Al-based mobile application to fight antibiotic resistance

The emergence of antimicrobial resistance (AMR) outpaces the development of new antimicrobial agents, posing a significant global health threat. The lack of data, particularly in areas with limited access to antimicrobial susceptibility testing (AST), hinders a full understanding of AMR's magnitude. While the Kirby-Bauer disk diffusion test is widely used for AST, it suffers from inter-operator variability and complex interpretation. Automatic reading systems mitigate these issues but are often unaffordable and inaccessible in resource-limited settings. This research addresses this gap by developing an AI-powered, offline smartphone application for antibiogram analysis, specifically designed for use in resource-constrained environments where access to sophisticated AST is limited or non-existent. The application's goal is to increase the availability and reliability of AST results, leading to improved patient care and a better understanding of AMR epidemiology.

Various methods exist for testing bacterial susceptibility to antibiotics, with the Kirby-Bauer disk diffusion test being the most prevalent. This method, while simple and inexpensive, is criticized for its labor-intensity, susceptibility to inter-operator variability in manual measurement of inhibition zones, and the need for expert interpretation. Automatic reading systems, while improving accuracy and consistency, are often expensive and require specialized hardware, rendering them unsuitable for resource-limited settings. Existing image processing algorithms for automatic measurement of inhibition zone diameters have been published, but many lack the user-friendliness and offline capability crucial for widespread adoption in resource-constrained environments.

This study introduces a fully offline mobile application for analyzing disk diffusion ASTs and providing interpreted results, entirely on a smartphone. The application combines original algorithms utilizing machine learning (ML) and image processing with a rule-based expert system for automatic AST analysis. The image processing module involves plate cropping, antibiotic disk detection (using a trained convolutional neural network), and inhibition zone diameter measurement using a novel algorithm called SWITCH (Spatial Weighted Intensity Threshold CHangepoint). This algorithm segments radial profiles to determine inhibition zone boundaries, accounting for irregular shapes and variations in contrast. An expert system, based on up-to-date EUCAST rules, performs coherence checks and provides interpreted results. The application also incorporates a user-friendly graphical interface for user interaction and correction of automatic measurements. The application's performance was evaluated on three sets of antibiograms prepared under laboratory conditions using both standard and blood-enriched agar. Comparisons were made against a hospital-standard automatic system (SIRscan) and manual measurement (gold standard). Additionally, the feasibility of an ML-based automatic detection of resistance mechanisms was explored.

The application's fully automatic measurement procedure achieved an overall agreement of 90% with the hospital-standard automatic system (SIRscan) and 98% with manual measurement (gold standard) in susceptibility categorization. This demonstrates the feasibility of accurate antibiotic resistance testing on a smartphone. Inter-operator variability was significantly reduced compared to manual measurement. The application's image processing library, developed in C++ and using OpenCV and TensorFlow, showed robustness to variations in image quality inherent in using smartphone cameras. The ML model for identifying antibiotic disks achieved 99.97% accuracy. The analysis of a single antibiogram takes less than 1 second on a PC, 1.5 seconds on a high-end smartphone, and 6.6 seconds on a low-end smartphone. The application's performance was consistent across different types of growth media (standard and blood-enriched). While ML models for detecting specific resistance mechanisms showed promising accuracy, the risk of overfitting was deemed too high for inclusion in the application's current version. The application successfully addresses the challenges of manual AST interpretation, and performs comparably to existing commercial systems.

This study successfully demonstrates the feasibility and accuracy of an AI-powered, offline smartphone application for performing and interpreting disk diffusion AST. The application's performance is comparable to existing, more expensive, automated systems and significantly improves upon the accuracy and consistency of manual methods. Its offline capability and reliance on readily available smartphone technology make it ideal for resource-limited settings, potentially addressing a major global health challenge. The user-friendly interface ensures ease of use, even for personnel with limited microbiology expertise. The development and deployment of this application represent a significant step towards improving access to accurate and timely AST worldwide, facilitating better patient care and enhanced epidemiological surveillance of AMR.

This research has successfully developed and validated an offline smartphone application for accurate and efficient analysis of antibiotic susceptibility tests. The application's high performance, ease of use, and suitability for resource-limited settings position it as a valuable tool in the global fight against antibiotic resistance. Future work will focus on further clinical validation in MSF hospitals and the open-source release of the application (AntigbioGo). Integration with global AMR surveillance systems will allow for the collection of valuable epidemiological data.

The study's primary limitation is the relatively small size of the training datasets used for the ML models. This raises the potential for overfitting and limits the generalizability of these models to diverse bacterial species and antibiotic combinations. While the acquisition guidelines help to standardize image quality, the inherent variability in smartphone cameras remains a potential source of error. The application's performance in real-world clinical settings, outside of the controlled laboratory environment, requires further evaluation. The current version does not include automatic detection of resistance mechanisms due to the risk of misinterpretations and potential clinical consequences.