TL;DR
This paper introduces a novel hybrid model, GM(1,1)-BP, that combines grey system theory and backpropagation neural networks to predict the trend of drug resistance in animal-derived pathogens. By preprocessing limited historical data with interpolation and cluster analysis, the GM(1,1)-BP model achieves significantly improved prediction accuracy compared to individual grey or neural network models, offering a valuable tool for monitoring and combating antimicrobial resistance.
What Problem Does This Paper Address?
Antimicrobial resistance (AMR) in animal-derived pathogens is a growing global threat to both animal and human health. Traditional methods for testing drug resistance are often slow, expensive, and inefficient, making it difficult to track and predict resistance trends. This lack of effective prediction hinders proactive measures to combat AMR. The paper specifically addresses the challenge of predicting AMR trends using limited, fragmented historical data, a common issue in real-world surveillance.
Methods
The study employs a multi-stage approach:
- Data Preprocessing:
- Multiple Regression Interpolation: Missing drug resistance data points are imputed using a multiple linear regression model based on existing time-series data. This helps to create a more complete dataset for analysis.
- K-Means Clustering Analysis: Drugs are classified into groups based on their resistance patterns. This helps to identify relationships between different drugs and their resistance trends, providing valuable input features for the prediction models.
- Prediction Models:
- Grey Model (GM(1,1) and GM(1,N)): These models are used for time-series forecasting, particularly effective with limited data and irregular distributions. GM(1,1) is a univariate model, while GM(1,N) can handle multivariate relationships.
- BP Neural Network: A standard multi-layer feedforward neural network known for its nonlinear mapping capabilities and ability to learn from data.
- GM(1,1)-BP Hybrid Model: This is the core innovation. It combines the strengths of the grey model (handling limited data and trends) with the BP neural network (nonlinear mapping and adaptive learning). The grey model's output is used as input for the BP neural network, with additional layers for data "ashing" and "albino" treatment to enhance robustness and information reduction.
Datasets
The study utilizes a dataset of Escherichia coli (E. coli) resistance data collected from the Chengdu Monitoring Station of Sichuan Province, China, spanning from 2003 to 2014. This dataset includes the number of E. coli strains resistant to seven specific drugs: ofloxacin, enrofloxacin, doxycycline, tetracycline, gentamicin, ceftiofur, and sulfafurazole. The paper notes that data for 2005 and 2006 were missing, and some data for other years were also incomplete, necessitating the interpolation step.
Key Insights and Ideas
- The Power of Hybrid Models: Combining different modeling approaches (grey theory and neural networks) can overcome the limitations of individual models, leading to superior performance. The GM(1,1)-BP model leverages the strengths of both for better AMR prediction.
- Data Preprocessing is Crucial: Interpolating missing data and clustering drugs based on resistance patterns significantly improves the quality of input data for prediction models, leading to more accurate results.
- AMR Trends are Complex: Simple grey models, while useful for initial trends, struggle with the complex, fluctuating nature of AMR over longer periods. Neural networks are better equipped to capture these nonlinear dynamics.
- Targeted Prediction: By identifying similar drug resistance patterns through clustering, researchers can use data from one drug to help predict resistance trends for another, especially when data is scarce.
Key Results
The study evaluated the performance of GM(1,1), GM(1,N), BP neural network, and the proposed GM(1,1)-BP model for predicting E. coli resistance to sulfamethoxazole.
- GM(1,1) and GM(1,N): These models showed lower prediction accuracy, with R² values below 0.5 and RMSE values higher than 0.1. They were better at capturing initial trends (GM(1,1)) or later trends (GM(1,N)) but struggled with overall prediction.
- BP Neural Network: While capturing the general trend, the BP model had a low variation range and an R² of only 0.5147, making accurate prediction difficult.
- GM(1,1)-BP Hybrid Model: This model demonstrated the best performance, achieving an R² of 0.7830 and an RMSE of 0.0527. The relative error for the year 2014 was 13.9%, significantly lower than the other models.
The paper also developed a web-based drug resistance prediction system that visualizes these prediction results, aiding in the analysis of AMR trends.
Limitations and Future Work
- Data Limitations: The study acknowledges that the prediction models are based on a "small sample" of drug resistance data, which inherently limits the scope and generalizability. While interpolation and clustering help, more comprehensive and diverse datasets would further enhance model robustness.
- Generalizability: The models were developed and tested using E. coli resistance data. While the authors suggest the approach can be applied to other animal-derived pathogens, further validation with different bacterial species and geographical regions is needed.
- Model Complexity: The GM(1,1)-BP model, while effective, involves multiple steps and parameters. Future work could focus on simplifying the model or exploring more automated feature selection and hyperparameter tuning.
- System Integration: The developed prediction system is a valuable tool, but its integration into existing surveillance systems and its adoption by relevant stakeholders would be crucial for its real-world impact.
Summary updated