Antimicrobial resistance (AMR) poses a significant threat to global health, limiting treatment options for infections. While understanding of AMR mechanisms has improved, rapid identification of resistant organisms remains challenging. Conventional phenotypic antimicrobial susceptibility testing (AST) is time-consuming, often resulting in empirical treatment with potentially inappropriate antimicrobials. This delay impacts patient outcomes. High-content imaging (HCI) offers a potential solution by providing detailed morphological data on bacterial populations. HCI integrates automated microscopy and analysis to measure numerous cellular characteristics. While HCI has been used to study morphological changes under drug exposure, its potential for predicting AMR without drug exposure is largely unexplored. *Salmonella Typhimurium*, a significant enteric pathogen, is increasingly resistant to ciprofloxacin, a commonly used treatment. This study investigates the use of HCI combined with machine learning to predict ciprofloxacin susceptibility in *S. Typhimurium* isolates, aiming to provide a rapid and accurate AMR prediction method without the need for traditional AST.

Existing literature extensively covers AMR mechanisms, including chromosomal mutations, plasmid-borne genes, and inducible resistance. However, the lack of rapid methods for distinguishing susceptible and resistant organisms remains a significant challenge. Conventional AST methods, such as the Vitek2 and BD Phoenix, while automated, still rely on bacterial growth in the presence of antimicrobials, making them time-consuming. High-content imaging (HCI) has emerged as a promising technique for studying bacterial morphology and response to antimicrobials. Previous studies have used HCI to analyze morphological changes under drug exposure, identifying potential links between morphology and the mechanism of action (MoA) of antimicrobials. However, the direct relationship between morphological characteristics and AMR remains underdefined, particularly the capability of predicting susceptibility without drug exposure. This gap highlights the need for this study.

This study utilized four *S. Typhimurium* isolates (two clinical, two laboratory strains) with varying ciprofloxacin susceptibilities. Isolates were exposed to four ciprofloxacin concentrations (0x, 1x, 2x, and 4x MIC) over 24 hours, with samples taken every two hours for high-content imaging (HCI). The HCI platform captured 65 morphological, intensity, and texture features per bacterial cell. Time-kill curves were generated to assess growth dynamics under ciprofloxacin exposure. Statistical analyses, including principal coordinate analysis (PCoA), were employed to visualize data and identify clustering patterns. Machine learning algorithms, specifically random forests and neural networks, were trained on the HCI data to classify isolates based on ciprofloxacin susceptibility, both with and without prior drug exposure. The performance of these classifiers was evaluated using various metrics, including accuracy, sensitivity, specificity, precision, F1 score, and AUC. In silico AMR analysis was also performed using ARIBA and ResFinder to identify resistance genes present in isolates.

Time-kill curves revealed distinct growth trajectories for the four *S. Typhimurium* isolates under ciprofloxacin exposure. HCI data showed temporal changes in bacterial morphology, with significant elongation followed by a decrease in length under ciprofloxacin treatment. The resistant strains exhibited less pronounced changes in cell length. Hierarchical clustering of HCI data revealed two main clusters – one primarily containing 0x and 1x MIC data points and another containing 2x and 4x MIC data points, indicating concentration and time of exposure as primary factors affecting cellular response. A random forest classifier trained on the combined data from all strains exhibited an OOB error rate of 0.25, identifying key features associated with treatment conditions. Cell length was identified as a crucial feature, exhibiting a time-dependent trend and correlation with ciprofloxacin concentration. Comparison of SL1344 and its GyrA mutant (SL1344gyrA) revealed distinct morphological differences, even at high ciprofloxacin concentrations. This distinction was further pronounced at 4x MIC after 20 hours. A random forest classifier perfectly distinguished the two isolates, highlighting the impact of GyrA mutations on morphology. Remarkably, even without ciprofloxacin exposure, ciprofloxacin-resistant and -susceptible isolates exhibited distinct morphological features at 22 hours. PCoA of the 0x MIC-22h data clearly segregated resistant and susceptible isolates. A random forest model identified ten key features associated with resistance, including fluorescence intensity and texture measurements. Testing a broader dataset (16 strains) using various machine learning classifiers (Naive Bayes, SVM, Random Forest, CatBoost, and Neural Network) indicated that only five features were necessary to accurately distinguish susceptible and resistant isolates. A neural network achieved the best performance (accuracy of 0.87 ± 0.08 on the test set), highlighting the potential of using these five key features for rapid AMR prediction without drug exposure. Analysis of the neural network revealed that SYTOX Green intensity (related to membrane permeability) and DAPI intensity (related to DNA content) were crucial for predicting resistance.

This study demonstrates a novel approach to rapid AMR prediction using HCI and machine learning. The ability to predict ciprofloxacin susceptibility without drug exposure represents a significant advancement. Unlike conventional AST methods, HCI offers high-throughput, high-resolution analysis of bacterial populations at the single-cell level. This method provides valuable insights into the morphological impact of antimicrobials and could potentially identify new drug targets. The findings suggest that ciprofloxacin resistance is associated with inherent differences in membrane permeability, even in the absence of drug exposure. While the study focuses on *S. Typhimurium* and ciprofloxacin, the generalizability of this approach to other Gram-negative bacteria and antimicrobials warrants further investigation. The use of interpretable machine learning models facilitates understanding of the relationship between morphological features and AMR probability.

This research successfully demonstrates the potential of combining HCI with machine learning for rapid and accurate prediction of ciprofloxacin resistance in *S. Typhimurium*. The identification of key morphological features that distinguish resistant and susceptible isolates without drug exposure offers a novel approach to AST. This technique may be adaptable to other bacterial species and antimicrobials, ultimately contributing to improved diagnostics and antimicrobial stewardship.

The study's limitations include the use of pure bacterial cultures in liquid media, which may not perfectly reflect clinical settings. The use of a proprietary analysis software also limited parameter flexibility. Further refinement of the image segmentation and analysis algorithms, potentially using deep learning architectures, could improve accuracy and single-cell resolution. The study primarily focused on *S. Typhimurium* and ciprofloxacin, and further research is needed to assess the generalizability across various bacterial species and antimicrobial agents. Finally, the current system's complexity and cost could hinder immediate clinical adoption, but this could be addressed through development of more accessible diagnostic devices.