An Acoustic Feature-Based Ensemble Learning Approach for Chicken Health Detection

Early disease detection in commercial poultry farms is critical for preventing outbreaks and minimizing economic losses. Conventional inspection is labor-intensive and frequently results in delayed diagnosis. This paper proposes an ensemble machine learning system for automated binary broiler health classification and evaluates the feasibility of non-invasive vocalization-based monitoring using acoustic analysis. Audio recordings were collected from 17–30-day-old broiler chickens in two closed-house commercial facilities: an academic research farm at the Faculty of Animal Science, Universitas Gadjah Mada (Yogyakarta), and a commercial farm in Blitar, Indonesia. Individual birds were recorded in isolated pens to eliminate background noise and ensure signal quality. A total of 22 acoustic features were extracted, comprising Mel-Frequency Cepstral Coefficients (13 features), time-domain features (5 features), and frequency-domain features (4 features). Three machine learning algorithms (SVM, Random Forest, and Logistic Regression) were evaluated across seven feature combinations using 5-fold cross-validation. Random Forest with MFCC features achieved the best individual performance (96.49% F1-score). An ensemble classifier with weighted soft voting was developed, integrating SVM (Time+MFCC), Logistic Regression (Time+Frequency+MFCC), and Random Forest (MFCC), with optimal weights determined through grid search, achieving 98.29% F1-score and 98.25% accuracy, outperforming the individual models. The high classification F1-score and accuracy demonstrate the feasibility of acoustic-based health monitoring for broiler chickens under controlled recording conditions to support early disease detection.