Machine Learning Maps Arsenic Danger in Ghana’s Mining-Hit Rivers

Communities along three of Ghana’s most important rivers are drinking water laced with arsenic far more often than most people realize, and a new study now shows exactly where the danger concentrates—and which artificial intelligence tools can predict it. Research published in BMC Environmental Science examined the Pra, Ankobra, and Tano river basins of southwestern Ghana, where roughly two-thirds of all water samples exceeded the World Health Organization’s drinking-water guideline of 10 micrograms of arsenic per liter. The work, led by Senyefia Bosson-Amedenu of Takoradi Technical University, compared three machine learning approaches and found that a random forest model delivered the most reliable and interpretable forecasts of both arsenic concentrations and the probability that a given sample would breach the safety threshold.

The stakes could hardly be higher. The Pra, Ankobra, and Tano basins drain the Birimian greenstone belts, geological formations rich not only in gold but also in arsenic-bearing sulfides such as arsenopyrite. Over the past decade, the explosive growth of artisanal and small-scale gold mining—known locally as galamsey—has torn apart channel and floodplain sediments, accelerating the oxidation of these sulfides and flushing arsenic into rivers and adjacent alluvium. The people living along these corridors depend on the same waters for drinking, smallholder irrigation, and fisheries, which turns a hydrological problem into a direct public health and livelihood crisis. Routine monitoring campaigns have long reported frequent guideline exceedances, but the spatial pattern of that risk has been difficult to pin down with confidence.

The research team established 88 fixed monitoring sites across the three basins—42 on the Pra, 24 on the Tano, and 22 on the Ankobra—positioned to capture upstream-to-downstream gradients, confluences, municipal water abstraction points, and zones of mining disturbance. Each site was visited monthly for roughly 20 months, producing 1,752 observations that covered at least one full hydrological cycle of wet and dry seasons. Field teams collected mid-channel grab samples in acid-washed bottles, filtered and acid-preserved them on ice, and shipped them to the laboratory within 24 hours, where arsenic was quantified by inductively coupled plasma mass spectrometry or graphite furnace atomic absorption spectroscopy. In situ probes recorded temperature, pH, and specific conductance, while GPS units logged precise coordinates to within a few meters.

Before any modeling began, the researchers audited their data with unusual rigor. The dataset showed zero percent missingness, so no imputation was needed. Screening for extreme values flagged only longitude, where stations strung along narrow river corridors produced leverage-prone outliers; the team winsorized that single predictor at the 2.5th and 97.5th percentiles while deliberately leaving the arsenic measurements untouched, so that the public health burden reflected in exceedance statistics would not be artificially softened. The raw numbers told a sobering story: mean arsenic concentration was 11.95 micrograms per liter, the median 11.12, and the overall exceedance rate stood at 63 percent. By basin, Ankobra fared worst with 76.5 percent of samples over the guideline, followed by Pra at 61.3 percent and Tano at 51.4 percent.

The modeling comparison pitted a random forest against a support vector machine with a radial basis kernel and a feed-forward neural network. To eliminate so-called lucky-split bias, the team used five-times-repeated five-fold nested cross-validation, in which outer folds estimated generalization performance while inner folds tuned hyperparameters. They also tested four different train-validation-test partitions with 30 repeated stratified splits each. The random forest emerged as the most accurate or statistically tied-for-best model across nearly every configuration. At the 70-20-10 partition, it achieved a median mean absolute error of 3.86 micrograms per liter and a root mean square error of 4.93, with an area under the receiver operating curve of 0.73 for classifying exceedances. The support vector machine closed the gap at larger training sizes, but the neural network lagged consistently, its discrimination actually deteriorating as training data grew—likely a symptom of limited sample size and a sparse predictor set rather than a fundamental flaw of deep learning itself.

Crucially, the winning model was not a black box. Using SHAP analysis, a model-agnostic technique that quantifies each feature’s contribution to individual predictions, the researchers found that water temperature dominated the model’s logic by a wide margin, followed by pH and latitude. This hierarchy is chemically plausible: temperature-linked hydrochemistry is known to modulate arsenic mobility, and the strong latitudinal signal reflects broad geological and land-use gradients across the basins. Site-level location effects captured residual heterogeneity that coordinates alone could not explain, while calendar month contributed only modestly. Seasonal-trend decomposition of the 2020 to 2026 monthly series confirmed that picture, revealing weak seasonality with a peak-to-trough amplitude of roughly half a microgram per liter and only a slight downward drift of about 0.1 micrograms over the period. In other words, arsenic risk in these rivers is governed far more by spatial position and water chemistry than by the calendar.

Perhaps the study’s most methodologically important contribution lies in how it corrected for a subtle but pervasive bias in environmental monitoring. The Pra basin hosts nearly half of all stations, so naive hotspot mapping would inevitably paint broad swaths of the Pra corridor as dangerous simply because more measurements were taken there. The team countered this with three complementary techniques: an effort-adjusted exposure index that multiplies each site’s arsenic level by the inverse of local station density, a density-weighted and row-standardized Getis-Ord Gi* statistic with a 25-kilometer neighborhood that downweights oversampled areas, and a relative-risk surface that divides the kernel density of exceedances by the kernel density of sampling effort. After these corrections, the sprawling hot corridors largely vanished, leaving only sparse, localized hotspots—precise targets for field audits and enforcement rather than blanket crackdowns.

The corrected maps reshuffled the policy priorities in an instructive way. Once monitoring effort was accounted for, the Pra basin carried the largest aggregate exposure burden, with a median effort-adjusted exposure index of 321.45, reflecting the broad spatial extent of contamination across its heavily settled corridor. The Ankobra, by contrast, retained the highest probability that any single measurement would exceed the guideline, at 76.5 percent. The Tano ranked lower on both metrics but was emphatically not risk-free. The authors therefore recommend a dual-track strategy: invest in basin-scale mitigation and source control—riparian buffers, tailings containment, treatment upgrades—in the Pra, while running high-frequency compliance surveillance and rapid-response protocols in the Ankobra, where unsafe readings are most likely at any given visit. The Tano warrants sustained but proportionate monitoring as resources allow.

The study is candid about its limits. The predictor set was intentionally parsimonious—temperature, pH, discharge where available, coordinates, and time—because hydrochemical variables such as iron, dissolved organic carbon, alkalinity, and redox potential were not consistently measured across sites. The SHAP diagnostics suggest those missing chemistry variables, especially redox proxies, would add real explanatory power, and the authors call for their routine collection. Roughly 20 months of data is also too short to resolve interannual or climate-driven trends with confidence, and population-weighted exposure estimates built on 100-meter raster products can misrepresent the linear villages that hug riverbanks. The team recommends spatially balanced sampling designs, sentinel stations in sparse reaches, and spatial block cross-validation as the path forward.

What makes this work resonate beyond Ghana is its template. By pairing rigorous, repeated nested validation with transparent model explanation and bias-aware spatial statistics, the researchers have produced something rare in environmental machine learning: predictions that are accurate, defensible, and directly actionable. The framework aligns with several United Nations Sustainable Development Goals, from clean water and sanitation to health and climate action, and it transfers readily to other contaminants and other data-limited regions where monitoring stations cluster unevenly. For the millions of people who draw their daily water from the Pra, Ankobra, and Tano, the message is concrete: the danger is real, it is mappable, and the tools now exist to target remediation where it will save the most health—provided the monitoring network itself becomes as balanced as the science it supports.

Subject of Research: Machine learning prediction of arsenic exposure risk in Ghana's riverine ecosystems

Article Title: Comparative performance of machine learning models in predicting arsenic exposure risk in Ghana’s riverine ecosystems

Article References: Bosson-Amedenu, S., Boakye, E., Ayitey, E., & Addor, J. A. (2025). Comparative performance of machine learning models in predicting arsenic exposure risk in Ghana’s riverine ecosystems. BMC Environmental Science, 2(1), Article 26. https://doi.org/10.1186/s44329-025-00040-8

Image Credits: AI Generated

DOI: 10.1186/s44329-025-00040-8

Keywords: arsenic contamination, machine learning, random forest, Ghana, galamsey, river basins, water quality, SHAP interpretability, hotspot analysis, WHO guideline, Pra River, Ankobra River