Predicting Human Salmonella Antimicrobial Resistance from Poultry Antibiotic Residues: A Simulation-Based Machine Learning Proof-of-Concept with One Health Integration

In the invisible world of microscopic life, bacteria are constantly adapting. When they are exposed to medicines designed to kill them, the strongest survivors learn to resist, passing those defenses to their offspring. This phenomenon, known as antimicrobial resistance, is a growing crisis that threatens to make common infections difficult or impossible to treat. A major concern for scientists is how this resistance spreads from animals to people. In many places, farmers use antibiotics to keep poultry healthy and help them grow. However, traces of these drugs can remain in the meat we eat. The prevailing theory is that these tiny, leftover amounts of medicine act as a constant, low-level pressure on bacteria like Salmonella, forcing them to evolve resistance before they ever reach a human host. If this is true, then the amount of drug residue found in a chicken could serve as an early warning sign for the resistance levels we might see in sick patients.

The challenge is that proving this link is incredibly difficult. To do so, researchers would need to track a specific piece of chicken from a farm, measure the exact amount of antibiotic residue inside it, and then find the exact same strain of bacteria in a human patient to see if that person's infection is resistant. Such perfectly matched data rarely exists in the real world. Because of this gap, a researcher at the Syrian Virtual University, Yara Al-Hallak, took a different approach. Instead of waiting for data that does not yet exist, she built a controlled, artificial world inside a computer to test if modern tools could spot the connection if it were there. She created a simulation of 2,500 hypothetical human cases, each linked to a specific level of antibiotic residue in poultry. In this digital experiment, she programmed the rules so that higher residue levels would make the bacteria more likely to be resistant, mimicking the biological theory she wanted to test. She then asked a computer to learn the pattern from this data, using a method called machine learning, which allows software to find complex relationships without being explicitly told how to solve the puzzle.

The study focused on four common antibiotics and four different regions, creating a scenario where the computer had to decide if a human infection was resistant or not based on the drug type, the location, and the amount of residue present. To make the test realistic and difficult, the researcher intentionally designed the data so that the drug levels in resistant and non-resistant bacteria overlapped significantly. This meant the computer could not simply look at a single number and guess the answer; it had to understand the broader relationship between the residue and the outcome. The computer was given three different ways to solve the problem: a simple method that only looked at the type of drug and location, a standard statistical method, and a more advanced method known as a Random Forest, which is designed to handle complex, non-linear patterns.

When the results were analyzed, the simple method that ignored the residue levels performed no better than random guessing, achieving only about 63 percent accuracy. This confirmed that knowing just the type of antibiotic or the region was not enough to predict resistance. However, when the computer was allowed to use the actual residue levels and the bacterial drug sensitivity data, its performance jumped dramatically. The most advanced model, the optimized Random Forest, correctly identified the outcome in nearly 80 percent of the cases. It was particularly good at spotting the resistant cases, correctly flagging them 91 percent of the time. The computer also learned to weigh the evidence correctly; it determined that the amount of antibiotic residue was the single most important factor, far outweighing the type of drug or the geographic location.

To ensure the computer was not just memorizing the answers, the researchers tested it using a technique that splits the data into different groups to verify the results hold up consistently. The model remained stable and accurate across these tests. Furthermore, the researchers used a tool called SHAP to look inside the "black box" of the computer's decision-making. This tool revealed exactly why the computer made its choices, showing that high residue levels consistently pushed the prediction toward resistance, while low levels pushed it toward susceptibility. This transparency proved that the model was learning the specific biological rule that was programmed into the simulation, rather than finding a random trick in the data.

Despite these successes, the study comes with a crucial caveat. The entire dataset was synthetic; it was created by rules, not by measuring real patients or real bacteria. Therefore, the results do not prove that antibiotic residues in poultry are actually causing resistance in humans. Instead, they prove that the computer pipeline used in the study is capable of finding such a link if it exists. The research demonstrates that if scientists can one day gather real-world data that links poultry residue levels to human clinical cases, this same method could be used to screen for risks quickly and explainably. For now, the work stands as a proof of concept, showing that the tools are ready for the day when the real data becomes available to help protect public health.