Dataset for "Machine Learning Models for Hierarchical Forecasting of Diabetes Cases in Ghana"

Background: In any nation, including Ghana, public health planning and budget allocation depend heavily on accurate diabetes case projections at all administrative levels. In hierarchical health data, non-linear trends and cross-level dependencies are frequently missed by traditional time series approaches. Methods: This work suggests a machine learning method for XGBoost-based hierarchical diabetes case prediction in Ghana. A bottom-up method was used to reconcile base-level projections for 16 regions Level 2 into three geographical belts and a national total. ETS and ARIMA models were used as benchmarks for XGBoost. The XGBoost model added exogenous temporal elements, such as lag variables, quarter, and month of the year, to enhance performance. The first 60 months of the dataset were used for training, and the latter 12 months were set aside for out-of-sample validation. At every level of the hierarchy, forecast accuracy was assessed. Results: When it came to reconciled national-level projections, XGBoost outperformed ARIMA and ETS. Additionally, XGBoost demonstrated significant prediction performance, based on the root mean square error (RMSE), mean absolute error (MAE), and mean absolute square error (MASE), for the three geographical belts at the unreconciled level. All levels of prediction accuracy were enhanced by the addition of time-based elements. Coherence between region, belt, and national projections was guaranteed by the bottom-up reconciliation method. Conclusion: The results show that XGBoost is a useful technique for hierarchical diabetic case forecasting when paired with bottom-up reconciliation and designed time characteristics. At the regional and national levels in Ghana, this method offers more precise and logical forecasts to assist focused health initiatives and resource planning.