Machine Learning Could Transform Disease Surveillance in Nigeria, Review Finds

Machine learning could reshape how Nigeria detects and responds to disease outbreaks, but only if the country first confronts a set of stubborn structural obstacles, according to a narrative review published in BMC Artificial Intelligence. The paper, led by Lukman Jibril Aliyu and colleagues, examines how artificial intelligence techniques could be woven into Nigeria’s Integrated Disease Surveillance and Response (IDSR) framework, which has coordinated surveillance since 1998 but remains hampered by inconsistent data quality, limited private sector participation, and inadequate digital infrastructure. The authors argue that machine learning’s capacity for predictive analytics, real-time data processing, and automated pattern recognition offers transformative potential, yet adoption within Nigeria’s surveillance system lags considerably behind global advances.

Nigeria’s surveillance system operates through a multi-tiered structure in which data flows from primary health facilities to Local Government Areas, then to State Ministries of Health and the Nigeria Centre for Disease Control (NCDC). Disease Surveillance and Notification Officers play a central role, often working under resource constraints that include insufficient funding, shortages of trained personnel, and logistical challenges in remote areas. Historically, manual reporting led to delays and inconsistencies, but electronic systems have shown promise: the electronic IDSR pilot improved reporting timeliness from 43 percent to 73 percent, with a near doubling of reporting facilities. Even so, incomplete documentation, lack of interoperability between health data systems, and disparities in adoption between urban and rural areas continue to create gaps in national-level disease monitoring.

The review catalogs concrete examples of machine learning already tested in Nigerian settings. A 2022 pilot in Lagos used machine learning to predict malaria outbreaks with a 78 percent accuracy rate, but it was limited to urban centers with reliable internet and covered only 12 percent of the state’s health facilities. The Auto-Visual Acute Flaccid Paralysis Detection and Reporting system, which leverages mobile technology and artificial intelligence for polio surveillance, increased case reporting by 45 percent across 10 states from 2017 to 2020, yet its reliance on mobile networks excluded 30 percent of rural areas. These cases illustrate scalability constraints tied directly to infrastructure and funding, the authors note, even as they demonstrate the feasibility of AI-powered forecasting and automated diagnostics in the Nigerian context.

To illustrate a practical modeling workflow, the team constructed a synthetic weekly dataset for malaria surveillance covering 10 representative Local Government Areas in Northern Nigeria over two years, 2023 and 2024. Each record contained malaria case counts, total rainfall in millimeters, average temperature, and a health access index scaled from 0 to 1. Weekly malaria cases were modeled as a function of rainfall, temperature, and health access, plus a random noise term, with higher rainfall assumed to increase cases following a short lag and better health access assumed to reduce incidence. The authors are explicit that this exercise is a methodological illustration rather than empirical evidence: because the outcome variable was generated using predefined relationships with the predictors, the reported metrics should be interpreted strictly as internal simulation diagnostics, not as indicators of real-world predictive performance.

Under a temporally separated validation design, with 2023 data used for training and 2024 data reserved for testing, a decision tree regression model with a maximum depth of six achieved an R² of approximately 0.66 and a root mean square error of approximately 3.61 within the synthetic dataset. The team then benchmarked five machine learning models—Linear Regression, Decision Tree, Random Forest, Gradient Boosting, and Support Vector Machine—under identical temporal validation conditions. While ensemble methods and support vector machines marginally outperformed the decision tree on test R², the differences were modest, supporting the choice of the decision tree as the primary modeling approach for its interpretability advantage at minimal accuracy penalty. A sensitivity analysis that perturbed the rainfall coefficient across a plausible range of plus or minus 33 percent of its base value found the model maintained stable performance, with R² values ranging from 0.922 to 0.980 and root mean square error values remaining within a narrow band of 2.38 to 2.47.

Interpretability received particular attention. The researchers applied SHapley Additive exPlanations (SHAP), a technique that assigns each feature a contribution value for each individual prediction, revealing how much each input pushes the model’s output up or down relative to a baseline. In the simulation, rainfall and health access emerged as the most influential features, while temperature had a comparatively smaller effect—consistent with the notion that temperature in the region is generally always conducive to malaria transmission, with less week-to-week variability. In weeks with high rainfall, SHAP values for the rainfall feature were strongly positive, and Local Government Areas with higher health access showed negative contributions for that feature, meaning the model predicted fewer cases than it would have if health access were low. The authors emphasize that such transparent explanations are crucial for deploying artificial intelligence in healthcare, enabling public health officials to verify that a model’s drivers make sense before acting on its predictions.

Beyond the simulation, the review surveys machine learning methodologies suited to surveillance work. Supervised learning, trained on labeled datasets such as previous outbreaks, is useful for case classification, syndromic pattern recognition, and predictive modeling for diseases like cholera and Lassa fever. Unsupervised learning, applied where outcomes are unknown, allows clustering of cases or anomaly detection, particularly valuable for novel syndromes or underreported diseases. The paper details relevant architectures: decision trees and random forests for structured epidemiological data, support vector machines for high-dimensional clinical and demographic data, recurrent neural networks for temporal health data such as weekly IDSR reports, and convolutional neural networks adapted for spatial epidemiology by treating spatial disease data as images. The authors also discuss rigorous evaluation strategies, including k-fold and time-series cross-validation, noting that time-series cross-validation is especially important in epidemiology where data has temporal dependencies.

The review situates Nigeria’s efforts against international benchmarks. Digital and AI-assisted contact tracing in South Korea and Taiwan significantly reduced delays in case identification and quarantine, with Taiwan achieving a median response time of roughly three days from exposure to isolation. In India, AI-driven syndromic surveillance flagged COVID-19 hotspots earlier than conventional reporting systems. Rwanda’s AI-enhanced surveillance infrastructure enabled real-time tracking of COVID-19 case clusters, facilitating swift quarantine measures. Globally, systems such as HealthMap, which scrapes news media and official health reports using natural language processing, and WHO’s Epidemic Intelligence from Open Sources, which fuses multiple open-source data streams, illustrate what scalable real-time analytics can achieve. By comparison, Nigeria’s IDSR framework relies heavily on manual reporting, limited private sector integration, and lacks scalable real-time analytics and diverse data sourcing. Benchmarking against Rwanda and Kenya, the authors identify key gaps including weak AI-health policy integration, fragmented data systems, limited dedicated funding, and fewer scalable AI hubs, while cautioning that the comparative analysis is based on published literature and should be interpreted with caution given differences in reporting systems and data availability across countries.

The barriers to implementation are substantial. Many Nigerian healthcare facilities lack electronic health records and rely on paper-based systems, making it difficult to train machine learning models effectively. Only 15 percent of facilities use interoperable electronic records, and the fragmentation delays data aggregation by up to three weeks, undermining real-time potential; the authors suggest adopting FHIR standards could reduce this lag to 48 hours, as seen in Kenya’s e-health rollout. Nigeria also faces a shortage of professionals with the technical expertise needed to develop, deploy, and maintain machine learning-based health systems, alongside high costs, weak enforcement of data protection regulations—the National Data Protection Regulation mandates consent and anonymization but only 10 percent of health facilities comply—and resistance among healthcare professionals wary of new technologies. Most machine learning models are trained on datasets from high-income countries, which may not capture Nigeria’s epidemiological patterns, leading to biased predictions without locally relevant datasets.

To move forward, the authors propose a tiered strategic framework. Tier 1 foundational reforms, spanning years one and two, include equipping Disease Surveillance and Notification Officers in priority states with low-cost digital reporting tools, mandating phased HL7/FHIR adoption starting with tertiary and secondary facilities, enforcing the National Data Protection Regulation with independent compliance audits, establishing a National AI-for-Health Working Group, and integrating AI literacy into medical and public health curricula. Tier 2, in years three through five, calls for expanding rural connectivity through public-private partnerships, migrating surveillance data to secure cloud platforms, piloting retrospective validation studies for malaria, Lassa fever, and cholera using NCDC historical data, and tracking measurable indicators such as reducing the reporting lag from three weeks to 72 hours. Tier 3, in years five through ten, envisions deploying cloud-based machine learning surveillance at national scale across all 36 states and the Federal Capital Territory, developing federated learning frameworks that enable model training across 774 Local Government Areas without centralizing sensitive patient records, integrating genomic surveillance pipelines, and establishing cross-border collaboration through ECOWAS and Africa CDC. The authors stress that all findings remain illustrative and require real-world validation using NCDC or state-level surveillance data before deployment, and that Nigeria-specific cost-effectiveness studies using real surveillance data are strongly recommended before large-scale funding decisions are made.

Subject of Research: Applications of machine learning to disease surveillance in Nigeria, including a synthetic malaria modeling illustration and strategic recommendations

Article Title: Enhancing disease surveillance in Nigeria through machine learning: opportunities, challenges and strategic recommendations

Article References: Aliyu, L. J., Umar, A. B., Sani, S. K., Haruna, U. A., Danjuma, A. H., Zubairu, M., & Yahaya, H. (2026). Enhancing disease surveillance in Nigeria through machine learning: opportunities, challenges and strategic recommendations. BMC Artificial Intelligence, 2(1), Article 17. https://doi.org/10.1186/s44398-026-00036-7

Image Credits: AI Generated

DOI: 10.1186/s44398-026-00036-7

Keywords: machine learning, disease surveillance, Nigeria, malaria, IDSR, NCDC, SHAP, public health, outbreak detection, artificial intelligence, health data, LMIC