(H-302) Rapid Carbapenemase Detection and Classification Using Machine Learning for Enhanced Management of Carbapenemase-Producing Organisms in Healthcare Settings

Introduction:: Infections due to carbapenemase-producing organisms (CPOs) are associated with high mortality rates. CPOs typically encode carbapenemase genes on plasmids, which both confer resistance to carbapenems (reserved for multi-drug resistance organisms) and can be spread to other bacteria (Fig. 1A). There are also different types or classes of carbapenemases, each with its own unique antibiotic resistance profile. Identification of CPO and the class, therefore, has implications for the selection of appropriate antibiotic treatment as well as infection control practices in healthcare facilities. Existing methods for CPO detection include phenotypic, molecular, and lateral flow assays, but these are not performed by most clinical microbiology laboratories due to a combination of test complexity, prolonged workflows, subjective results, reagent instability, and high cost (1). A simple and rapid screening test is needed for timely CPO identification and improved management of these organisms in various healthcare settings. We have developed a rapid CPO testing platform that provides colorimetric carbapenemase detection and differentiation of enzyme class in 5-15 minutes. However, colorimetric tests can be problematic as they rely on operators correctly interpreting results that can have some ambiguity. To that end, we report a novel machine learning approach that objectively and accurately identifies CPOs and the enzyme class (Fig. 1B).

Materials and Methods::

Our assay is an improved version of Carba NP, a phenotypic detection method for carbapenems (Fig. 1C). The original Carba NP suffers from prolonged incubations, high reaction volumes, and interpretation errors, all which attribute to its high cost and run-to-run variability. Changes to the original Carba NP chemistry have shortened our assay performance time to 30 minutes from 2.5 hours, eliminated the need for 37°C incubations, and further reduced costs with no loss in sensitivity. Additionally, subtle color changes in our assay are standardized by measuring absorbance change leading to minimal variability between results. The assay chemistry includes four reactions: 1) a no-imipenem negative control, 2) imipenem plus Zn2+ to detect any carbapenemase, 3) imipenem plus tazobactam to selectively inhibit class A enzymes, and 4) imipenem plus EDTA to inhibit class B enzymes. The reactions produce disparate results for the diverse classes of carbapenems, which remain challenging to categorize visually. We evaluated the performance of our platform using a set of highly characterized multi-drug resistant strains of gram-negative bacteria and developed machine learning algorithms to automatically call assay results down to the carbapenemase class level. Recursive cross-validation on well-known multi-class classification algorithms was employed to evaluate the effectiveness of classifying our dataset (n=117) using machine learning, composed of absorbance readings at 0, 5, 15, and 30 minutes. Specifically, stratified k-fold cross-validation was used to determine the accuracy of five machine learning methods including logistic regression, decision tree, k-nearest neighbors, support vector machine, and Gaussian Naïve Bayes (Fig. 1D).

Results, Conclusions, and Discussions:: Results: Our results show excellent robustness across various models, with a peak accuracy of 0.965 (Fig. 1D). Additionally, iterative model testing showed excellent robustness across the different models (Fig. 1E). We also re-ran the analysis on smaller data sets with the latter time points eliminated to determine the accuracy of the classification at shorter time points. We found that machine learning allowed the assay time to be shortened to 15 minutes, with minimal loss in accuracy (K-nearest Neighbors, accuracy 0.931). The optimized assay with machine learning classification provides complimentary solutions to the problems found in current phenotypic methods and are applicable in different use cases, such as high-volume microbiology laboratories. In future work, we expect higher accuracy and further reduced assay times as we move forward with concentrating carbapenemases for detection in a rapid microfluidic for on-demand testing of surveillance samples.

Discussion: Machine learning applied to phenotypic assays can uncover hidden features, enabling early and confident calls. Future investigations will explore this potential further, likely reducing assay time. Additionally, the assay accurately identified classes, but also unexpectedly had the potential to identify different enzyme families within each class due to more subtle color variations. We found that different enzyme families within each class generate different signal patterns (Fig. 1C) which can be characterized using similar machine learning methods. The potential of this finding will be investigated in future studies. Use of carbapenemase family signatures would allow epidemiologic characterization of enzymes in a population and the rapid identification of a new signature, indicating potential introduction or emergence of a novel carbapenemase.

Conclusion: Understanding if a patient has a CPO infection and, if so, the class of carbapenemase produced, has important treatment implications. Additionally, current testing platforms have limitations such as prolonged/cumbersome workflows, reagent instability, and subjective results. Our modified assay and classification algorithm combine the advantages of available methods into simple, rapid, and sensitive screening tests for the varied use-cases where CPO detection is needed. Our rapid, low-cost test will increase CPO detection, decrease time-to-identification of a CPO infection, and facilitate timely initiation of effective antibiotics and isolation practices.

Acknowledgements (Optional): :

References (Optional): : References: 1) Tamma PD, Simner PJ. 2018. Phenotypic Detection of Carbapenemase-Producing Organisms from Clinical Isolates. J Clin Microbiol 56.