{
  "abstract": "Introduction The modified Rankin Scale (mRS) is an important ordinal scale used to assess degree of disability after a stroke. Its widely accepted usage gives it immense value in evaluating stroke outcomes. One proposed method for converting mRS to a continuous variable maps mRS to average EuroQol Group 5 (EQ-5D) scores among MR CLEAN trial participants.Objective To (1) train and test a machine learning model to predict stroke outcome, as assessed with different mRS transformations as outcome variables, and to (2) determine how modifying outcome scale affects feature importance.Methods A retrospective analysis was conducted on patients who received thrombectomy at Albany Medical Center (AMC) between 2019 and 2021. Three transformations were done to mRS: quadratic, exponential, and mapped average EQ-5D score. Four random forest models were generated using unchanged mRS and these three transformations. Predictor variables included history of TIA, MI, prior stroke, and renal insufficiency, distance from AMC, thrombectomy, Area Deprivation Index (ADI), age, race, and gender. Feature importances and R 2 were compared across the four models.Results Feature importance varied across models. Feature importance values are meaningful only as comparisons within the same model. In the EQ-5D model, age(0.54) was most influential, followed by distance (from AMC)(0.16) and ADI(0.15). The linear mRS model was driven by distance(0.53) and age(0.21) to a relatively lesser extent. Quadratic mRS was most influenced by distance(0.41) and prior stroke(0.21). The exponential mRS model valued distance(0.35), ADI(0.18), prior stroke(0.17), and age(0.11). R 2 was highest in unchanged mRS and Eq-5D mapped mRS, ahead of quadratic and exponential mRS.Conclusion EQ-5D mapped mRS as a continuous indicator for stroke outcome behaves markedly differently from traditional mRS. EQ-5D seems to be a better continuous representation of mRS than exponential and quadratic. Our findings agree with earlier work showing that EQ-5D mapped mRS may not provide real benefit over ordinal mRS. Future work may evaluate dichotomized mRS (0-1/2-6, 0-2/3-6,0-4/5-6), as an RF model outcome variable.Disclosures A. Goyal: None. M. Singh: None. O. Khazen: None. A. Custozzo: None. A. Paul: None.Abstract E-368 Figure 1Feature Importance Heatmap. Importance coefficients are meaningful only relative to other coefficients within the same model (values should only be compared relative to others in the same column)Abstract E-368 Figure 2Top 5 Features for Model Training. Compare values within the same model (labelled with the same color)",
  "authors": [
    {
      "affiliations": [
        "Albany Medical Center, Albany, NY"
      ],
      "name": "A Goyal"
    },
    {
      "affiliations": [
        "Albany Medical Center, Albany, NY"
      ],
      "name": "M Singh"
    },
    {
      "affiliations": [
        "Albany Medical Center, Albany, NY"
      ],
      "name": "O Khazen"
    },
    {
      "affiliations": [
        "Albany Medical Center, Albany, NY"
      ],
      "name": "A Custozzo"
    },
    {
      "affiliations": [
        "Albany Medical Center, Albany, NY"
      ],
      "name": "A Paul"
    }
  ],
  "title": "E-368 Evaluating the effect of linearizing transformations to modified rankin scale on random forest model prediction",
  "uid": "fe57ef92-0029-5851-9b70-2b0495e7a5a5"
}