-term accuracy matters when parameters evolve and public-health decisions must be made from limited recent observations.

Core claim

By reformulating five distinct COVID-19 models to use their common transmission terms between compartments while preserving original structure, and then applying Modified Patankar-Runge-Kutta methods to obtain unconditionally positive solutions of the nonlinear ODEs that also preserve the conservative part, the numerical solutions can be incorporated into a cost function to improve estimates of the non-autonomous hyperparameters; piecewise constant parameters fitting the data are obtained first, followed by WENO reconstruction to approximate the true time-dependent coefficients, resulting in 5-day predictions within a 10% error range as a proof-of-concept on a Ghana COVID-19 model.

What carries the argument

Modified Patankar-Runge-Kutta (MPRK) methods applied to reformulated nonlinear ODE systems for producing unconditionally positive approximations that preserve the conservative part, with the solutions inserted into a cost function for hyperparameter optimization followed by WENO reconstruction of time-dependent coefficients.

If this is right

  • The reformulated models retain their epidemiological meaning and admit positive numerical solutions via MPRK schemes.
  • Piecewise constant parameters fitting observed data can be extracted for each of the five models.
  • WENO reconstruction supplies approximations to the true time-dependent coefficients inside the ODEs.
  • Five-day-ahead predictions remain within a 10 percent error range when the method is applied to the Ghana case.
  • The same workflow applies across models originally formulated under differing epidemiological conditions.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same restructuring-plus-MPRK-plus-WENO pipeline could be tested on compartmental models for other infectious diseases that share similar transmission structures.
  • The cost-function step could be replaced or augmented by standard machine-learning hyperparameter optimizers to search larger parameter spaces automatically.
  • Positive-preserving solutions may reduce spurious negative values that appear in long integrations with conventional ODE solvers.
  • The resulting time-dependent parameter trajectories could be used for ongoing model updates as fresh incidence data arrive.

Load-bearing premise

Reformulating the models via common transmissions preserves their original epidemiological meaning, and the MPRK solutions can be inserted directly into the cost function to obtain hyperparameters whose forward predictions stay accurate outside the data window.

What would settle it

Applying the optimized Ghana model forward for five days after the fitting window and measuring whether prediction error exceeds 10 percent would directly test the accuracy claim.

read the original abstract

In this study, five distinct COVID-19 models developed in different countries, each designed to reflect the prevailing epidemiological condition at the time of formulation, are examined. The models are reformulated while still maintaining their original structure, using their common transmissions from one compartment to the other. Modified Patankar-Runge-Kutta (MPRK) methods are then applied to approximate the solutions of the resulting system of nonlinear ordinary differential equations (ODEs) representing each model to produce unconditionally positive approximations and to preserve the conservative part of the ODEs. In particular, we incorporate the numerical solution into a cost function to improve the estimates for the non-autonomous model hyperparameters. In a first step we obtain piecewise constant parameters that fit real data. Later we perform a WENO reconstruction in a post-process to approximate the true time-dependent coefficients inside the ODEs. As a proof-of-concept, we apply our approach to improve the parameters of a paper concerned with modeling COVID-19 in Ghana, where we can make 5-day predictions within a 10% error range.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 1 minor

Summary. The manuscript examines five COVID-19 compartmental models, reformulates each around common transmission terms between compartments while preserving original structure, applies Modified Patankar-Runge-Kutta (MPRK) schemes to obtain unconditionally positive numerical solutions that also conserve the appropriate quantities, and inserts those numerical solutions directly into a cost function to optimize the non-autonomous (time-dependent) hyperparameters. Piecewise-constant parameter values are first obtained by fitting to data; a subsequent WENO reconstruction step is used to recover continuous time-dependent coefficients. As a proof-of-concept the procedure is applied to a published Ghana COVID-19 model, with the claim that the resulting system produces 5-day-ahead predictions lying within a 10% error bound.

Significance. If the out-of-sample predictive accuracy claim can be substantiated with transparent data, cost-function definition, fitting-window length, and baseline comparisons, the integration of positivity-preserving integrators with post-hoc reconstruction for hyperparameter tuning could supply a practical route to improved short-term forecasting in non-autonomous epidemiological models. The approach also illustrates how numerical-method properties (positivity, conservation) can be exploited inside an optimization loop rather than applied only after parameter estimation.

major comments (3)
  1. [Abstract] Abstract: the central claim that 5-day predictions lie 'within a 10% error range' supplies no information on the data source or time span, the exact form of the cost function, the length of the fitting window, any baseline comparator, or whether the 5-day forecasts are genuinely out-of-sample. Without these elements the numerical result cannot be evaluated.
  2. [Abstract] Abstract (and method description): parameters are obtained by minimizing a cost function whose integrand contains the MPRK numerical solution of the ODE itself. The subsequent 5-day predictions are therefore generated from quantities already tuned to the same data; the manuscript provides no separate validation set, rolling-window test, or comparison against a hold-out period that would demonstrate that the WENO-reconstructed time-dependent coefficients generalize beyond the fitting interval.
  3. [Abstract] Abstract: the reformulation step that extracts 'common transmissions' is asserted to leave the original epidemiological meaning intact, yet no explicit verification is offered that the reparameterized system remains biologically interpretable or that the MPRK-augmented cost function yields hyperparameters whose forward trajectories remain epidemiologically plausible outside the training window.
minor comments (1)
  1. [Abstract] The abstract refers to 'five distinct COVID-19 models developed in different countries' but does not list the models or their original citations; adding these references would improve traceability.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed report. The comments correctly identify areas where the abstract and presentation require expansion to substantiate the claims. We address each point below and will revise the manuscript accordingly to improve transparency and validation.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim that 5-day predictions lie 'within a 10% error range' supplies no information on the data source or time span, the exact form of the cost function, the length of the fitting window, any baseline comparator, or whether the 5-day forecasts are genuinely out-of-sample. Without these elements the numerical result cannot be evaluated.

    Authors: We agree that the abstract is insufficiently detailed. In the revised manuscript we will expand the abstract to specify the Ghana COVID-19 data source and study period, the exact definition of the cost function that embeds the MPRK solution, the fitting-window length, baseline comparators, and confirmation that the 5-day forecasts are generated by forward integration after the optimization interval. revision: yes

  2. Referee: [Abstract] Abstract (and method description): parameters are obtained by minimizing a cost function whose integrand contains the MPRK numerical solution of the ODE itself. The subsequent 5-day predictions are therefore generated from quantities already tuned to the same data; the manuscript provides no separate validation set, rolling-window test, or comparison against a hold-out period that would demonstrate that the WENO-reconstructed time-dependent coefficients generalize beyond the fitting interval.

    Authors: The observation is accurate: the present version does not include an explicit hold-out or rolling-window validation. We will revise the methods and results sections to add such an analysis for the Ghana case study, thereby demonstrating generalization of the WENO-reconstructed coefficients. revision: yes

  3. Referee: [Abstract] Abstract: the reformulation step that extracts 'common transmissions' is asserted to leave the original epidemiological meaning intact, yet no explicit verification is offered that the reparameterized system remains biologically interpretable or that the MPRK-augmented cost function yields hyperparameters whose forward trajectories remain epidemiologically plausible outside the training window.

    Authors: Because the reformulation preserves the original compartmental structure and merely rewrites transmission terms in equivalent form, the epidemiological meaning of the parameters is unchanged by construction. Nevertheless, we acknowledge the absence of explicit verification. We will add a short subsection (or supplementary note) that confirms biological interpretability and shows that forward trajectories with the optimized parameters remain epidemiologically plausible beyond the fitting window. revision: yes

Circularity Check

0 steps flagged

No significant circularity; standard parameter fitting and forward simulation

full rationale

The paper describes reformulating existing epidemiological models, applying MPRK solvers to the resulting ODE systems, embedding those numerical solutions inside a cost function to obtain piecewise-constant hyperparameters that match observed Ghana COVID-19 data, then applying WENO reconstruction to recover time-dependent coefficients. The 5-day forward predictions are generated by integrating the reconstructed non-autonomous system beyond the fitting interval. None of the enumerated circularity patterns is present: the fitted parameters are not defined in terms of the predictions, the predictions are not statistically forced to equal the fitted values by construction, and no self-citation or imported uniqueness theorem is invoked to close the argument. The reported 10 % error on the 5-day horizon is an empirical claim about out-of-sample behavior that can be falsified by new data and is therefore not circular.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The approach depends on the assumption that MPRK integrators preserve positivity and conservation for the reformulated epidemiological ODEs and that fitting piecewise constants followed by WENO reconstruction yields parameters whose short-term forecasts are meaningful.

free parameters (2)
  • piecewise constant transmission and recovery rates
    Obtained by minimizing the cost function that incorporates the MPRK solution against observed data.
  • WENO reconstruction parameters
    Chosen in post-processing to produce smooth time-dependent coefficients from the piecewise fits.
axioms (1)
  • domain assumption MPRK methods produce unconditionally positive approximations and preserve the conservative part of the ODEs.
    Invoked when the numerical solution is inserted into the cost function for hyperparameter estimation.

pith-pipeline@v0.9.1-grok · 5721 in / 1529 out tokens · 38870 ms · 2026-06-28T15:40:14.231280+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

36 extracted references · 14 canonical work pages

  1. [1]

    : Novel coronavirus (2019-ncov): Situation report, 17-erratum

    Organization, W.H., et al. : Novel coronavirus (2019-ncov): Situation report, 17-erratum. In: Novel Coronavirus (2019-nCoV): Situation Report, 17-Erratum, (2020)

    2019

  2. [2]

    Covid 1(1), 39–70 (2021)

    Voorn, T., Jong, M.: Cope or perish? managing tipping points in developing coping strategies for emergency response during the first wave of the covid-19 outbreak in europe. Covid 1(1), 39–70 (2021)

    2021

  3. [3]

    Journal of hospital infection 104(3), 246–251 (2020)

    Kampf, G., Todt, D., Pfaender, S., Steinmann, E.: Persistence of coronaviruses on inanimate surfaces and their inactivation with biocidal agents. Journal of hospital infection 104(3), 246–251 (2020)

    2020

  4. [4]

    Scientific Reports 15 (2025) https://doi.org/10.1038/ s41598-025-93994-3

    Ramalingam, R., Gnanaprakasam, A., Boulaaras, S.: Stability and control analysis of covid- 19 spread in india using seir model. Scientific Reports 15 (2025) https://doi.org/10.1038/ s41598-025-93994-3

    2025

  5. [5]

    Nonlinear Dynamics 113, 971–988 (2024) https://doi.org/10

    Li, C.-Z., Lü, X., Gong, J.-J., Lei, Y.: Extended seir model of covid-19 spread focusing on compartmental flow in england. Nonlinear Dynamics 113, 971–988 (2024) https://doi.org/10. 1007/s11071-024-09748-9

    2024

  6. [6]

    Frontiers in Digital Health6 (2024), 1427233

    Alkhalefah, H., Preethi, D., Khare, N., Abidi, M.H., Umer, U.: Deep learning infused sirvd model for covid-19 prediction: Xgboost-sirvd-lstm approach. Frontiers in Medicine 11 (2024) https://doi.org/10.3389/fmed.2024.1427239

    work page doi:10.3389/fmed.2024.1427239 2024

  7. [7]

    ArXiv (2025) https://doi.org/ 10.48550/arxiv.2501.09298

    Qian, Y., Marty, Basu, A., O’dea, E., Wang, X., Fox, S., Rohani, P., Drake, J.M., Li, H.: Physics-informed deep learning for infectious disease forecasting. ArXiv (2025) https://doi.org/ 10.48550/arxiv.2501.09298

    work page doi:10.48550/arxiv.2501.09298 2025

  8. [8]

    AIP Conference Proceedings (2023) https://doi.org/10.1063/5.0115345

    Shivalingappa, N.M.H., Krishna, P.M.: Prediction of covid-19 transmission in india using deep learning neural network. AIP Conference Proceedings (2023) https://doi.org/10.1063/5.0115345

    work page doi:10.1063/5.0115345 2023

  9. [9]

    ArXiv abs/2503.12813 (2025) https://doi.org/10.48550/arxiv.2503.12813

    Alizadeh, M., Samaei, M.H., Seilsepour, A., Beheshti, M.T.: Epidemic forecasting with a hybrid deep learning method using cnn-lstm with woa-gwo parameter optimization: Global covid-19 case study. ArXiv abs/2503.12813 (2025) https://doi.org/10.48550/arxiv.2503.12813

    work page doi:10.48550/arxiv.2503.12813 2025

  10. [10]

    International Journal of Biomathematics (2024) https://doi.org/10.1142/s1793524524500955

    Adewole, M.O., Okposo, N., Abdullah, F.A., Ali, M.K.M.: Covid-19 dynamics and mutation: Linking intra-host and inter-hosts dynamics via agent-based modeling approach. International Journal of Biomathematics (2024) https://doi.org/10.1142/s1793524524500955

    work page doi:10.1142/s1793524524500955 2024

  11. [11]

    Bioengineering 12 (2025) https://doi.org/10.3390/ bioengineering12050514

    Cheng, Y., Cheng, R., Xu, T., Tan, X., Bai, Y.: Machine learning techniques applied to covid-19 prediction: A systematic literature review. Bioengineering 12 (2025) https://doi.org/10.3390/ bioengineering12050514

    2025

  12. [12]

    G., Naranjo, S., Rideout, W

    Asamoah, J.K.K., Owusu, M.A., Jin, Z., Oduro, F.T., Abidemi, A., Gyasi, E.O.: Global stability and cost-effectiveness analysis of COVID-19 considering the impact of the environment: using data from Ghana. Chaos, Solitons & Fractals 140, 110103 (2020) https://doi.org/10.1016/j. chaos.2020.110103

    work page doi:10.1016/j 2020

  13. [13]

    Computational and Mathematical Methods in Medicine 2021(1), 1250129 (2021)

    Diagne, M., Rwezaura, H., Tchoumi, S., Tchuenche, J.: A mathematical model of covid-19 with vaccination and treatment. Computational and Mathematical Methods in Medicine 2021(1), 1250129 (2021)

    2021

  14. [14]

    Decoding Infection and Transmission 3, 100051 (2025)

    Santosh, C., Shakya, A.: A novel approach to disease modeling: The seivr model with a vulnerable compartment. Decoding Infection and Transmission 3, 100051 (2025)

    2025

  15. [15]

    AIMS Math 9(3), 6281–6304 (2024)

    Rattanakul, C., Chaiya, I.: A mathematical model for predicting and controlling covid-19 transmission with impulsive vaccination. AIMS Math 9(3), 6281–6304 (2024)

    2024

  16. [16]

    Mathematics 11(1), 142 (2022) 16

    Haq, I.U., Ullah, N., Ali, N., Nisar, K.S.: A new mathematical model of covid-19 with quarantine and vaccination. Mathematics 11(1), 142 (2022) 16

    2022

  17. [17]

    https://github.com/IzginThomas/COVID-19 (2026)

    Azure, I., Izgin, T., Meister, A.: Reproducibility repository for ”Using Machine Learning to Enhance Hyperparameter Optimazation in Pandemic Modeling: Case study of COVID-19 Dynamics in Ghana” . https://github.com/IzginThomas/COVID-19 (2026). https://doi.org/10. 5281/zenodo.20298147

    2026

  18. [18]

    Burchard, H., Deleersnijder, E., Meister, A.: A high-order conservative Patankar-type discreti- sation for stiff systems of production-destruction equations. Appl. Numer. Math. 47(1), 1–30 (2003) https://doi.org/10.1016/S0168-9274(03)00101-6

    work page doi:10.1016/s0168-9274(03)00101-6 2003

  19. [19]

    Applied Numerical Mathematics 123, 159–179 (2018)

    Kopecz, S., Meister, A.: On order conditions for modified patankar–runge–kutta schemes. Applied Numerical Mathematics 123, 159–179 (2018)

    2018

  20. [20]

    arXiv preprint arXiv:2402.13788 (2024)

    Izgin, T.: A unifying theory for runge-kutta-like time integrators: convergence and stability. arXiv preprint arXiv:2402.13788 (2024)

    arXiv 2024

  21. [21]

    https://arxiv.org/abs/2312.01796 (2025) arXiv:2312.01796 [math.NA]

    Izgin, T., Ranocha, H.: Using Machine Learning to Design Time Step Size Controllers for Stable Time Integrators. https://arxiv.org/abs/2312.01796 (2025) arXiv:2312.01796 [math.NA]

    arXiv 2025

  22. [22]

    https://arxiv.org/abs/2602.14392

    Izgin, T., Meister, A., Shu, C.-W., Torlo, D.: Flux-Balanced Patankar-type Schemes for the Compressible Euler Equations (2026). https://arxiv.org/abs/2602.14392

    arXiv 2026

  23. [23]

    Öffner, P., Torlo, D.: Arbitrary high-order, conservative and positivity preserving Patankar-type deferred correction schemes. Appl. Numer. Math. 153, 15–34 (2020)

    2020

  24. [24]

    Communications in Applied Mathematics and Computational Science 20-1, 29–66 (2025) https://doi.org/10.2140/camcos.2025.20.29

    Izgin, T., Ketcheson, D.I., Meister, A.: Order conditions for Runge–Kutta-like methods with solution-dependent coefficients. Communications in Applied Mathematics and Computational Science 20-1, 29–66 (2025) https://doi.org/10.2140/camcos.2025.20.29

    work page doi:10.2140/camcos.2025.20.29 2025

  25. [25]

    Izgin, T., Kopecz, S., Meister, A.: On the stability of unconditionally positive and linear invari- ants preserving time integration schemes. SIAM J. Numer. Anal. 60(6), 3029–3051 (2022) https://doi.org/10.1137/22M1480318

    work page doi:10.1137/22m1480318 2022

  26. [26]

    ESAIM Math

    Izgin, T., Öffner, P.: A study of the local dynamics of modified Patankar DeC and higher order modified Patankar–RK methods. ESAIM Math. Model. Numer. Anal. 57(4), 2319–2348 (2023) https://doi.org/10.1051/m2an/2023053

    work page doi:10.1051/m2an/2023053 2023

  27. [27]

    Torlo, D., Öffner, P., Ranocha, H.: Issues with positivity-preserving Patankar-type schemes. Appl. Numer. Math. 182, 117–147 (2022) https://doi.org/10.1016/j.apnum.2022.07.014

    work page doi:10.1016/j.apnum.2022.07.014 2022

  28. [28]

    Total confirmed deaths due to COVID-19

    World Health Organization (2026) – processed by Our World in Data: “Total confirmed deaths due to COVID-19” [dataset]. World Health Organization, “COVID-19 Dashboard WHO COVID- 19 Dashboard - Daily cases and deaths” [original data]. https://ourworldindata.org/explorers/ covid?tableSearch=ghana&country=~GHA (Accessed: May 2, 2026)

    2026

  29. [29]

    Accessed: May 2, 2026 (2026)

    Worldometer: COVID-19 Coronavirus Pandemic. Accessed: May 2, 2026 (2026). https://www. worldometers.info/coronavirus/country/ghana/

    2026

  30. [30]

    Communications on Applied Mathematics and Computation 8, 868–892 (2026) https://doi.org/10.1007/s42967-025-00480-8

    Izgin, T.: A boot-strapping technique to design unconditionally positive dense output formu- lae for modified patankar-runge-kutta methods. Communications on Applied Mathematics and Computation 8, 868–892 (2026) https://doi.org/10.1007/s42967-025-00480-8

    work page doi:10.1007/s42967-025-00480-8 2026

  31. [31]

    Bull, A.D.: Convergence rates of efficient global optimization algorithms. J. Mach. Learn. Res. 12, 2879–2904 (2011)

    2011

  32. [32]

    In: Proceedings of the Thirtieth Conference on Uncertainty in Artificial Intelligence

    Gelbart, M.A., Snoek, J., Adams, R.P.: Bayesian optimization with unknown constraints. In: Proceedings of the Thirtieth Conference on Uncertainty in Artificial Intelligence. UAI’14, pp. 250–259. AUAI Press, Arlington, Virginia, USA (2014)

    2014

  33. [33]

    In: Pereira, F., Burges, C.J., Bottou, L., Weinberger, K.Q

    Snoek, J., Larochelle, H., Adams, R.P.: Practical bayesian optimization of machine learning algorithms. In: Pereira, F., Burges, C.J., Bottou, L., Weinberger, K.Q. (eds.) 17 Advances in Neural Information Processing Systems, vol. 25. Curran Associates, Inc. (2012). https://proceedings.neurips.cc/paper_files/paper/2012/file/05311655a15b75fab86956663e1819cd...

    arXiv 2012

  34. [34]

    Journal of Computational Physics 115(1), 200–212 (1994) https://doi.org/10.1006/jcph.1994.1187

    Liu, X.-D., Osher, S., Chan, T.: Weighted essentially non-oscillatory schemes. Journal of Computational Physics 115(1), 200–212 (1994) https://doi.org/10.1006/jcph.1994.1187

    work page doi:10.1006/jcph.1994.1187 1994

  35. [35]

    Journal of Computational Physics , volume =

    Jiang, G.-S., Shu, C.-W.: Efficient implementation of weighted eno schemes. Journal of Computational Physics 126(1), 202–228 (1996) https://doi.org/10.1006/jcph.1996.0130

    work page doi:10.1006/jcph.1996.0130 1996

  36. [36]

    Journal of Computational Physics 229(9), 3091–3120 (2010) https://doi.org/10

    Zhang, X., Shu, C.-W.: On maximum-principle-satisfying high order schemes for scalar conser- vation laws. Journal of Computational Physics 229(9), 3091–3120 (2010) https://doi.org/10. 1016/j.jcp.2009.12.030 18

    2010