2100 Predicting Hospitalization and Clinical Outcomes in Diffuse Axonal Injury Using Machine Learning Lesion Segmentation

INTRODUCTION: Diffuse Axonal Injury (DAI) is graded on MRI using the Adams Classification, which has limited prognostic utility. Machine Learning (ML) is a promising tool to detect more granular lesions to predict clinical outcomes. METHODS: The study included patients diagnosed with DAI using MRI who were hospitalized between July 2012 and March 2022 without midline shift or surgical intervention. ML was used to segment the DAI lesions on SWI images on MRI and correlated with anatomic parcellation maps to identify lesion location, number, and volume. Admission GCS and Head Injury on ISS were collected as well as mortality, length of hospital stay (LOS), days in the ICU, and days requiring ventilation. Multiple linear regression analyses with t-statistics were used with p=0.05 to assess correlation of lesion burden with the clinical variables. RESULTS: Our study included 229 patients (193 male; mean age 38.9 ± 16.6 years). Admission GCS was associated with lesions in the calcarine cortex (p=0.04) and subcallosal area (p=0.04). Head Injury on ISS was associated with lesions in the posterior insula (p=0.03), and number (p>0.02) and volume (p=0.001) of lesions. Mortality was associated with overall number (p=0.03) and volume of lesions (p=0.002). Marginally increased LOS was observed for lesions in the anterior insula (p=0.03). A greater number of ICU days was observed for lesions in the basal forebrain (p=0.003) and supplementary motor cortex (p=0.04). Higher number of ventilation days was required for lesions in the basal forebrain (p=0.02) with relatively fewer for those in the putamen (p=0.006), inferior frontal gyrus (p=0.04), supplementary motor cortex (p=0.04), and supramarginal gyrus (p=0.05). CONCLUSIONS: A ML algorithm segmenting granular lesions in DAI correlates with factors related to hospitalization and mortality.