|
Our research should be seen in the light of the worldwide increase of antimicrobial resistance (AMR), which is a serious threat to human health. To prevent the spread of AMR, fast reliable diagnostics tools that facilitate optimal antibiotic stewardship are of urgent need. Raman spectroscopy (RS) is a promising tool for rapid label- and culture-free identification and antimicrobial susceptibility testing (AST) in a single step. To take full advantage of RS for bacterial identification machine learning (ML) analysis is essential. Many limitations must be addressed before RS will be a practical platform for point-of-care diagnostics applications in clinics and hospitals. RS is sensitive to factors such as the growth stage, changes in measurement environment and inconsistency in sample preparation. We address the issues of sample preparation, changes in measurement environment and limited data availability. We reduce sample preparation to merely transferring the bacteria to the measurement environment, hereby minimizing the issue of sample inconsistency and the additional benefit of removing sample preparation. To alleviate the situation of limited data availability for ML model training, we have developed a novel spectral transformer (ST) ML model that is efficient after training on both small- and large RS bacteria datasets. We explicit demonstrated that our ST outperforms a state-of-the-art domain-specific residual CNN both in terms of accuracy with 7.5%. Where we attain more than 96% classification accuracy on a dataset consisting of 15 different classes and 95.6% classification accuracy for six MR–MS bacteria species.
|
