Limits...
Predicting Outcome 12 Months after Mild Traumatic Brain Injury in Patients Admitted to a Neurosurgery Service

View Article: PubMed Central - PubMed

ABSTRACT

Objective: Accurate outcome prediction models for patients with mild traumatic brain injury (MTBI) are key for prognostic assessment and clinical decision-making. Using multivariate machine learning, we tested the unique and added predictive value of (1) magnetic resonance imaging (MRI)-based brain morphometric and volumetric characterization at 4-week postinjury and (2) demographic, preinjury, injury-related, and postinjury variables on 12-month outcomes, including global functioning level, postconcussion symptoms, and mental health in patients with MTBI.

Methods: A prospective, cohort study of patients (n = 147) aged 16–65 years with a 12-month follow-up. T1-weighted 3 T MRI data were processed in FreeSurfer, yielding accurate cortical reconstructions for surface-based analyses of cortical thickness, area, and volume, and brain segmentation for subcortical and global brain volumes. The 12-month outcome was defined as a composite score using a principal component analysis including the Glasgow Outcome Scale Extended, Rivermead Postconcussion Questionnaire, and Patient Health Questionnaire-9. Using leave-one-out cross-validation and permutation testing, we tested and compared three prediction models: (1) MRI model, (2) clinical model, and (3) MRI and clinical combined.

Results: We found a strong correlation between observed and predicted outcomes for the clinical model (r = 0.55, p < 0.001). The MRI model performed at the chance level (r = 0.03, p = 0.80) and the combined model (r = 0.45, p < 0.002) were slightly weaker than the clinical model. Univariate correlation analyses revealed the strongest association with outcome for postinjury factors of posttraumatic stress (Posttraumatic Symptom Scale-10, r = 0.61), psychological distress (Hospital Anxiety and Depression Scale, r = 0.52), and widespread pain (r = 0.43) assessed at 8 weeks.

Conclusion: We found no added predictive value of MRI-based measures of brain cortical morphometry and subcortical volumes over and above demographic and clinical features.

No MeSH data available.


Association between predicted and observed outcome. (A) Imaging model mean squared error (MSE): 4.05, (B) clinical model MSE: 2.58, and (C) model with imaging + clinical data MSE: 2.91.
© Copyright Policy
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC5385465&req=5

Figure 2: Association between predicted and observed outcome. (A) Imaging model mean squared error (MSE): 4.05, (B) clinical model MSE: 2.58, and (C) model with imaging + clinical data MSE: 2.91.

Mentions: Figure 2 shows the association between predicted and observed outcomes. With an MSE of 2.58 and a correlation between observed and predicted outcomes (r = 0.55), the clinical model (model B) yielded significantly better prediction accuracy than the imaging model (MSE 4.05, r = 0.03). Predictions in model A (MRI data only) did not perform significantly above chance (p = 0.80, p value obtained across 10,000 permutations), whereas those in model B (clinical data only) were well above chance (p < 0.001). The prediction accuracy of the imaging model with age incorporated as a feature did not improve (MSE: 4.06). Adding MRI and clinical data in one model (model C) did not increase performance (model 3, MSE 2.91, r = 0.45, p < 0.002) compared to the clinical model.


Predicting Outcome 12 Months after Mild Traumatic Brain Injury in Patients Admitted to a Neurosurgery Service
Association between predicted and observed outcome. (A) Imaging model mean squared error (MSE): 4.05, (B) clinical model MSE: 2.58, and (C) model with imaging + clinical data MSE: 2.91.
© Copyright Policy
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC5385465&req=5

Figure 2: Association between predicted and observed outcome. (A) Imaging model mean squared error (MSE): 4.05, (B) clinical model MSE: 2.58, and (C) model with imaging + clinical data MSE: 2.91.
Mentions: Figure 2 shows the association between predicted and observed outcomes. With an MSE of 2.58 and a correlation between observed and predicted outcomes (r = 0.55), the clinical model (model B) yielded significantly better prediction accuracy than the imaging model (MSE 4.05, r = 0.03). Predictions in model A (MRI data only) did not perform significantly above chance (p = 0.80, p value obtained across 10,000 permutations), whereas those in model B (clinical data only) were well above chance (p < 0.001). The prediction accuracy of the imaging model with age incorporated as a feature did not improve (MSE: 4.06). Adding MRI and clinical data in one model (model C) did not increase performance (model 3, MSE 2.91, r = 0.45, p < 0.002) compared to the clinical model.

View Article: PubMed Central - PubMed

ABSTRACT

Objective: Accurate outcome prediction models for patients with mild traumatic brain injury (MTBI) are key for prognostic assessment and clinical decision-making. Using multivariate machine learning, we tested the unique and added predictive value of (1) magnetic resonance imaging (MRI)-based brain morphometric and volumetric characterization at 4-week postinjury and (2) demographic, preinjury, injury-related, and postinjury variables on 12-month outcomes, including global functioning level, postconcussion symptoms, and mental health in patients with MTBI.

Methods: A prospective, cohort study of patients (n&thinsp;=&thinsp;147) aged 16&ndash;65&thinsp;years with a 12-month follow-up. T1-weighted 3&thinsp;T MRI data were processed in FreeSurfer, yielding accurate cortical reconstructions for surface-based analyses of cortical thickness, area, and volume, and brain segmentation for subcortical and global brain volumes. The 12-month outcome was defined as a composite score using a principal component analysis including the Glasgow Outcome Scale Extended, Rivermead Postconcussion Questionnaire, and Patient Health Questionnaire-9. Using leave-one-out cross-validation and permutation testing, we tested and compared three prediction models: (1) MRI model, (2) clinical model, and (3) MRI and clinical combined.

Results: We found a strong correlation between observed and predicted outcomes for the clinical model (r&thinsp;=&thinsp;0.55, p&thinsp;&lt;&thinsp;0.001). The MRI model performed at the chance level (r&thinsp;=&thinsp;0.03, p&thinsp;=&thinsp;0.80) and the combined model (r&thinsp;=&thinsp;0.45, p&thinsp;&lt;&thinsp;0.002) were slightly weaker than the clinical model. Univariate correlation analyses revealed the strongest association with outcome for postinjury factors of posttraumatic stress (Posttraumatic Symptom Scale-10, r&thinsp;=&thinsp;0.61), psychological distress (Hospital Anxiety and Depression Scale, r&thinsp;=&thinsp;0.52), and widespread pain (r&thinsp;=&thinsp;0.43) assessed at 8&thinsp;weeks.

Conclusion: We found no added predictive value of MRI-based measures of brain cortical morphometry and subcortical volumes over and above demographic and clinical features.

No MeSH data available.