{
  "abstract": "Aims and Objectives Paracetamol overdose accounts for over 100,000 UK Emergency Department (ED) attendances annually. Acetylcysteine is effective, but does not prevent all cases of liver injury. To address this, novel therapies for paracetamol overdose are emerging, but we lack methods for prospectively identifying who will benefit, meaning we cannot currently deliver them efficiently and effectively at scale.We report a machine learning-derived risk stratification tool specifically designed for this decision, using only routinely ordered blood tests, allowing confident decision-making without increasing clinical workload for Emergency Medicine (EM) clinicians.Method and Design Analysing 4,705 patients across three hospitals (2008-2024), we developed stratified prediction models using elastic net regression, recognizing that patients presenting without liver injury (ALT≤50U/L) require distinct risk algorithms to early injury (51-1000U/L). Models combined age, sex, and 15 biomarkers routinely measured in ED overdose assessment (paracetamol level, FBC, U&Es, LFTs, and INR). Performance was assessed using a strictly held-out test cohort and benchmarked against the current gold-standard (ALT×APAP product). Study approvals: ethics, 25-EMREC-070; Caldicott Guardian, 24167.Results and Conclusion The stratified model outperformed ALT×APAP (AUROC 0.933 vs 0.830). Superiority was confirmed by Integrated Discrimination Improvement (0.154; p<0.001), Net Reclassification Index (0.675; p<0.001), and McNemar’s test (p<0.001). At a sensitivity of 75.9%, the model achieved 96.4% specificity and 99.4% NPV. Per 1,000 treated patients, the model selects 54 for escalated care, identifying 19 of the 25 who subsequently develop significant liver injury. ALT×APAP would flag 213 patients for equivalent sensitivity.Our model is immediately adoptable to support EM clinician-driven decision-making, and uses only currently available data. Consequently, 946 of 1,000 patients will not receive escalated care, while the remainder are a high-yield cohort for additional interventions. The tool provides objective patient selection criteria for novel therapies and clinical trials, and requires no additional blood tests, specialist investigations, or delays awaiting further results.",
  "authors": [
    {
      "affiliations": [
        "University of Edinburgh"
      ],
      "name": "Chris Humphries"
    }
  ],
  "title": "4470 A machine learning tool for identifying high-risk paracetamol overdose patients using routine ED data",
  "uid": "934d82b3-8616-58f6-a3c1-72df59c377e6"
}