Jalloh Saidu Wurie, Malenje Boniface, Imboga Herbert, Hodges Mary H
Department of Mathematics (Data Science Option), Pan African University Institute for Basic Sciences Technology and Innovation, Kiambu, 00200, Juja, Kenya.
Institute for Development (IfD), New England Ville, Freetown, Sierra Leone.
Malar J. 2025 May 20;24(1):158. doi: 10.1186/s12936-025-05389-4.
Malaria continues to pose a public health challenge in Sierra Leone, where timely and accurate forecasting can guide more effective interventions. Although seasonal models such as Seasonal Autoregressive Integrated Moving Average (SARIMA) have traditionally been employed for disease forecasting, Artificial Neural Networks (ANNs) have gained attention for capturing complex temporal patterns that linear models may not fully capture.
This study compares the forecasting performance of SARIMA and ANN models in forecasting malaria cases using malaria case data from 2018 to 2023. A baseline SARIMA model was developed and improved with exogenous climatic variables (precipitation, maximum temperature, and mean relative humidity) to form a SARIMAX approach. In parallel, an ANN was trained solely on historical malaria cases, without adding climatic variables.
SARIMA offered reasonable predictive capabilities but was outperformed by the ANN, which captured complex temporal patterns more effectively, decreasing forecast errors and improving its coefficient of determination . The SARIMA model achieved an MAPE of 12.01%, which improved further to an MAPE of 11.45% with the inclusion of climatic variables. A strong positive correlation between precipitation (r = 0.68) and malaria cases was observed, while maximum temperature showed a moderate negative correlation (r = ), and mean relative humidity demonstrated a moderate positive correlation (r = 0.55). The ANN model outperformed both the baseline SARIMA and SARIMAX models with the lowest MAPE of 6.68%.
These findings underscore the ANN's ability to capture non-linear dynamics, even without explicit climate inputs. These results reinforce the value of machine learning modelling approaches in guiding malaria control strategies, particularly in high-burden settings like Sierra Leone.
疟疾在塞拉利昂仍然是一项公共卫生挑战，及时准确的预测可以指导更有效的干预措施。尽管季节性自回归积分滑动平均（SARIMA）等季节性模型传统上一直用于疾病预测，但人工神经网络（ANN）因能够捕捉线性模型可能无法完全捕捉的复杂时间模式而受到关注。
本研究使用2018年至2023年的疟疾病例数据，比较了SARIMA和ANN模型在预测疟疾病例方面的预测性能。开发了一个基线SARIMA模型，并利用外部气候变量（降水量、最高温度和平均相对湿度）对其进行改进，以形成一种SARIMAX方法。同时，仅根据历史疟疾病例对一个ANN进行训练，不添加气候变量。
SARIMA具有合理的预测能力，但ANN的表现优于它，ANN能更有效地捕捉复杂的时间模式，减少预测误差并提高其决定系数。SARIMA模型的平均绝对百分比误差（MAPE）为12.01%，纳入气候变量后进一步提高到11.45%。观察到降水量与疟疾病例之间存在强正相关（r = 0.68），而最高温度呈中度负相关（r = ），平均相对湿度呈中度正相关（r = 0.55）。ANN模型的表现优于基线SARIMA和SARIMAX模型，最低MAPE为6.68%。
这些发现强调了ANN即使在没有明确气候输入的情况下捕捉非线性动态的能力。这些结果强化了机器学习建模方法在指导疟疾控制策略方面的价值，特别是在像塞拉利昂这样的高负担地区。
