A Checkpoint-Based Hybrid Transformer-CNN Framework for Early Prediction of Escherichia coli Infection Risk in Broiler Chickens

Technical Summary: Checkpoint-Based Hybrid Transformer-CNN Framework for Early E. coli Infection Risk Prediction

Problem Statement
Colibacillosis, caused by pathogenic Escherichia coli, remains a significant constraint in commercial broiler production, leading to reduced growth, increased mortality, and poor feed conversion. Current management practices often rely on manual observation and retrospective production reviews, which are reactive; interventions typically occur only after clinical signs or production losses are evident. Furthermore, existing monitoring studies often rely on single data sources (either image-based or sensor-based), failing to capture the complex interactions between environmental stressors, production metrics, and visual bird condition that precede severe disease. There is a need for a proactive, multimodal system capable of detecting gradual risk progression before severe symptoms appear.

Methodology
The study proposes a checkpoint-based Hybrid Transformer-CNN framework designed for early risk prediction. The methodology integrates four data modalities: environmental, production, behavioural, and visual data, structured around six specific production checkpoints (Day 3, 7, 14, 21, 28, and 31).

1. Data Collection and Preprocessing:

   • Temporal Data: Environmental variables (temperature, humidity, ammonia, CO2), production metrics (feed/water intake, body weight, mortality, FCR), and behavioural indicators (activity level, ventilation status) were collected from commercial broiler contexts (2022–2025). These were cleaned, normalized, and organized into sequential observations.
   • Visual Data: Poultry images were resized to 224x224 pixels, normalized, and augmented (rotation, flipping, zooming, brightness adjustment) to address class imbalance, particularly for low- and medium-risk categories.
   • Labeling: Flocks were classified into low-, medium-, and high-risk categories based on combined indicators.

2. Model Architecture:

   • Transformer Encoder: A Transformer model was developed to analyze the sequential temporal data. It utilized input embeddings, positional encoding, multi-head self-attention, and feed-forward layers to model the progression of risk across the six checkpoints. The model was trained using the Adam optimizer and cross-entropy loss on a dataset of approximately 90,000 observations (72,000 training, 18,000 test).
   • Convolutional Neural Network (CNN): A CNN was employed for visual health assessment, classifying images into healthy, low-risk, medium-risk, high-risk, and non-broiler classes. This component served as supportive evidence for visible health conditions rather than a direct pathogen detector.
   • Multimodal Fusion: Decision-level fusion combined probability outputs from both models into a weighted risk score.
   • Explainability and Interface: The system incorporated Grad-CAM to highlight image regions influencing predictions. A Streamlit dashboard was developed to present risk scores, confidence levels, contributing factors, alerts, and veterinary advisory prompts.

Key Contributions

• Hybrid Architecture: The integration of a Transformer for temporal sequence modeling and a CNN for visual assessment creates a multimodal system that leverages complementary evidence sources.
• Checkpoint-Based Monitoring: The framework introduces a structured approach to monitoring flock health across six distinct production stages, allowing for the detection of deterioration trends over time rather than isolated snapshots.
• Explainable Decision Support: Unlike "black-box" models, this framework provides interpretability through Grad-CAM visualizations, confidence scores, and specific advisory prompts, translating model outputs into actionable management information.
• Proactive Surveillance: The system shifts poultry health monitoring from retrospective classification to proactive early-warning, aiming to identify risk before severe clinical signs manifest.

Results

• Transformer Performance: On 18,000 test observations, the Transformer model achieved an accuracy of 99.96%. It demonstrated high precision, recall, and F1-scores (all >99%) across low, medium, and high-risk categories, with perfect recall (100%) for the low-risk category.
• CNN Performance: On 249 evaluation images, the CNN achieved an overall accuracy of 95.58%. It performed strongly on healthy (100%), high-risk (96.85%), and non-broiler (100%) classes. However, performance was lower for low-risk (50%) and medium-risk (50%) classes, attributed to fewer training examples and subtle visual boundaries in intermediate risk states.
• System Integration: The dashboard successfully synthesized these outputs. In a documented example, the system identified a high-risk flock (score 0.85) and correctly flagged contributing factors such as high ammonia, humidity, and ventilation issues, generating appropriate alerts and veterinary advice.

Significance and Claims
The authors claim that this framework advances precision poultry management by unifying temporal flock records with visual assessment into an explainable early-warning system. The study posits that while environmental and production records reveal gradual deterioration, image data provides necessary visual confirmation of bird condition. By integrating these sources, the framework reduces dependence on any single data type.

The paper emphasizes that the system supports earlier intervention, better environmental oversight, and data-driven health decisions. However, the authors maintain a modest stance regarding deployment, noting that the high Transformer accuracy suggests strong separability within the specific dataset but requires external validation across different housing systems, climates, and management routines. Similarly, they acknowledge that the CNN's performance on intermediate risk classes is limited by data imbalance and the lack of E. coli-specific image datasets, suggesting that future work must address larger, more balanced datasets and live IoT integration before broad operational adoption.