Development of a Dog Health Score Using an Artificial Intelligence Disease Prediction Algorithm Based on Multifaceted Data

**Background:** Dog owners often lack veterinary expertise, making it difficult to promptly assess their pets' health status. Dogs communicate primarily through behaviors, and abnormal patterns (e.g., scratching, licking, swallowing, sleeping changes) can indicate diseases such as skin conditions, digestive disorders, or chronic kidney disease. Previous studies have used wearable sensors to monitor human health, but canine health assessment faces challenges due to variability in breed, age, and weight. This study aimed to develop a quantitative Health Score using AI analysis of behavioral data from activity sensors, validated against veterinarian diagnoses.

**Methods:** The study was conducted over nine months (May 2022 to February 2023) with 30 dogs recruited from dog parks, dog cafes, and veterinarian recommendations. Inclusion criteria: dogs exhibiting normal daily life, no surgical history in prior three months, and under 10 years old (determined by dental and joint assessments). All dogs underwent a health checkup at a veterinary hospital before the experiment. Dogs were categorized by weight (small, medium, large) and age. Activity sensors (33 mm × 38 mm, 18.3 mm thick, 15 g) were attached to collars, featuring Bluetooth 5.0, IP67 waterproof certification, a 2000 degrees per second acceleration sensor, and gyro-sensor resolution of 50 Hz. Data were validated using concurrent smartphone camera recordings. Four behavioral types were monitored: scratching, licking, swallowing (food and water), and sleeping. Initially, three months of baseline daily behavior data were collected from each dog. Over 600 instances of these four behaviors were processed through a fuzzy associative memory algorithm for training. Scratching and licking were classified into four tiers: occasional, average, frequent, and severe. Swallowing was quantified by number per hour; sleep data were assessed based on average daily duration (12-16 hours typical for dogs). The Health Score was calculated on a scale from 1 to 10 using Equation (1): H = Σ(i=0 to n-1) (wa + wb + wc + wd), where i is activity sensor data, W is data weight for each behavior, and j represents Health Score calculation data. Scores 1-5 indicated health conditions requiring medical attention; scores 6-10 indicated healthy state. Validation was performed by comparing AI-generated scores with classifications by three veterinarians using percent agreement formula: percent agreement % = 100 × (a + d) / (a + b + c + d), where a = both positive, b/c = discordant, d = both negative.

**Key Results:** The 30 dogs had an average age range of 3-6 years (16 male, 14 female), with average weight ranging from 3 kg to 30 kg. Breeds included retrievers, Jindo, Shih Tzu, poodles, beagles, mixed breeds, pit bulls, Siberian huskies, and Maltese. Average scratching frequencies: retrievers 122 times, Jindo 130 times, Shih Tzu 144 times; these exceeded standard by 18, 22, and 16 times respectively. Poodles showed swallowing frequency of 74 times (higher than average). Beagles and mixed breeds had swallowing frequencies of 26 and 23 times. All breeds except retrievers, Shih Tzus, and Jindos averaged more than 12 hours of sleep. The AI algorithm assigned health scores of 3-4 for retrievers, pit bulls, and Shih Tzus, indicating need for veterinary examination, which aligned with veterinarians' assessments. AI scores also corresponded with veterinarians' diagnoses for Siberian huskies, poodles, Maltese, and other breeds. A discrepancy occurred for one beagle: AI suggested health score of 4 (recommending health test), while veterinarian assessed it as healthy. Overall agreement between AI-evaluated results and veterinarians' analyses was 87.5%.

**Clinical Implications:** The Health Score system provides a reliable, quantitative tool for dog owners without veterinary expertise to assess their pet's health status based on daily behavioral monitoring. The 87.5% concordance rate with veterinarian diagnoses supports its validity. The system uses small, lightweight sensors that do not interfere with dogs' daily activities. However, limitations include: predictions based solely on data analysis without comprehensive veterinary data (e.g., coat quality, age); only eight breeds studied out of over 300 recognized breeds; and small sample size (30 dogs) over nine months. The Health Score reflects current health status and does not diagnose or predict diseases; it is presented as a range rather than a precise quantitative value to prevent misinterpretation due to fluctuating behavior patterns. Future research should include more diverse breeds, larger cohorts, advanced sensors, and fuzzy hierarchical techniques to improve prediction accuracy. This approach has potential for enhancing pet healthcare through proactive, data-driven health monitoring.