INS NYC 2024 Program

Poster	Poster Session 03 Program Schedule 02/15/2024 09:30 am - 10:40 am Room: Majestic Complex (Posters 61-120) Poster Session 03: Neurotrauma \| Neurovascular Final Abstract #78 Neural Network-Based Classification of Pediatric mTBI: Advancing Diagnostic Accuracy Deepan Tripathy, Georgia State University, Atlanta, India Michael Lee, Georgia State University, Atlanta, United States Catherine Lebel, University of Calgary, Calgary, Canada Adrian Onicas, IMT School for Advanced Studies Lucca, Lucca, Italy Nishard Abdeen, University of Ottawa, Ottawa, Canada Miriam Beauchamp, CHU Sainte-Justine Hospital Research Center, Montréal, Canada Christian Beaulieu, University of Alberta, Edmonton, Canada Bruce Bjornson, University of British Columbia, Vancouver, Canada William Craig, University of Alberta and Stollery Childrens Hospital, Edmonton, Canada Mathieu Dehaes, Institute of Biomedical Engineering, University of Montreal, Montréal, Canada Quynh Doan, University of British Columbia, Vancouver, Canada Sylvain Deschenes, University of Montreal, Montréal, Canada Stephen Freedman, University of Calgary, Calgary, Canada Bradley Goodyear, University of Calgary, Calgary, Canada Jocelyn Gravel, University of Montréal, Montréal, Canada Andr ́ee-Anne Ledoux, University of Ottawa, & Childrens Hospital of Eastern Ontario Research Institute, Ottawa, Canada Roger Zemek, University of Ottawa, & Childrens Hospital of Eastern Ontario Research Institute, Ottawa, Canada Keith Yeates, University of Calgary, Calgary,, Canada Dong Ye, Georgia State University, Atlanta, United States Ashley Ware, Georgia State University, Atlanta, United States Category: Concussion/Mild TBI (Child) Keyword 1: child brain injury Keyword 2: neuroimaging: structural Objective: Millions of children in North America sustain mild traumatic brain injury (mTBI; i.e., concussion) annually. Although mTBI typically results in subtle and diffuse brain tissue alterations, no diagnostic or prognostic biomarkers currently are available for routine clinical use. However, combining advanced neuroimaging and machine learning techniques could help to identify biomarkers, leading to greater understanding of the neurobiology of mTBI and enhancing clinical care. This study sought to examine whether a machine learning detection framework could use brain macrostructural metrics to accurately classify pediatric mTBI relative to mild orthopedic injury (OI). Participants and Methods: Children (N=967) aged 8-16.99 years with concussion or mild OI were recruited from 5 pediatric hospital emergency departments across Canada within 48h post-injury and completed a post-acute MRI scan (~10 days post-injury). Regional cortical volume and thickness, and subcortical volume, were derived for 57 unique brain regions based on the MICCAI Pediatric Atlas using the Advanced Normalization Tools (ANTs) segmentation pipeline, and then harmonized for scanner differences using COMBAT. For each participant, harmonized brain metrics, age at injury, and sex were then combined and used as textual classification features for a machine learning neural network that consisted of three layers. The first two layers had 64 and 32 nodes, respectively, and the third layer had an output layer with a SoftMax activation function for classification. This machine learning model uses adaptive learning to fine-tune the network at each layer to optimize binary classification prediction of groups. To ensure the robustness and generalization of the learning-based machine learning model, a 5-fold cross-validation strategy was employed. A second modeling approach used a supervised Naive Bayes algorithm, a probabilistic approach that does not have a learning network component and classifies groups using traditional classification techniques (e.g., logistic regression), for comparison to the learning-based machine learning model. Results: After rigorous quality assurance of MRI scans, the final sample included 541 children (354 mTBI/187 OI); mean age=12.42 years, SD=2.38; 59% male). Overall group classification accuracy for the neural network approach was 84%, with 91% sensitivity and 71% specificity. The supervised Naive Bayes algorithm yielded a classification accuracy of 65%, with 99% sensitivity and 1% specificity. Conclusions: Our innovative machine learning framework utilized network learning and successfully classified pediatric mTBI relative to mild OI with higher accuracy, including better specificity and sensitivity, than the traditional classification model. Specifically, the classification obtained from the neural network machine learning approach outperformed the basic supervised approach by 19%. These findings are clinically relevant and suggest that combining machine learning and advanced structural MRI techniques could help to identify a neuroimaging biomarker of pediatric mTBI. Explanatory studies are necessary next steps for understanding the specific neurobiological features that best differentiated groups, e.g., specific regional metrics. This underscores the need for future research to implement explainable artificial intelligence. Future study should also investigate whether the inclusion of multimodal neuroimaging metrics enhances classification of concussion as compared with mild OI in children.

Poster

02/15/2024
09:30 am - 10:40 am
Room: Majestic Complex (Posters 61-120)

Poster Session 03: Neurotrauma | Neurovascular

Final Abstract #78

Neural Network-Based Classification of Pediatric mTBI: Advancing Diagnostic Accuracy

Deepan Tripathy, Georgia State University, Atlanta, India
Michael Lee, Georgia State University, Atlanta, United States
Catherine Lebel, University of Calgary, Calgary, Canada
Adrian Onicas, IMT School for Advanced Studies Lucca, Lucca, Italy
Nishard Abdeen, University of Ottawa, Ottawa, Canada
Miriam Beauchamp, CHU Sainte-Justine Hospital Research Center, Montréal, Canada
Christian Beaulieu, University of Alberta, Edmonton, Canada
Bruce Bjornson, University of British Columbia, Vancouver, Canada
William Craig, University of Alberta and Stollery Childrens Hospital, Edmonton, Canada
Mathieu Dehaes, Institute of Biomedical Engineering, University of Montreal, Montréal, Canada
Quynh Doan, University of British Columbia, Vancouver, Canada
Sylvain Deschenes, University of Montreal, Montréal, Canada
Stephen Freedman, University of Calgary, Calgary, Canada
Bradley Goodyear, University of Calgary, Calgary, Canada
Jocelyn Gravel, University of Montréal, Montréal, Canada
Andr ́ee-Anne Ledoux, University of Ottawa, & Childrens Hospital of Eastern Ontario Research Institute, Ottawa, Canada
Roger Zemek, University of Ottawa, & Childrens Hospital of Eastern Ontario Research Institute, Ottawa, Canada
Keith Yeates, University of Calgary, Calgary,, Canada
Dong Ye, Georgia State University, Atlanta, United States
Ashley Ware, Georgia State University, Atlanta, United States

Category: Concussion/Mild TBI (Child)

Keyword 1: child brain injury
Keyword 2: neuroimaging: structural

Objective:

Millions of children in North America sustain mild traumatic brain injury (mTBI; i.e., concussion) annually. Although mTBI typically results in subtle and diffuse brain tissue alterations, no diagnostic or prognostic biomarkers currently are available for routine clinical use. However, combining advanced neuroimaging and machine learning techniques could help to identify biomarkers, leading to greater understanding of the neurobiology of mTBI and enhancing clinical care. This study sought to examine whether a machine learning detection framework could use brain macrostructural metrics to accurately classify pediatric mTBI relative to mild orthopedic injury (OI).

Participants and Methods:

Children (N=967) aged 8-16.99 years with concussion or mild OI were recruited from 5 pediatric hospital emergency departments across Canada within 48h post-injury and completed a post-acute MRI scan (~10 days post-injury). Regional cortical volume and thickness, and subcortical volume, were derived for 57 unique brain regions based on the MICCAI Pediatric Atlas using the Advanced Normalization Tools (ANTs) segmentation pipeline, and then harmonized for scanner differences using COMBAT. For each participant, harmonized brain metrics, age at injury, and sex were then combined and used as textual classification features for a machine learning neural network that consisted of three layers. The first two layers had 64 and 32 nodes, respectively, and the third layer had an output layer with a SoftMax activation function for classification. This machine learning model uses adaptive learning to fine-tune the network at each layer to optimize binary classification prediction of groups. To ensure the robustness and generalization of the learning-based machine learning model, a 5-fold cross-validation strategy was employed. A second modeling approach used a supervised Naive Bayes algorithm, a probabilistic approach that does not have a learning network component and classifies groups using traditional classification techniques (e.g., logistic regression), for comparison to the learning-based machine learning model.

Results:

After rigorous quality assurance of MRI scans, the final sample included 541 children (354 mTBI/187 OI); mean age=12.42 years, SD=2.38; 59% male). Overall group classification accuracy for the neural network approach was 84%, with 91% sensitivity and 71% specificity. The supervised Naive Bayes algorithm yielded a classification accuracy of 65%, with 99% sensitivity and 1% specificity.

Conclusions:

Our innovative machine learning framework utilized network learning and successfully classified pediatric mTBI relative to mild OI with higher accuracy, including better specificity and sensitivity, than the traditional classification model. Specifically, the classification obtained from the neural network machine learning approach outperformed the basic supervised approach by 19%. These findings are clinically relevant and suggest that combining machine learning and advanced structural MRI techniques could help to identify a neuroimaging biomarker of pediatric mTBI. Explanatory studies are necessary next steps for understanding the specific neurobiological features that best differentiated groups, e.g., specific regional metrics. This underscores the need for future research to implement explainable artificial intelligence. Future study should also investigate whether the inclusion of multimodal neuroimaging metrics enhances classification of concussion as compared with mild OI in children.