{
  "abstract": "Background and Importance Hospital pharmacists frequently assess adverse drug reactions (ADRs) in patients, a task that is time consuming and requires expert knowledge, especially in patients with polypharmacy. Large language models (LLMs) such as ChatGPT may accelerate the assessment of ADRs. With each new release of LLMs, performance is expected to improve, but systematic evaluation is lacking.Aim and Objectives The aim of this study was to investigate the performance of ChatGPT-4o compared to ChatGPT-5 in producing an accurate overview of ADRs.Material and Methods Thirty commonly prescribed cardiovascular drugs were analysed using an engineered prompt restricting both models to a national drug database. Across five iterations per drug, each model extracted ADRs and categorised them as common (1–10%) or very common (>10%). Outputs were manually validated.Accuracy (correct ADRs vs database), hallucinations (false ADRs), and omission errors (missed ADRs) were calculated. Descriptive statistics were used, and mean accuracy between ChatGPT releases was compared using independent t-tests.Results ChatGPT-5 substantially outperformed ChatGPT-4o. The mean accuracy of correctly summarising ADRs was 77.9% (SD 21.7%) for ChatGPT-4o vs. 97.4% (SD 5.6%) for ChatGPT-5 (p<0.001). In ChatGPT-4o, 7 of 30 drugs (23.3%) were fully accurate across all iterations compared to 21 drugs (70%) in ChatGPT-5.The most common errors were omissions in ChatGPT-4o (mean 19.4%, SD: 19.5%) vs 2.5% (SD: 5.5%) in ChatGPT-5. Hallucination errors (mean 0.5%, SD: 1.8%) only occurred in ChatGPT-4o.Conclusion and Relevance ChatGPT-5 clearly outperformed ChatGPT-4o in summarising ADRs of the tested drugs. Omission errors were the most frequently occurring errors, with less errors in ChatGPT-5. Hallucinations were not observed with ChatGPT-5. These findings suggest that ChatGPT-5, may offer a possible reliable tool for hospital pharmacists in detecting ADRs. Future research should address other medication groups and evaluate real-world clinical application.Conflict of Interest No conflict of interest",
  "authors": [
    {
      "affiliations": [
        "Maastricht University Medical Centre+, Department of Clinical Pharmacy and Toxicology, Maastricht, The Netherlands"
      ],
      "name": "R Van Der Zanden"
    },
    {
      "affiliations": [
        "Maastricht University Medical Centre+, Department of Clinical Pharmacy and Toxicology, Maastricht, The Netherlands"
      ],
      "name": "M Slikkerveer"
    },
    {
      "affiliations": [
        "Erasmus University Medical Centre, Department of Hospital Pharmacy, Rotterdam, The Netherlands"
      ],
      "name": "R Gündogan"
    },
    {
      "affiliations": [
        "Erasmus University Medical Centre, Department of Hospital Pharmacy, Rotterdam, The Netherlands"
      ],
      "name": "A Abdulla"
    },
    {
      "affiliations": [
        "Maastricht University Medical Centre+, Department of Clinical Pharmacy and Toxicology, Maastricht, The Netherlands"
      ],
      "name": "F Karapinar-Çarkit"
    },
    {
      "affiliations": [
        "Maastricht University Medical Centre+, Department of Clinical Pharmacy and Toxicology, Maastricht, The Netherlands"
      ],
      "name": "J Driessen"
    }
  ],
  "title": "5PSQ-078 Performance of ChatGPT-4o versus ChatGPT-5 in summarising adverse drug reactions",
  "uid": "72e10c9a-720b-556f-8fa9-f20856c741e5"
}