Abstract
The Sport Concussion Assessment Tool (SCAT) could be improved by identifying critical subsets that maximize diagnostic accuracy and eliminate low information elements.
To identify optimal SCAT subsets for acute concussion assessment.
Using Concussion Assessment, Research, and Education (CARE) Consortium data, we compared student-athletes' and cadets' preinjury baselines (n = 2178) with postinjury assessments within 6 h (n = 1456) and 24 to 48 h (n = 2394) by considering demographics, symptoms, Standard Assessment of Concussion (SAC), and Balance Error Scoring System (BESS) scores. We divided data into training/testing (60%/40%) sets. Using training data, we integrated logistic regression with an engineering methodology-mixed integer programming-to optimize models with ≤4, 8, 12, and 16 variables (Opt-k). We also created models including only raw scores (Opt-RS-k) and symptom, SAC, and BESS composite scores (summary scores). We evaluated models using testing data.
At <6 h and 24 to 48 h, most Opt-k and Opt-RS-k models included the following symptoms: do not feel right, headache, dizziness, sensitivity to noise, and whether physical or mental activity worsens symptoms. Opt-k models included SAC concentration and delayed recall change scores. Opt-k models had lower Brier scores (BS) and greater area under the curve (AUC) (<6 h: BS = 0.072-0.089, AUC = 0.95-0.96; 24-48 h: BS = 0.085-0.093, AUC = 0.94-0.95) than Opt-RS-k (<6 h: BS = 0.082-0.087, AUC = 0.93-0.95; 24-48 h: BS = 0.095-0.099, AUC = 0.92-0.93) and summary score models (<6 h: BS = 0.14, AUC = 0.89; 24-48 h: BS = 0.15, AUC = 0.87).
We identified SCAT subsets that accurately assess acute concussion and improve administration time over the complete battery, highlighting the importance of eliminating "noisy" elements. These findings can direct clinicians to the SCAT components that are most sensitive to acute concussion.
