Predicting Fasciolosis in Uruguayan 🇺🇾 Bovine Carcasses: An End-to-End Machine Learning Study

Description

This repository implements a supervised machine learning pipeline to predict fasciolosis (liver fluke infection) in bovine carcasses processed at a major Uruguayan abattoir in 2016. Using routinely collected post-slaughter data, the model leverages Light Gradient Boosting Machine (LightGBM) with Bayesian hyperparameter optimization, feature engineering, and rigorous evaluation. The approach prioritizes sensitivity (recall) for disease detection, using the three most informative features: carcass weight, dental age, and fat score. All pipeline components (data preprocessing, imputation, feature ranking, model tuning, and performance assessment) are designed for full reproducibility and operational relevance in abattoir surveillance workflows.

Overview

• Target: Binary indicator of fasciolosis presence/absence in bovine carcasses.
• Key predictive features:
  • Carcass weight (continuous)
  • Dental age (age_teeth; ordinal)
  • Fat score (ordinal)
• Additional optional variables explored but not used in final model:
  • Productive purpose (categorical)
  • Carcass conformation score (numeric)
  • Month (derived from slaughter date)
• Model: LightGBM classifier with Bayesian hyperparameter tuning (maximizing AUC-ROC).
• Training & Evaluation:
  • Stratified 5-fold cross-validation on training data for hyperparameter tuning.
  • Held-out test set evaluation on multiple metrics including AUC-ROC, recall, precision, F1-score, accuracy, and confusion matrix.
• Operating point: Sensitivity-oriented (high recall) threshold to reduce missed infected carcasses at the cost of increased false positives.

Graphical summary

Figure: Receiver Operating Characteristic (ROC) curve of the tuned LightGBM classifier predicting fasciolosis presence in Uruguayan bovine carcasses (test set). The model achieves an AUC-ROC of 0.6519, indicating moderate discriminative ability.

Data

• Source: Publicly available abattoir dataset [1].

• Target variable: Fasciolosis presence (fasciola binary indicator).

• Features:

  Feature	Type	Description
  carcass_weight	Continuous	Post-slaughter carcass weight
  age_teeth	Ordinal	Dental age classification
  fat_score	Ordinal	Fat cover score
  carcass_conformation_score	Numeric	Conformation score (not in final model)
  productive_purpose	Categorical	Production purpose (one-hot encoded, optional)
  month	Ordinal	Derived from slaughter date (1-12)

• Preprocessing steps:

  • Standardized variable names.
  • Missing values imputed with median (numerical) or mode (categorical).
  • One-hot encoding for categorical variables with unknown category handling.

Methods

1. Data preparation and preprocessing

   • Clean and standardize column names.
   • Derive month from date field.
   • Impute missing values.
   • Encode categorical variables (one-hot for productive_purpose when included).

2. Feature ranking

   • Random Forest classifier-based feature importance ranking on training data.
   • Group one-hot encoded variables back to original feature level.
   • Select top three features (carcass_weight, age_teeth, fat_score) to ensure parsimony and interpretability.

3. Modeling and hyperparameter tuning

   • Use LightGBM binary classifier with balanced class weights.
   • Bayesian optimization over:
     • num_leaves, learning_rate, n_estimators, min_child_samples, subsample, colsample_bytree, reg_lambda, reg_alpha
   • Optimization target: maximize mean AUC-ROC via 5-fold stratified cross-validation on training set.
   • Early stopping employed during cross-validation.
   • Final model trained on full training data restricted to top 3 features with best hyperparameters.

4. Evaluation

   • Held-out test set evaluation: AUC-ROC, accuracy, precision, recall, F1-score.
   • Confusion matrix analysis and class-wise metrics for interpretability.
   • Emphasis on recall (sensitivity) to detect infected carcasses.

Results

• Feature importance (Random Forest baseline):
  • carcass_weight: 0.876
  • age_teeth: 0.082
  • fat_score: 0.017
• Hyperparameter tuning:
  • Best AUC-ROC (cross-validation): ~0.6510
  • Optimal hyperparameters:
    • num_leaves=16
    • learning_rate≈0.2534
    • n_estimators=767
    • min_child_samples=63
    • subsample≈0.6889
    • colsample_bytree≈0.7264
    • reg_lambda≈1.4116
    • reg_alpha≈2.0948
    • class_weight=balanced
• Test set performance:
  • AUC-ROC: 0.6519
  • Accuracy: 0.5822
  • Precision (infected class): 0.4451
  • Recall (infected class): 0.7263
  • F1-score: 0.5519
  • Confusion matrix:
    • True Negative (TN): 1959
    • False Positive (FP): 1935
    • False Negative (FN): 585
    • True Positive (TP): 1552
• Interpretation: Model favors recall (sensitivity), suitable for minimizing missed infections in surveillance contexts.

Discussion and Future Work

• Dominant signal from carcass weight aligns with known biological impacts of fasciolosis on animal physiology.
• Age and fat score provide supplemental but weaker signals.
• The model is best suited for abattoir-stage detection, where carcass physiology is apparent.
• Pre-slaughter or herd-level prediction will require upstream environmental and management data to avoid label leakage.
• Potential improvements:
  • Expanding feature set (multi-year temporal data, farm/environment covariates).
  • Alternative interpretability approaches (SHAP values, permutation importance).
  • Investigate other learners (XGBoost, CatBoost) or ensemble methods.
  • Implement probability calibration and threshold tuning for operational cost-benefit optimization.
  • Incorporate precision–recall and cost-sensitive metrics aligned with real-world trade-offs.

---

License

This project is released under the MIT License.

---

Questions?

For questions, suggestions, or collaboration, please open an issue in this repository or contact me at jprmaulion[at]gmail[dot]com.