Optimizing the Combination of Common Clinical Concussion Batteries to Predict Persistent Postconcussion Symptoms in a Prospective Cohort of Concussed Youth.

BACKGROUND: Studies have evaluated individual factors associated with persistent postconcussion symptoms (PPCS) in youth concussion, but no study has combined individual elements of common concussion batteries with patient characteristics, comorbidities, and visio-vestibular deficits in assessing an optimal model to predict PPCS. PURPOSE: To determine the combination of elements from 4 commonly used clinical concussion batteries and known patient characteristics and comorbid risk factors that maximize the ability to predict PPCS. STUDY DESIGN: Cohort study; Level of evidence, 2. METHODS: We enrolled 198 concussed participants-87 developed PPCS and 111 did not-aged 8 to 19 years assessed within 14 days of injury from a suburban high school and the concussion program of a tertiary care academic medical center. We defined PPCS as a Post-Concussion Symptom Inventory (PCSI) score at 28 days from injury of ≥3 points compared with the preinjury PCSI score-scaled for younger children. Predictors included the individual elements of the visio-vestibular examination (VVE), Sport Concussion Assessment Tool, 5th Edition (SCAT-5), King-Devick test, and PCSI, in addition to age, sex, concussion history, and migraine headache history. The individual elements of these tests were grouped into interpretable factors using sparse principal component analysis. The 12 resultant factors were combined into a logistic regression and ranked by frequency of inclusion into the combined optimal model, whose predictive performance was compared with the VVE, initial PCSI, and the current existing predictive model (the Predicting and Prevention Postconcussive Problems in Pediatrics (5P) prediction rule) using the area under the receiver operating characteristic curve (AUC). RESULTS: A cluster of 2 factors (SCAT-5/PCSI symptoms and VVE near point of convergence/accommodation) emerged. A model fit with these factors had an AUC of 0.805 (95% CI, 0.661-0.929). This was a higher AUC point estimate, with overlapping 95% CIs, compared with the PCSI (AUC, 0.773 [95% CI, 0.617-0.912]), VVE (AUC, 0.736 [95% CI, 0.569-0.878]), and 5P Prediction Rule (AUC, 0.728 [95% CI, 0.554-0.870]). CONCLUSION: Among commonly used clinical assessments for youth concussion, a combination of symptom burden and the vision component of the VVE has the potential to augment predictive power for PPCS over either current risk models or individual batteries.

Maximizing the Accuracy of Adolescent Concussion Diagnosis Using Individual Elements of Common Standardized Clinical Assessment Tools.

CONTEXT: Multiple clinical evaluation tools exist for adolescent concussion with various degrees of correlation, presenting challenges for clinicians in identifying which elements of these tools provide the greatest diagnostic utility. OBJECTIVE: To determine the combination of elements from 4 commonly used clinical concussion batteries that maximize discrimination of adolescents with concussion from those without concussion. DESIGN: Cross-sectional study. SETTING: Suburban school and concussion program of a tertiary care academic center. PATIENTS OR OTHER PARTICIPANTS: A total of 231 participants with concussion (from a suburban school and a concussion program) and 166 participants without concussion (from a suburban school) between the ages of 13 and 19 years. MAIN OUTCOME MEASURE(S): Individual elements of the visio-vestibular examination (VVE), Sport Concussion Assessment Tool, fifth edition (SCAT5; including the modified Balance Error Scoring System), King-Devick test (K-D), and Postconcussion Symptom Inventory (PCSI) were evaluated. The 24 subcomponents of these tests were grouped into interpretable factors using sparse principal component analysis. The 13 resultant factors were combined with demographic and clinical covariates into a logistic regression model and ranked by frequency of inclusion into the ideal model, and the predictive performance of the ideal model was compared with each of the clinical batteries using the area under the receiver operating characteristic curve (AUC). RESULTS: A cluster of 4 factors (factor 1 [VVE saccades and vestibulo-ocular reflex], factor 2 [modified Balance Error Scoring System double-legged stance], factor 3 [SCAT5/PCSI symptom scores], and factor 4 [K-D completion time]) emerged. A model fit with the top factors performed as well as each battery in predicting concussion status (AUC = 0.816 [95% CI = 0.731, 0.889]) compared with the SCAT5 (AUC = 0.784 [95% CI = 0.692, 0.866]), PCSI (AUC = 0.776 [95% CI = 0.674, 0.863]), VVE (AUC = 0.711 [95% CI = 0.602, 0.814]), and K-D (AUC = 0.708 [95% CI = 0.590, 0.819]). CONCLUSIONS: A multifaceted assessment for adolescents with concussion, comprising symptoms, attention, balance, and the visio-vestibular system, is critical. Current diagnostic batteries likely measure overlapping domains, and the sparse principal component analysis demonstrated strategies for streamlining comprehensive concussion assessment across a variety of settings.

Neural networks for clustered and longitudinal data using mixed effects models.

Although most statistical methods for the analysis of longitudinal data have focused on retrospective models of association, new advances in mobile health data have presented opportunities for predicting future health status by leveraging an individual's behavioral history alongside data from similar patients. Methods that incorporate both individual-level and sample-level effects are critical to using these data to its full predictive capacity. Neural networks are powerful tools for prediction, but many assume input observations are independent even when they are clustered or correlated in some way, such as in longitudinal data. Generalized linear mixed models (GLMM) provide a flexible framework for modeling longitudinal data but have poor predictive power particularly when the data are highly nonlinear. We propose a generalized neural network mixed model that replaces the linear fixed effect in a GLMM with the output of a feed-forward neural network. The model simultaneously accounts for the correlation structure and complex nonlinear relationship between input variables and outcomes, and it utilizes the predictive power of neural networks. We apply this approach to predict depression and anxiety levels of schizophrenic patients using longitudinal data collected from passive smartphone sensor data.