The AI model predicted the eligibility of IVOS on a daily basis by using routinely collected clinical data. It was seen to mimic decision-making on antimicrobial ward rounds. 

The model was trained using public datasets like Medical Information Mart for Intensive Care (MIMIC) and Electronic Intensive Care Unit (eICU) followed by testing at Imperial College Healthcare NHS Trust. 

In terms of performance, it showed good accuracy, with an Area Under the Curve (AUC) ranging from 0.78 to 0.80 with low false positive rates across data sets.

The model was tested on 500 patients and the AI tool demonstrated consistent performance when applied prospectively which showed potential for real-world deployment. 

The tool was deployed as a web based Steward AI Platform which is accessible through a web based mobile app. The app enabled the clinicians to input key patient parameters (e.g., diagnosis, temperature, C-reactive protein) into a simple form. Advanced inputs were optional.

The app also offers real time predictions with explainability. It recommends if it's safe to switch from IV to oral antibiotics, and explains the reasoning behind its decision. 

The outputs can be annotated and shared via text/image format and also integrate them into medical records. 

The evaluation was conducted with Imperial College NHS Trust using pharmacist-reviewed gold-standard decisions from point prevalence surveys.

It was seen that there was an initial drop in the Area under the Curve (AUC) due to a mismatch between observed prescribing practices and optimal decisions

After alterations the updated model of the application aligned better with expert decisions and achieved lower false positive rates.