{
  "abstract": "Background Mechanical thrombectomy improves functional outcomes in acute large-vessel occlusion stroke, yet early prognostication remains challenging. Existing models often require post-procedural data—24-hour NIHSS, laboratory indices, final infarct volumes—unavailable when clinicians must counsel families immediately after the procedure. Complex machine learning models further limit bedside adoption by requiring computational tools. We developed a machine learning-derived bedside risk score using only variables available upon procedure completion.Methods We conducted a single-center retrospective study of 421 consecutive thrombectomy patients (January 2022-June 2024). All post-operative variables were excluded to ensure predictions could be generated immediately. Gradient boosting machine models with recursive feature elimination (GBM-RFE) identified optimal predictors from pre- and intra-operative variables. An ordinal logistic regression model predicted the full modified Rankin Scale (mRS) spectrum (0-6), and coefficients were transformed into integer points (scaling factor of 20) to yield a simplified arithmetic score. Data were split 80/20 into discovery (n=337) and validation (n=84) cohorts. Performance was assessed by AUC, sensitivity, and specificity for functional independence (mRS ≤2) at 90 days.Results Of 421 patients, 36.6% achieved functional independence at 90 days. GBM-RFE identified 7 optimal features; NIHSS, age, and BMI dominated. The ordinal model achieved AUC 0.859 (95% CI 0.777-0.942) with 96.2% sensitivity and 69.0% specificity. Two variables with coefficients rounding to zero were eliminated (BMI, first-pass success), yielding a 5-variable equation: Risk Score = (2 × NIHSS) + Age + (22 × Cancer History) + (2 × Number of Passes) − (7 × TICI Score). At threshold ≥86, the score achieved AUC 0.837 (95% CI 0.750-0.923) with 96.2% sensitivity and 69.0% specificity ( figure 1). The 0.022 AUC difference from the full model confirms minimal information loss from simplification.Conclusions This explainable post-procedural risk score enables immediate prognostication after mechanical thrombectomy without computational tools. By restricting inputs to the procedural note, the score addresses a critical temporal gap in existing models. High sensitivity (96.2%) prioritizes identification of at-risk patients, supporting intensified monitoring and realistic family counseling. External validation is needed.Disclosures J. Keller: None. I. Bediako: None. E. Ozcan: None. A.Q. Wu: None. C. Sollenberger: None. Z. Hoglund: None. J. Catapano: None. J. Burkhardt: None. S. Kandregula: None. R. Rahmani: None. V. Srinivasan: None.Abstract E-190 Figure 1Abstract E-190 Table 1Model development and performance comparisonMetricGBM-RFEOrdinal Logistic RegressionRisk ScoreAUC0.8100.8590.83795% CI0.715 - 0.9060.777 - 0.9420.750 - 0.923Sensitivity69.2%96.2%96.2%Specificity81.0%69.0%69.0%",
  "authors": [
    {
      "affiliations": [
        "Neurosurgery, Perelman School of Medicine at the University of Pennsylvania, Philadelphia, PA"
      ],
      "name": "J Keller"
    },
    {
      "affiliations": [
        "Neurosurgery, Perelman School of Medicine at the University of Pennsylvania, Philadelphia, PA"
      ],
      "name": "I Bediako"
    },
    {
      "affiliations": [
        "Neurosurgery, Perelman School of Medicine at the University of Pennsylvania, Philadelphia, PA"
      ],
      "name": "E Ozcan"
    },
    {
      "affiliations": [
        "Neurosurgery, Perelman School of Medicine at the University of Pennsylvania, Philadelphia, PA"
      ],
      "name": "AQ Wu"
    },
    {
      "affiliations": [
        "Neurosurgery, Perelman School of Medicine at the University of Pennsylvania, Philadelphia, PA"
      ],
      "name": "C Sollenberger"
    },
    {
      "affiliations": [
        "Neurosurgery, Perelman School of Medicine at the University of Pennsylvania, Philadelphia, PA"
      ],
      "name": "Z Hoglund"
    },
    {
      "affiliations": [
        "Neurosurgery, Perelman School of Medicine at the University of Pennsylvania, Philadelphia, PA"
      ],
      "name": "J Catapano"
    },
    {
      "affiliations": [
        "Neurosurgery, Perelman School of Medicine at the University of Pennsylvania, Philadelphia, PA"
      ],
      "name": "J Burkhardt"
    },
    {
      "affiliations": [
        "Neurosurgery, Perelman School of Medicine at the University of Pennsylvania, Philadelphia, PA"
      ],
      "name": "S Kandregula"
    },
    {
      "affiliations": [
        "Neurosurgery, Perelman School of Medicine at the University of Pennsylvania, Philadelphia, PA"
      ],
      "name": "R Rahmani"
    },
    {
      "affiliations": [
        "Neurosurgery, Perelman School of Medicine at the University of Pennsylvania, Philadelphia, PA"
      ],
      "name": "V Srinivasan"
    }
  ],
  "title": "E-190 Machine-learning derived bedside risk score for immediate post-procedural prognostication after mechanical thrombectomy",
  "uid": "a46f8d30-b7c2-56ea-b798-4a85c73ef371"
}