As companion animals are increasingly regarded as family members, interest in pet healthcare and disease prevention has grown significantly. Among various health issues affecting companion animals, obesity is recognized as one of the most prevalent an...
As companion animals are increasingly regarded as family members, interest in pet healthcare and disease prevention has grown significantly. Among various health issues affecting companion animals, obesity is recognized as one of the most prevalent and critical conditions, contributing to diabetes, cardiovascular diseases, arthritis, skin disorders, and reduced life expectancy. Despite its importance, canine obesity is still primarily assessed through Body Condition Score (BCS) evaluations or simple weight measurements, both of which have limitations in terms of objectivity, consistency, and continuous monitoring. Therefore, there is a growing need for data-driven and AI-based approaches that can support preventive healthcare and provide more accurate obesity assessment. This study proposes DOG-OB (Dog Obesity Classification Framework), an AI-based canine obesity classification framework that utilizes multivariate data collected from real-life environments. Unlike conventional approaches that rely primarily on body weight or subjective visual assessment, DOG-OB integrates multiple factors,
including weight, activity level, feeding behavior, and individual characteristics, to classify obesity status more comprehensively. The proposed framework encompasses the entire process of data collection, preprocessing, feature construction, machine learning-based classification, and performance evaluation.
The experimental dataset was collected from a dog daycare center between February and July 2025. A total of 1,024 records were obtained from eight companion dogs. After preprocessing, including missing value removal and IQR-based outlier detection, 944 recordswere retained for analysis. Input variables consisted of age, sex, weight, indoor activity time, outdoor activity time, feeding frequency, and food intake. The target variable, FAT_GRADE, was defined as a four-class obesity grade based on veterinarian-assessed Body Condition Scores (BCS), including obese, overweight, normal, and underweight categories. To evaluate the proposed framework, Random Forest was selected as the final classification engine of DOG-OB. Model performance was assessed using a 7:3 holdout validation strategy and stratified 5-fold cross-validation. Multiple evaluation metrics, including Accuracy, Precision, Recall, F1-score, and ROC-AUC, were employed to validate classification performance. Experimental results demonstrated that DOG-OB achieved an Accuracy of 0.9398, a Macro F1-score of 0.9329, a Weighted F1-score of 0.9395, and a ROC-AUC of 0.9949. These results indicate excellent classification capability across the four obesity categories. Feature
importance analysis revealed that weight and age were the most influential variables for obesity classification, while feeding and activity-related variables also contributed meaningfully to prediction performance. Furthermore, ablation studies confirmed that the integration of multiple lifestyle variables significantly improved classification accuracy compared to single-variable approaches. The proposed DOG-OB framework demonstrates the feasibility of utilizing real-world multivariate data for canine obesity classification and provides a practical foundation for preventive pet healthcare services. By combining explainable machine learning techniques with lifestyle-based health indicators, this study contributes to the development of intelligent pet healthcare systems and offers a scalable framework for future obesity monitoring and health management applications. Future work will focus on expanding the dataset, incorporating additional physiological sensors, and developing time-series prediction models for long-term health monitoring.