Skip to main navigation Skip to search Skip to main content

Exploring Feature Ambiguity in TBI Diagnosis Using Missing Aware Machine Learning Models

  • Hamzah Dabool
  • , Hany Alashwal*
  • , Abdalla Z. Mohamed
  • , Ahmed A. Moustafa
  • *Corresponding author for this work

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › Research › peer-review

14 Downloads (Pure)

Abstract

Traumatic Brain Injury (TBI) presents significant diagnostic challenges due to overlapping symptomatology with other neurological and psychological conditions. In this study, we investigate the potential of machine learning—specifically, the Extreme Gradient Boosting (XGBoost) algorithm—for classifying TBI versus control cases using the ADNI-DOD dataset, which includes 200 participants (80 TBI and 120 controls). A key aspect of our methodology is the preservation of missing data, which comprises 35.21% of the dataset. Rather than applying traditional imputation techniques, we leverage XGBoost’s native capability to handle missing values, thereby maintaining the integrity and variability of real-world clinical data. Model evaluation was performed using stratified 3-fold cross-validation with Bayesian optimization used to tune hyperparameters such as learning rate, max depth, and subsampling ratios. The final model achieved a mean accuracy of 71.85%, precision of 68.41%, recall of 59.05%, F1-score of 63.31%, specificity of 80.65%, and an AUC of 77.57%. These results indicate stronger performance in correctly identifying control subjects, while sensitivity to TBI cases remained comparatively limited—reflecting the inherent clinical ambiguity and overlapping symptom profiles within the dataset. Feature importance analysis revealed that trauma-related and emotional health variables (e.g., PCLM TotalScore, CAPS Current, and CDGLOBAL) were among the most influential predictors. This pattern underscores the diagnostic complexity of TBI, where symptoms often intersect with those associated with PTSD, depression, chronic stress, or age-related cognitive changes. Overall, our findings highlight the utility of interpretable machine learning approaches for uncovering clinically meaningful patterns in heterogeneous datasets and emphasize the need for multimodal data integration and longitudinal tracking to support more precise TBI characterization.
Original languageEnglish
Title of host publicationBDE 2025 - Proceedings of the 2025 7th International Conference on Big Data Engineering
PublisherAssociation for Computing Machinery (ACM)
Pages9-13
Number of pages5
ISBN (Electronic)9798400719363
DOIs
Publication statusPublished - 20 Dec 2025
Event2025 7th International Conference on Big Data Engineering - University of the Ryukyus, Okinawa, Japan
Duration: 24 Sept 2025 → 26 Sept 2025

Publication series

NameBDE 2025 - Proceedings of the 2025 7th International Conference on Big Data Engineering

Conference

Conference2025 7th International Conference on Big Data Engineering
Abbreviated titleBDE 2025
Country/TerritoryJapan
CityOkinawa
Period24/09/25 → 26/09/25

UN SDGs

This output contributes to the following UN Sustainable Development Goals (SDGs)

  1. SDG 3 - Good Health and Well-being
    SDG 3 Good Health and Well-being

Fingerprint

Dive into the research topics of 'Exploring Feature Ambiguity in TBI Diagnosis Using Missing Aware Machine Learning Models'. Together they form a unique fingerprint.

Cite this