A deep-learning-based early warning system for abnormal eye conditions in chickens

In current poultry production practice, farmers are required to frequently enter their poultry houses and visually inspect their chickens to assess flock health. This practice is not only time-consuming and labor-intensive but also increases the risk of introducing contaminants into poultry houses and increasing disease transmission. In Taiwan, a typical poultry house houses approximately 15,000-20,000 red-feathered native chickens. Routine inspections are performed three times daily, and each inspection requires approximately one hour to complete. In the present study, a deep-learning-based early warning system was developed for detecting abnormal eye conditions in chickens. A pan-tilt-zoom camera was used to capture images from different areas of a poultry house and automatically detect and quantify the proportion of abnormal eye conditions. After these images were annotated into two categories, namely normal and abnormal, data augmentation techniques were applied to expand the training dataset. A You Only Look Once v7 deep learning model was then used to train the system, achieving a precision of 0.944, a recall of 0.831, and an F1 score of 0.884. The trained model was subsequently deployed in an enclosed commercial poultry farm to monitor three production cycles of red-feathered native chickens. Time-lag analysis demonstrated that the strongest associations occurred at lag periods of 3-7 days, with the highest correlation (ρ = 0.7304) observed at a 5-day lag in the July-August cycle. These findings indicate that increases in abnormal eye conditions may precede mortality by several days, supporting their role as an early warning indicator of flock health deterioration. A fixed-grid threshold scan identified an optimal abnormal eye proportion of approximately 4.5%, which achieved high sensitivity with acceptable false positive rates and demonstrated cross-cycle consistency. Overall, the proposed early warning system can serve as a digital chicken health monitoring tool that enables timely alerts and interventions, thereby reducing labor burden and potential economic losses.