{
  "abstract": "Background We aimed to evaluate the impact of implementing an artificial intelligence (AI)-enabled acute ischaemic stroke triage system on workflow efficiency and transfer optimisation in a large academic healthcare network.Methods A prospectively maintained database was reviewed comparing equivalent time periods before and after AI-enabled triage platform implementation (January 2021–December 2022). The primary analysis compared workflow metrics between AI-enabled and non-AI spokes during the same calendar period (2022) to control for temporal confounding. Benjamini-Hochberg correction was applied for multiple comparisons, and analyses were adjusted for age and baseline National Institutes of Health Stroke Scale. Evaluated outcomes included door-in-door-out (DIDO) times, door-to-puncture (DTP) times, endovascular therapy (EVT) utilisation rates, cost analysis and clinical outcomes at discharge.Results The study included 4548 admissions with 844 EVT patients (394 pre-implementation, 450 post-implementation) across four hub centres. In the primary same-period analysis (2022), AI-enabled spokes demonstrated significantly shorter DIDO times compared with non-AI spokes (median 103 (92–118) vs 134 (103–162) min; adjusted difference −41.6 min (95% CI −60.9 to −24.1); p<0.001, Q<0.001) and shorter DTP times (21 (14–43) vs 40 (18–65) min; adjusted difference −10.9 min (95% CI −17.9 to −3.7); p=0.003, Q=0.009). A difference-in-differences analysis demonstrated that DIDO improvements were specific to AI-enabled spokes (−27 min; 95% CI −62 to −4; p=0.029). EVT utilisation was also significantly higher in AI-enabled versus non-AI spokes where AI-enabled spokes had increased EVT rates by +17.8% (39.3% to 57.1%) compared with +1.1% in non-AI spokes (41.3% to 42.4%, P interaction=0.006). DTP improvements were more pronounced at community hubs (86 (48–108) to 51 (22–77) min; adjusted difference −24.9 min; p=0.021, Q=0.041) compared with academic hubs (60 (23–87) to 55 (22–73) min; adjusted difference −15.5 min; p<0.001, Q=0.002). Subgroup analyses demonstrated consistent DIDO benefits across age, stroke severity and sex strata with no significant treatment effect heterogeneity (all P-interaction >0.05). Probabilistic cost analysis estimated savings of $3.6 million (95% CI $1.5M to $6.1M) per 1000 AI-enabled spoke transfers. Clinical outcomes, including functional status and mortality at discharge, were similar between groups (all Q>0.05).Conclusion Implementation of an AI-enabled triage platform was associated with significant reductions in workflow times and increased EVT utilisation, with effects specific to AI-enabled spokes rather than secular trends alone. The proportion of transfers who did not proceed to EVT decreased in AI-enabled spokes, though counterfactual outcomes for non-transferred patients remain unknown. Clinical outcomes at discharge were unchanged.",
  "authors": [
    {
      "affiliations": [
        "Departments of Neurology and Neurosurgery, University of Pittsburgh School of Medicine, Pittsburgh, Pennsylvania, USA"
      ],
      "name": "Mohamed F Doheim"
    },
    {
      "affiliations": [
        "Departments of Neurology and Neurosurgery, University of Pittsburgh School of Medicine, Pittsburgh, Pennsylvania, USA"
      ],
      "name": "Matthew Starr"
    },
    {
      "affiliations": [
        "Departments of Neurology and Neurosurgery, University of Pittsburgh School of Medicine, Pittsburgh, Pennsylvania, USA"
      ],
      "name": "Nirav R Bhatt"
    },
    {
      "affiliations": [
        "Departments of Neurology and Neurosurgery, University of Pittsburgh School of Medicine, Pittsburgh, Pennsylvania, USA"
      ],
      "name": "Marcelo Rocha"
    },
    {
      "affiliations": [
        "Departments of Neurology and Neurosurgery, University of Pittsburgh School of Medicine, Pittsburgh, Pennsylvania, USA"
      ],
      "name": "Alhamza R Al-Bayati"
    },
    {
      "affiliations": [
        "Departments of Neurology and Neurosurgery, University of Pittsburgh School of Medicine, Pittsburgh, Pennsylvania, USA"
      ],
      "name": "Abdullah Sultany"
    },
    {
      "affiliations": [
        "University of Pittsburgh Medical Centre, Pittsburgh, Pennsylvania, USA"
      ],
      "name": "Charles Romero"
    },
    {
      "affiliations": [
        "University of Pittsburgh Medical Centre, Pittsburgh, Pennsylvania, USA"
      ],
      "name": "Cynthia L Kenmuir"
    },
    {
      "affiliations": [
        "University of Pittsburgh Medical Centre, Pittsburgh, Pennsylvania, USA"
      ],
      "name": "Stephanie Henry"
    },
    {
      "affiliations": [
        "Departments of Neurology and Neurosurgery, University of Pittsburgh School of Medicine, Pittsburgh, Pennsylvania, USA"
      ],
      "name": "Raul G Nogueira"
    }
  ],
  "title": "Impact of an artificial intelligence–driven triage system on workflow and transfer efficiency: stratified analysis of 4548 admissions to four thrombectomy hubs receiving transfers from sixty spokes",
  "uid": "518fa08c-b8ff-5e9d-8333-16cfb6e166dd"
}