Hybrid Probabilistic Forecasting of Under-Five Malaria Admissions in Ghana: A Gaussian Process Regression with Holt-Winters Smoothing Journal: Artificial Intelligence In Medicine
T. Ansah-Narh, Y. Asare Afrane, J. Bremang Tandoh · 2026
Read Hybrid Probabilistic Forecasting of Under-Five Malaria Admissions in Ghana: A Gaussian Process Regression with Holt-Winters Smoothing Journal: Artificial Intelligence In Medicine by T. Ansah-Narh, Y. Asare Afrane, J. Bremang Tandoh (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-00834.txt.
Introduction
Ansah-Narh and co-authors forecast malaria admissions for young children in Ghana in this 2026 paper. Forecasts there are hard: seasons swing hard, reports are unsure, and transmission shifts over time. District teams need methods that are both rigorous about doubt and sturdy on thin data.
Their hybrid pairs two tools. Gaussian Process Regression captures wavy patterns and gives doubt ranges. Holt-Winters smoothing steadies long-range forecasts and keeps the seasonal shape. They fit ten years of district data, from 2014 to 2023, and test with rolling-origin checks. The hybrid scores R-squared of 0.9906, against 0.8213 for Holt-Winters alone. Most residuals fall inside the two-sigma bounds, as they should. Forecasts for 2024 to 2028 see average monthly admissions rise from about 8,000 to 12,200. Northern high-burden districts keep stable shares despite large swings in counts.
Catalog record
| CREATOR | T. Ansah-Narh, Y. Asare Afrane, J. Bremang Tandoh |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 83 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2606.00834
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Hybrid Probabilistic Forecasting of Under-Five Malaria Admissions in Ghana: A Gaussian Process Regression with Holt-Winters Smoothing Journal: Artificial Intelligence In Medicine T. Ansah-Narh, Y. Asare Afrane, J. Bremang Tandoh Abstract Accurate malaria forecasting remains a major challenge in sub-Saharan Africa, where strong seasonality, reporting uncertainty, and non-stationary transmission dynamics reduce the reliability of conventional epidemiological time-series models. In Ghana, district-level malaria surveillance requires forecasting frameworks that are not only predictive but also probabilistically rigorous, operationally interpretable, and robust under limited-data conditions. This study proposes a hybrid probabilistic forecasting framework that integrates Gaussian Process Regression (GPR) with Holt–Winters exponential smoothing for modelling monthly malaria admissions among children under five years of age. Within this framework, GPR captures complex non-linear temporal behaviour and predictive uncertainty, while Holt–Winters smoothing stabilises long-horizon forecast trajectories and preserves recurrent seasonal structure. Using ten years of district-level malaria surveillance data (2014–2023), model performance was evaluated using a rolling-origin expanding-window validation framework to ensure assessment on sequential unseen future observations. Full-sample analyses were additionally performed to examine each model’s ability to reproduce the underlying temporal dynamics of malaria transmission. The proposed hybrid model achieved substantially improved predictive performance ( R 2 = 0.9906 R^{2}=0.9906 ) relative to the best standalone baseline model ( R 2 = 0.8213 R^{2}=0.8213 for Holt–Winters), while maintaining well-calibrated residual behaviour, with approximately 94.2 % 94.2\% of residuals contained within ± 2 σ \pm 2\sigma uncertainty bounds. The results further demonstrate that integrating probabilistic machine learning with seasonal smoothing reduces forecast volatility while preserving epidemiologically meaningful seasonal amplitudes, thereby improving forecast stability and interpretability over extended horizons. Forecasts for 2024–2028 project average monthly malaria admissions ranging from approximately 8,000 8{,}000 to 12,200 12{,}200 cases, consistent with observed seasonal transmission patterns and stable predictive intervals across the forecast horizon. Spatio-temporal variability analysis additionally revealed pronounced ecological heterogeneity in malaria burden across Ghana, with northern high-burden districts exhibiting comparatively stable relative transmission patterns despite large absolute fluctuations in admissions. These findings highlight the importance of uncertainty-aware forecasting systems capable of supporting adaptive malaria preparedness under heterogeneous transmission conditions. The proposed framework provides a scalable probabilistic forecasting approach for malaria early warning, seasonal preparedness, and short- to medium-term operational planning in endemic settings. Although constrained by reporting quality and the absence of environmental covariates, this study establishes a methodological foundation for integrating probabilistic machine learning and classical seasonal smoothing within epidemiological forecasting systems designed for resource-constrained public-health environments. † † graphicalabstract: † † highlights: Hybrid GPR–Holt–Winters model enhances malaria forecast accuracy. Probabilistic framework improves uncertainty quantification and stability. Spatial analysis reveals high, stable burdens in northern districts. Robust residual calibration with 94.2\% within \pm 2\sigma limits. 1 Introduction Malaria remains a major public health burden across sub-Saharan Africa, driven by persistent challenges such as increasing drug resistance, limited access to healthcare, and fragile health systems [ 14 , 29 , 35 ] . In 2021 alone, the region accounted for 247 million malaria cases globally, underscoring its disproportionate share of the disease burden [ 30 ] . Ghana is among the countries most affected, reporting approximately 5.4 million malaria cases in 2021 , equivalent to an incidence rate of 164.3 cases per 1,000 population and 12,557 associated deaths 1 1 1 https://www.afro.who.int/sites/default/files/2023-08/Ghana.pdf . Despite national and global initiatives aimed at malaria control and eventual elimination, the disease remains endemic throughout Ghana. Children under five years of age continue to be the most vulnerable group, reflecting broader regional patterns [ 4 , 10 ] . Malaria transmission in Ghana is not uniform; it exhibits strong seasonal and spatial variability, with environmental factors such as rainfall, temperature, and humidity influencing transmission peaks across different ecological zones [ 11 , 3 ] . To strengthen malaria control efforts, forecasting has become an essential tool for anticipating outbreaks, informing public health planning, and guiding the timely allocation of resources. Accurate and interpretable forecasts at monthly resolution can support early warning systems and localised interventions [ 27 ] . However, developing reliable forecasting tools remains challenging due to the non-linear and non-stationary dynamics of malaria transmission, particularly in regions like Ghana, which has high district-level variability. Traditional statistical models, such as Autoregressive Integrated Moving Average (ARIMA) and its seasonal extension, SARIMA, have long been used for malaria forecasting due to their simplicity and ease of implementation. However, these models rely on linearity and stationarity assumptions that are often violated in epidemiological time series. These models are particularly effective for stationary time series with linear dynamics and well-defined seasonal components [ 12 ] . Similarly, exponential smoothing methods like Holt-Winters are frequently used for short-term forecasting in health surveillance systems where seasonality is prominent [ 36 ] . However, these models exhibit several limitations when applied to epidemiological data, which are often characterised by non-linear trends, abrupt regime shifts, and unobserved exogenous influences such as environmental variability and intervention policies [ 37 , 20 ] . Moreover, classical time series models do not inherently provide a full probabilistic description of future observations, limiting their ability to express forecast uncertainty in a principled manner. This becomes a critical drawback in public health decision-making, where risk quantification and uncertainty bounds are essential for planning and response. Recent advances in machine learning and probabilistic modelling have brought powerful tools such as Gaussian Process Regression (GPR) into the spotlight. GPR is a non-parametric Bayesian method capable of learning from limited data, modelling complex seasonality, and providing full predictive distributions rather than point estimates [ 38 ] . Its ability to incorporate uncertainty directly into forecasts makes it particularly suitable for epidemiological forecasting, where confidence intervals are essential for risk assessment and policy-making [ 31 ] . Nonetheless, the application of GPR to malaria forecasting in African contexts remains underexplored, and its performance relative to classical statistical models is not yet fully understood. More critically, existing studies have largely treated probabilistic machine learning models and classical time-series approaches as competing alternatives rather than complementary components. This has left a specific methodological gap: there is limited work on hybrid frameworks that explicitly address the tendency of GPR to produce increasingly diffuse predictive intervals while preserving the seasonal structure captured by classical methods such as Holt-Winters. In epidemiological settings characterised by strong seasonality and operational decision constraints, this limitation is non-trivial. Purely probabilistic models may yield forecasts with high uncertainty that are difficult to operationalise, while classical approaches impose structure at the cost of flexibility. A principled integration of these paradigms therefore defines a distinct methodological niche, enabling forecasts that balance uncertainty quantification with seasonal stability. To address this gap, we propose a comparative study of forecasting models using district-aggregated monthly malaria incidence data from Ghana, covering the period from January 2014 to December 2023. The data include admissions for both children under five and individuals aged five and above, along with spatial coordinates (latitude and longitude) for each district. For this study, we focus primarily on the under-five population due to their heightened vulnerability and high incidence burden. These data are collated from national health surveillance systems and represent a rich temporal dataset amenable to rigorous modelling. Our study pursues three main objectives. First, we aim to construct forecasting models using four complementary approaches: Linear Regression, Holt-Winters Exponential