Application of Machine Learning in the Detection of Antimicrobial Resistance
Antimicrobial resistance (AMR) has become one of the significant global threats to both human and animal health, intensifying the need for rapid and precise AMR diagnostic methods. Traditional antimicrobial susceptibility testing (AST) is time-consuming, low throughput, and limited to cultivable bac...
Saved in:
Main Author: | |
---|---|
Contributors: | |
Format: | Doctoral Thesis |
Language: | English |
Published: |
Philipps-Universität Marburg
2023
|
Subjects: | |
Online Access: | PDF Full Text |
Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Summary: | Antimicrobial resistance (AMR) has become one of the significant global threats to both human and animal health, intensifying the need for rapid and precise AMR diagnostic methods. Traditional antimicrobial susceptibility testing (AST) is time-consuming, low throughput, and limited to cultivable bacteria. Machine learning offers a promising avenue for automated AMR prediction. However, most existing models emphasize features related only to known resistance genes and variants, relying heavily on AMR reference databases, and thus may overlook new AMR-related features. To address the above challenges, my first study introduces genome-wide machine learning models to detect AMR without dependence on prior AMR knowledge efficiently. Specifically, I assessed various models, including logistic regression (LR), support vector machine (SVM), random forest (RF), and convolutional neural network (CNN), for predicting resistance against four antibiotics. The findings illustrated that these models can effectively predict AMR with label encoding, one-hot encoding, and frequency matrix chaos game representation (FCGR) encoding on whole-genome sequencing data. Generally, RF and CNN outperformed LR and SVM. Importantly, I identified specific mutations associated with AMR for each antibiotic.
Moreover, current AMR studies focus on single-drug resistance prediction, ignoring the cumulative nature of antimicrobial resistance over time, which makes rapid identification of multi-drug resistance (MDR) a challenge. Therefore, in my second study, in order to overcome these limitations, I constructed five multi-label classification (MLC) models for MDR problems. The findings revealed that the ECC (Ensemble Classifier Chains) model surpassed the other MLC methods, demonstrating marked effectiveness in predicting MDR.
Furthermore, the constraints of limited training samples and data imbalances present significant barriers to the generalization and accuracy of AMR models. To overcome these challenges, in my third study, I have proposed a deep transfer learning model based on a CNN architecture. First, I pre-train the model on four datasets, then the best-performing model is used as the source model for transfer learning, and the model is retrained on small datasets by transferring the architecture and weights from the source model. The results showed that the deep transfer learning model improves model performance for AMR prediction on small and imbalanced datasets.
In an era where data security and privacy are crucial, federated learning (FL) and swarm learning (SL) present solutions by maintaining data locally during training, which reduces the necessity to transfer sensitive information to a centralized server and improves efficiency by distributing computational load. Moreover, swarm learning achieves decentralization by not requiring a central server to manage the parameters compared to federated learning, which further improves the security of the data. Thus, in my fourth study, I delve into the application of swarm learning specifically within the context of AMR. |
---|---|
DOI: | 10.17192/z2024.0050 |