This study investigates machine learning for forecasting medicine consumption in hospitals to optimize resource allocation and logistics. We use two approaches: a unified approach that combines data from multiple hospitals and a separated approach that forecasts for individual hospitals. We explored both K-means clustering and manual pair clustering based on consumption trends. While K-means clustering did not yield improvements, manual clustering identified specific pairs of medicines with significantly enhanced forecast accuracy (e.g., Medicine 15 at Hospital 1: MAPE decreased from 19.70% to 3.30%). However, the unified approach did not consistently benefit all hospitals (e.g., Medicine 9). This underscores the need to balance accuracy gains in some hospitals against potential losses in others. Overall, manual clustering within the separated approach shows promise. Future work should explore advanced automated clustering techniques like Dynamic Time Warping (DTW) and leverage larger datasets for further validation.