Hybrid AI models sharpen multi-disease outbreak forecasting in Zambia

Researchers at the University of Zambia have built a forecasting system that models six infectious diseases at once, combining a mechanistic SEIR epidemic model with a transformer-based neural network, and report that the hybrid approach predicts weekly case counts more accurately than single-disease models for most of the pathogens they tested. The study, published in BMC Artificial Intelligence on 2 July 2026, was motivated by a persistent gap in Zambian epidemiological modelling: most existing models track one disease in isolation, even though COVID-19, cholera, influenza, measles, anthrax and mumps circulated simultaneously in the country between 2020 and 2023.

The team, led by Grey Chibawe with Mayumbo Nyirenda and Jackson Phiri, argued that single-disease frameworks miss disease interactions, co-infections and immune suppression effects, which can lead to inaccurate forecasts and inefficient allocation of scarce public health resources. Their central question was straightforward: does explicitly modelling several co-circulating pathogens together, rather than one at a time, actually improve prediction accuracy? To answer it, they assembled weekly surveillance data for all six diseases from 2020 to 2023, along with environmental variables that plausibly shape transmission.

The data came from several sources. COVID-19 records were drawn from the websites of the Centre for Infectious Disease Research in Zambia, the US Centers for Disease Control and the World Health Organization, and the researchers described these daily readings as consistent and complete. Data for cholera, influenza, measles, anthrax and mumps came from the Zambia National Public Health Institute, and these datasets contained frequent gaps. Daily records were aggregated to weekly counts, and residual missingness was handled with a continuity-preserving imputation strategy: gaps of two weeks or less were interpolated, while longer gaps were filled using seasonal averages or multiple imputation rather than being treated as zero incidence. Environmental data on temperature, rainfall and humidity, obtained in daily format from the VisualCrossing repository, were converted to weekly maxima to align with the epidemiological series.

On the mechanistic side, the researchers extended the classical SEIR framework, which partitions a population into susceptible, exposed, infectious and recovered compartments, by adding disease-specific exposed and infectious compartments plus an explicit co-infection compartment for individuals infected with two pathogens. For a two-disease system, the equations track primary infection with each pathogen and secondary infection of an already exposed or infectious individual by the other disease. The authors verified that the extended system conserves total population size, confirming that the model does not artificially create or lose individuals as it runs. Interaction between diseases was captured through co-infection coefficients, estimated from the surveillance time series using regression models with lagged terms.

Because sparse and noisy weekly data can produce biologically implausible interaction estimates, the team implemented a literature-informed shrinkage strategy. Penalised regression with weakly informative priors centred on zero was used, with prior means shifted toward antagonism for respiratory virus pairs where viral interference is documented, and toward synergy for pairs such as measles and influenza where post-infection vulnerability is described in the literature. For pathogen pairs with no mechanistic support, such as COVID-19 with mumps or anthrax, the prior was tightened around zero so that estimates would only deviate meaningfully if strongly supported by the data. All coefficients were reported with 95 percent confidence intervals derived from bootstrap resampling, and rolling twelve-month estimation windows showed the signs remained directionally stable, although confidence intervals widened during periods of sparse reporting.

The machine learning component was a transformer neural network layered on top of the SEIR outputs. Its inputs included SEIR-based weekly predictions, lagged case counts, the environmental variables and seasonal Fourier components encoding the epidemiological week. The architecture used multi-head self-attention with four heads and a key dimension of 64, layer normalisation, dense layers with ReLU activation and a regression output. The network was trained on data from 2020 to 2022 and tested on 2023 data, with early stopping to prevent overfitting and Bayesian optimisation to tune hyperparameters. The authors chose transformers over recurrent architectures for three reasons: parallel processing speeds training on multi-disease time series, self-attention captures long-range cross-disease dependencies, and non-sequential memory avoids vanishing gradient limitations.

The results supported the multi-disease hypothesis, with important caveats. In the dual-disease experiment pairing COVID-19 with influenza, extending the baseline SEIR model to two diseases reduced the root mean squared error from 593.138 to 557.065, although mean absolute error and R-squared did not improve. The hybrid model outperformed both SEIR-only and transformer-only configurations, cutting RMSE from 409.267 for the single-disease hybrid to 387.845 for the dual-disease hybrid. When the framework was expanded to all six diseases, hybrid models outperformed single-disease models for COVID-19, influenza, cholera, mumps and measles, with the COVID-19 hybrid model improving RMSE from 0.541 to 0.210. Anthrax was the exception: its dataset was particularly sparse, and the hybrid configuration performed worse for that disease. Validation in the reverse configuration, predicting influenza instead of COVID-19, yielded an even lower RMSE of 28.952, which the authors took as evidence of flexibility and generalisability.

The transformer-only models told a more cautionary story. Some purely data-driven models produced negative R-squared values, which the researchers attributed to a mismatch between model flexibility and the limited size of the weekly dataset, underscoring the risk of overfitting when deep learning is applied to short, irregular surveillance series. In the dual-disease setting, the multi-disease transformer actually did slightly worse than its single-disease counterpart, with RMSE of 387.198 versus 376.891, suggesting the network alone struggled to extract a limited co-infection signal. Only when all six diseases were integrated did the multi-disease transformer consistently beat its single-disease version across all outcomes, indicating that the architecture benefited substantially from the full cross-disease feature set. The hybrid design, by embedding epidemiological structure within the learning pipeline, appeared to stabilise training in exactly the situations where the standalone network faltered.

An interpretability analysis using SHapley Additive exPlanations offered a glimpse into what the hybrid model relied on. For COVID-19 prediction, the most influential feature was mumps incidence lagged by two weeks, with a SHAP value of 109.16, followed by COVID-19 itself lagged by one week at 105.82. Influenza lagged cases, weekly humidity and cholera lagged cases also ranked highly, while the SEIR-predicted feature contributed only marginally at 0.43. The authors stressed that SHAP values are model-specific feature attributions, not causal biological effects, but they noted that the prominence of mumps lags is intriguing: prior literature links mumps vaccination and antibody levels to COVID-19 severity, and mumps incidence fell during pandemic lockdowns in Pakistan and India, consistent with shared non-pharmaceutical interventions. The finding, they wrote, merits deeper investigation rather than being taken as established biology.

The authors were candid about the limitations. The four-year study window, weekly aggregation and sparse data for some diseases constrained the effective sample size for deep learning, increased overfitting risk and limited the generalisability of the estimated co-infection parameters beyond the Zambian surveillance context. The study period also coincided with COVID-19 waves, lockdowns, school closures and vaccination campaigns that altered transmission and reporting across diseases; these interventions were not explicitly parameterised but were indirectly embedded in the time-series structure. Sensitivity analyses varying the co-infection coefficients by plus or minus 20 percent and comparing imputation methods showed forecasting accuracy varied by less than 5 percent, suggesting the results are robust to moderate parameter and data-handling uncertainty. The team recommends that Zambia and similar settings adopt multi-disease hybrid forecasting, with collaboration between computer scientists and epidemiologists, and future work integrating vaccination rates and policy timelines as explicit model inputs.

Subject of Research: Hybrid SEIR and transformer machine learning models for multi-disease infectious disease forecasting in Zambia

Article Title: Hybrid multi-disease SEIR and transformer models for infectious disease forecasting in Zambia

Article References: Chibawe, G., Nyirenda, M., & Phiri, J. (2026). Hybrid multi-disease SEIR and transformer models for infectious disease forecasting in Zambia. BMC Artificial Intelligence, 2(1), Article 14. https://doi.org/10.1186/s44398-026-00031-y

Image Credits: AI Generated

DOI: 10.1186/s44398-026-00031-y

Keywords: infectious disease forecasting, SEIR model, transformer neural networks, Zambia, co-infection, COVID-19, cholera, machine learning, epidemiology, SHAP interpretability, surveillance data, public health