Evaluating ChatGPT-5 for Detecting Potential Drug-Drug Interactions in Intensive Care: A Comparative Analysis with a Clinical Decision Support System

Background:

Polypharmacy is a frequent challenge in intensive care units (ICUs), where critically ill patients are exposed to multiple concurrent medications. This situation significantly increases the risk of potential drug-drug interactions (pDDIs), which may contribute to adverse drug events, prolonged ICU stays, and higher morbidity and mortality rates. Ensuring timely and accurate detection of pDDIs is therefore a cornerstone of patient safety in critical care settings. Traditional rule-based clinical decision support systems (CDSSs), such as the UpToDate Drug Interaction Checker, provide standardized alerts but may have limitations in contextual interpretation and adaptability. Recently, large language models (LLMs), such as ChatGPT-4.0, have emerged as advanced tools with natural language processing capabilities, potentially offering a novel approach to medication safety.

Objective:

This study aims to compare the performance of ChatGPT-4.0 with the UpToDate Drug Interaction Checker in identifying, classifying, and interpreting potential drug-drug interactions within real ICU patient medication orders.

Methods:

A retrospective dataset of ICU patient orders will be systematically analyzed using both ChatGPT-4.0 and the UpToDate Drug Interaction Checker. Each potential interaction will be assessed for sensitivity, specificity, accuracy, and clinical relevance. Discrepancies between the two systems will be documented and evaluated by independent critical care experts. Statistical analysis will be performed to compare detection rates and the qualitative depth of interaction explanations provided by each tool.

Expected Outcomes:

The study is expected to determine whether ChatGPT-4.0, as an AI-based system, can enhance the detection of clinically meaningful drug-drug interactions compared to traditional CDSS. The results may inform future integration of generative AI into ICU clinical workflows and contribute to safer pharmacotherapy practices in critical care.

Conclusion:

By directly comparing a state-of-the-art LLM with a widely used rule-based system, this study seeks to highlight the strengths, weaknesses, and potential clinical implications of generative AI in the domain of drug safety.

Inclusion Criteria:

• Patients aged 18 years or older
• Admission to the intensive care unit (ICU) for at least 48 hours
• Receipt of five or more medications concurrently during ICU stay
• Availability of complete clinical data and medication lists

Exclusion Criteria:

• Cases with incomplete medication or interaction data
• Patients receiving experimental or unproven drugs
• Pediatric patients or those with pregnancy

The study population consists of adult patients (≥18 years) admitted to the intensive care unit (ICU) for at least 48 hours at a tertiary care hospital. Eligible patients received four or more concurrent medications during their ICU stay, and only those with complete clinical and medication records were included. Patients with incomplete drug or interaction data, those receiving experimental or unproven drugs, as well as pediatric or pregnant patients, were excluded

This is a point-prevalence study using ICU patient records. Due to ethical and privacy considerations, individual participant data will not be shared outside the study team. De-identified individual participant data underlying the study findings may be shared upon reasonable request.