Sekar Parameswara Meilia Soegiharto, Yuita Arum Sari, Sigit Adinugroho
Accurate assessment of hospital food leftovers is critical in monitoring patient nutrition and minimizing inefficiency; however, the current method of visual estimation by observer is subjective, carried out without impartiality and time-consuming. Hence, this article has proposed a lightweight convolutional neural network (CNN) architecture, based on MobileNetV2, to classify hospital food leftovers images according to the Comstock scale. As a highly configurable model designed for low-resource devices such as mobile tablets, the proposed Lightweight CNN enables real-time and impartial assessment of food leftovers. The experimental results reveal that the proposed Siamese MobileNetV2 architecture achieves an overall accuracy of 73.58%, MAE equals 0.547, and Cohen’s Kappa equals to 0.641. Even with a small dataset and class imbalance, our model yields objective predictions at low computational cost, therefore proposing promising applications in hospital. This method shows potential for AI-based nutrition care and reducing food waste in hospitals – can be embedded into hospital information system to assist dietitians in interpreting patients’ consumption patterns and developing personalized menu plans. The results also confirm that Lightweight CNNs can process real-world clinical image data and facilitate the translation of scientific research into clinical practice. Our future study will concentrate on expanding the dataset, multimodal analysis and deploying real-time to mobile devices. © 2025 IEEE.
Faculty of Computer Science, Brawijaya University, Malang, Indonesia
