Visual Image Analysis for Predicting Hospital Food Leftovers Percentage Using Random Forest Regressor

Monitoring hospital food consumption is an important component of nutritional care and food service management. Standard methods such as direct observation and weighed records remain the reference standard but they take a lot of work, subjective, and difficult to apply consistently in large scale monitoring. The proposed approach used a Random Forest Regressor trained on visual features that combine color, texture, and area information extracted from paired before and after meal images. The experiment was conducted using the LeFood-Set dataset, which contains 524 pairs of hospital meal images along with corresponding weight annotations. Two experimental scenarios are implemented: raw images and segmented images, where segmentation isolates the food region to enhance the relevance of extracted features. Experimental results show that the model trained on segmented images achieved better performance, with MAE = 11.35, MSE = 239.87, and R² = 0.834, outperforming the raw-image model (MAE = 17.43, MSE = 439.13, R² = 0.697). Feature importance analysis further shows that area and texture based features contribute most to prediction accuracy. This shows that segmentation not only improves precision but also increases model interpretability. These results confirm that integrating computer vision and machine learning provides a reliable framework for objective nutrition monitoring and future real-time food waste assessment in hospital settings.