Use of artificial intelligence in diagnosis and prognosis of traumatic brain injury: a scoping review

Abstract

Traumatic Brain Injury (TBI) has been increasingly recognized as a leading cause of death and disability worldwide.

Objective

To summarize clinical applications of artificial intelligence, including machine learning and deep learning, in the diagnosis and prognosis of traumatic brain injury.

Methods

The authors conducted a scoping review of original clinical research studies on humans published in English after January 1, 2014. A search was performed using PubMed, including PMC, MEDLINE, and Bookshelf. The search terms were applied to the title field and included: (TBI) AND (Artificial Intelligence OR Machine Learning OR Deep Learning). Studies meeting inclusion criteria were screened and selected for review. The reference lists of the included studies were also screened to identify any additional eligible articles.

Results

Of 493 studies identified, seven met the inclusion criteria and were included in the analysis, which summarizes study title, publication year, study objective, key findings, and conclusions.

Conclusion

Artificial intelligence shows promise in aiding diagnosis and improving prognostic insights in traumatic brain injury. Although few clinical trials have been conducted, early results are encouraging. Future progress will require more clinical studies and efforts to address the current limitations of AI tools in medicine.

Introduction

Over the past decade, Traumatic Brain Injury (TBI) has emerged as a leading cause of death and disability worldwide. Of those with head injury, between 70 and 90% will be diagnosed with mild TBI. An estimated 69 million people experience TBI annually, with individuals aged 15–24 years representing the highest number of emergency department (ED) visits for this condition [1, 2]. In the United States, EDs manage more than 25 million injury-related visits, including suspected TBIs. However, it is individuals over 75 years of age who suffer the highest mortality rates following hospital admission [1, 2]. The impact of TBI extends beyond individual health, contributing to significant economic consequences. The global costs associated with TBI exceed $400 billion annually, placing a substantial financial burden on both families and societies worldwide [1–3].

Traumatic brain injury (TBI) results from both primary and secondary mechanisms. Primary injury occurs from external forces, such as falls, road accidents, sports injuries, and assaults. Secondary injury follows biochemical and cellular processes like inflammation and oxidative stress, which can worsen brain damage over time. TBI’s increasing prevalence underscores its significance as a public health concern. It is classified by severity using the Glasgow Coma Scale (GCS): mild (GCS 13–15), moderate (GCS 9–12), and severe (GCS 3–8), and can also be categorized as closed or penetrating [3]. Symptoms commonly include headache and dizziness, with the duration depending on severity. Other potential symptoms involve cognitive issues (e.g., attention deficits, memory loss), language difficulties (e.g., comprehension, communication), and emotional changes (e.g., depression, personality shifts) [3]. Repeated TBIs can increase the risk of neurodegenerative diseases later in life [3].

Early and accurate diagnosis is crucial in forecasting and treating TBI. Time sensitive and targeted interventions based on severity of the injury are essential for optimizing an individual’s health outcomes. However, current diagnostic methods face several challenges and limitations. Commonly used techniques, such as clinical assessments or neuroimaging tools like CT scans and MRI scans, often fail to detect subtle or mild brain injuries. Consequently, misdiagnosis or delayed treatment due to these limitations can adversely affect patient outcomes [4].

Artificial Intelligence (AI) holds significant potential for addressing the challenges and limitations in diagnosing traumatic brain injury (TBI) due to its ability to enhance diagnostic capabilities. AI utilizes sophisticated algorithms to process vast amounts of medical data. One such system, machine learning (ML), is a subset of AI. ML operates independently of pre-established algorithms, developing its own functions and validating them against known outcomes. These functions offer an alternative approach to data processing, often with predictive accuracy that surpasses the original algorithms used in training [5]. AI-powered tools can analyze complex patterns in imaging data and predict outcomes in ways that physicians may not be able to, thus supporting clinical decision making. This improves the diagnosis and treatment of TBI, ultimately leading to better health outcomes for patients [6].

Methods

Search strategy and study selection

The scoping review utilized the 2020 PRISMA-Scr guidelines and was conducted in June 2024 using PubMed, which includes the PMC, MEDLINE, and Bookshelf databases. The search query was: “(TBI OR traumatic brain injury) AND (Artificial Intelligence OR Machine Learning OR Deep Learning).” This search focused on articles where the keywords were central to the study. Only articles retrieved through this search were included in the review [Figure 1]. Additional publications were identified by reviewing the studies cited as references and systematic reviews identified through the search.

Fig. 1.

PRISMA diagram for the current scoping review

Eligibility criteria

Eligibility was determined using the Population, Intervention, Control, Outcome (PICO) framework. Studies were included if they met the following criteria:

• Population: Participants with or being assessed for TBI.

• Intervention: Use of AI, machine learning, or deep learning algorithms to process medical images.

• Control: Studies with a control group or alternative predictive model.

• Outcome: Outcomes must be measured against the control group or compared with other predictive models.

Only studies in English, published within the last 10 years, and focused on human subjects with TBI were included. Editorials, case reports, conference abstracts, reviews and practice guidelines on the management of TBI were excluded.

Study selection

The literature search retrieved 493 articles. Of these, 456 were excluded because they were not clinical trials. Twelve other articles were found by searching the bibliographies of the included research articles and relevant reviews. All 12 were excluded after assessing the full text. Seven studies were included in this review. Figure 1 shows the flow diagram for the process of exclusion and inclusion of the publications based on relevance.

After removing duplicates, all abstracts were independently screened by two reviewers (C.M. and M.G.) for eligibility. Full texts of selected abstracts were also assessed by the same two independent reviewers (C.M. and M.G). Included were articles that met the eligibility criteria. Conflict assessments were resolved through discussion. When needed, a third author (S.N.) was solicited to reach a consensus. Of the selected articles, data was taken to make a chart including all of the reviewed studies: titles, year, objective, key findings, and conclusions.

Data extraction and quality assessment

Data were extracted from the selected articles by two independent reviewers (C.M. and M.G.) Studies were classified based on their methodological rigor, and only those meeting predefined quality criteria were included in the final analysis.

The following information was extracted from each study:

• Study characteristics (e.g., year of publication, country, study design).

• Objectives of the study.

• Conclusion of the study.

• Details of the AI models used (e.g., algorithm type, data input, and method of processing).

• Outcomes measured (e.g., diagnostic accuracy, predictive value).

Exclusion criteria

Studies were excluded if they did not meet the eligibility criteria outlined above. Specifically, studies that were not clinical, involved non-human subjects, lacked a control group or alternative predictive model, or were published in languages other than English were not considered. Additionally, studies without clear definitions of outcomes or those that did not report sufficient data for analysis were excluded.

Results

A total of 7 studies that met the inclusion criteria were identified [Figure1]. Each of the studies is summarized below [Table 1] by study title, publication year, study objective, key findings and conclusion.

Table 1.

Summary of AI models for TBI diagnosis and prognosis

Title	Year	Objective & Methods	Results	Conclusion
Statistical machine learning to identify traumatic brain injury (TBI) from structural disconnections of white matter networks Mitra, et al. [9]	2016	Objective: The study aimed to utilize statistical machine learning to identify traumatic brain injury (TBI) by analyzing structural disconnections in white matter networks. Methods: The researchers used higher-order diffusion models to map white matter connections between 116 cortical and subcortical regions. Probabilistic tractography was employed to generate tracts, and mean fractional anisotropy (FA) measures were encoded in connectivity matrices. Network-based statistical analysis identified differences between TBI patients and controls. Principal component analysis and random forest classification were used for the final classification.	Results: The analysis revealed altered connectivity in several intra- and inter-hemispheric white matter pathways associated with diffuse axonal injury (DAI). The machine learning model achieved a mean classification accuracy of 68.16% ± 1.81% and a mean sensitivity of 80.0% ± 2.36% in correctly classifying TBI patients. Predictive Value: The study demonstrated that statistical machine learning approaches applied to structural connectomes could effectively identify patients with diffuse axonal injury, with a notable classification accuracy and sensitivity.	The findings highlight the potential of using machine learning and structural connectivity analysis to improve the identification and understanding of TBI, particularly in cases where traditional imaging methods may not reveal clear abnormalities.
Prediction of early TBI mortality using a Machine Learning approach in a LMIC population Amorin, et al. [12]	2019	Objective: To design and compare predictive models for early mortality in traumatic brain injury (TBI) patients in a low- and middle-income country (LMIC) population using machine learning techniques. Methods: A prospective registry was established in São Paulo, Brazil, including TBI patients admitted to the ICU. Predictors evaluated included gender, age, pupil reactivity, Glasgow Coma Scale (GCS), hypoxia, hypotension, CT findings, trauma severity score, and laboratory results. Various machine learning models were developed and compared.	The overall 14-day mortality was 22.8%. The Naive Bayes model achieved the highest prediction performance for overall mortality with an area under the curve (AUC) of 0.906. Significant predictors included GCS at admission, prehospital GCS, age, and pupil reaction. For predicting ICU length of stay, the Conditional Inference Tree model performed best with a root mean square error (RMSE) of 1.011. Predictive Value: The machine learning models demonstrated high predictive performance, particularly the Naive Bayes model for mortality prediction and the Conditional Inference Tree model for ICU length of stay.	Machine learning models can effectively predict early mortality and ICU length of stay in TBI patients in LMICs, potentially aiding in treatment decisions and family counseling.
Development of a Multimodal Machine Learning-Based Prognostication Model for Traumatic Brain Injury Using Clinical Data and Computed Tomography Scans: A CENTER-TBI and CINTER-TBI Study Hibi, et al. [8]	2020	Objective: To develop a multimodal machine learning-based prognostication model for traumatic brain injury (TBI) using clinical data and computed tomography (CT) scans, leveraging data from the CENTER-TBI and CINTER-TBI studies. Methods: This study was a retrospective analysis of the CENTER-TBI dataset (n = 1016). A machine learning-driven binary classifier was developed to predict 6-month post-injury outcomes. Comparison were made of the model’s performance with conventional models based on clinical variables and CT scoring systems with external validation using the CINTER-TBI dataset (n = 348)	The machine learning model achieved an AUC of 0.846 (95% CI: 0.843–0.849) in internal validation. External validation showed an AUC of 0.859 (95% CI: 0.857–0.862), outperforming the clinical variables model (AUC = 0.809, 95% CI: 0.798–0.820). Predictive Value: The machine learning model demonstrated superior prognostic performance compared to traditional models, indicating its potential for more accurate and reliable TBI outcome predictions without the need for manual CT assessments	The study established a machine learning-based model that enhances the efficiency and reliability of TBI prognosis using CT scans, with significant implications for earlier intervention and improved patient outcomes.
Machine Learning Enhanced Mechanistic Simulation Framework for Functional Deficit Prediction in TBI Schroder, et al. [11]	2021	Objectives: To develop a machine learning-enhanced mechanistic simulation framework for predicting functional deficits in traumatic brain injury (TBI) patients using resting state functional MRI (rsfMRI) data. Methods: The study utilized a numerical framework combining pre-calculated cases with a machine learning layer to predict resting state network disruption following head impacts. The framework was tested against rsfMRI data from nine TBI patients scanned within 24 h of injury.	The machine learning predictions closely matched full simulation results in a dummy fall case. The framework demonstrated potential for predicting rsfMRI alterations from paramedical data and reconstructing accidents through rsfMRI measurements Predictive Value: The framework showed promise in efficiently predicting functional deficits and reconstructing head impact scenarios, although further clinical data is required for full validation.	The approach opens the door to on-the-fly prediction of rsfMRI alterations and reverse-engineered accident reconstruction, potentially enhancing clinical assessment and intervention for TBI patients.
Abnormal structural and functional network topological properties associated with left prefrontal, parietal, and occipital cortices significantly predict childhood TBI-related attention deficits: A semi-supervised deep learning study Cao et al. [10]	2023	Objective: The study aimed to investigate the topological alterations in both structural and functional brain networks in children with traumatic brain injury (TBI) and their predictive power for post-TBI attention deficits using a semi-supervised deep learning model. Methods: The study utilized functional magnetic resonance imaging (fMRI) data during a sustained attention processing task and diffusion tensor imaging (DTI) data from 110 subjects (55 children with TBI and 55 group-matched controls). A total of 60 topological properties were selected as brain features for building the model.	The deep learning model differentiated children with TBI from controls with an average accuracy of 82.86%. The most important brain features for accurate classification were functional and structural nodal topological properties associated with the left frontal, inferior temporal, postcentral, and medial occipitotemporal regions. Predictive value: Post hoc regression-based machine learning analyses showed that neuroimaging features associated with the left postcentral area, superior frontal region, and medial occipitotemporal regions significantly predicted elevated inattentive and hyperactive/impulsive symptoms.	The findings suggest that deep learning techniques can help identify robust neurobiological markers for post-TBI attention deficits. The left superior frontal, postcentral, and medial occipitotemporal regions may serve as reliable targets for diagnosis and interventions of TBI-related attention problems in children.
Contribution of CT-Scan Analysis by Artificial Intelligence to the Clinical Care of TBI Patients Brossard et al. [4]	2021	Objective: The objective was to evaluate the contribution of artificial intelligence (AI) in analyzing CT scans for the clinical care of patients with traumatic brain injury (TBI) Methods: The methods involved a retrospective analysis of CT scans using AI algorithms to identify and quantify TBI-related abnormalities. The study focused on the automated detection of intracranial lesions, classification of hemorrhage types, and measurement of midline shift and hematoma volume.	Demonstrated that using machine learning on CT scans and clinical data enhances prognostication for traumatic brain injury patients. Compared to traditional methods relying on clinical variables or CT scoring systems, the ML model showed better predictive accuracy. The ML model operates without needing manual CT assessments. The results demonstrated that AI could accurately identify and quantify various TBI-related abnormalities. The AI algorithms showed high specificity and positive predictive value in classifying different types of hemorrhages and measuring midline shift and hematoma volume. Predictive value: The study reports significant predictive value of the AI models with the algorithms achieving high area under the curve (AUC) scores in predicting clinical outcomes based on CT data. This suggests that AI can enhance the prognostication of TBI patients, potentially leading to earlier interventions and improved patient outcomes	The study highlighted that AI-assisted analysis improved the reproducibility and accuracy of TBI assessments compared to manual evaluations. Using machine learning on CT scans can predict outcomes in traumatic brain injury patients better than traditional methods. This advancement could standardize prognostic assessments and improve patient care and outcomes in TBI management. In conclusion, the study by Brossard et al. supports the integration of AI in the clinical workflow for TBI management. AI-assisted CT scan analysis can provide valuable tools for clinicians, improving the accuracy and efficiency of TBI diagnosis and prognosis.
Prediction analysis of TBI 24-h survival outcome based on machine learning Y. Yang, et al. [13]	2024	Objective: The study aimed to predict the 24-hour survival outcome of patients with traumatic brain injury (TBI) using machine learning techniques. Methods: The analysis included 1224 samples with clinical indicators such as age, gender, blood pressure, and MGAP. The study utilized data visualization, single factor analysis, feature engineering, and various machine learning models including random forest (RF), K-Nearest Neighbors (KNN), logistic regression (LR), and deep neural network (DNN). The Synthetic Minority Over-sampling Technique (SMOTE) was used to address the imbalance in the dataset.	Results: The study found that while the accuracy of all models was high, the recall rate was initially low. After resampling, the recall rate of positive samples improved significantly. The optimal model was identified as the RF model, which achieved a recall rate of 0.67 and an AUC of 0.87. Predictive Value: The RF model demonstrated the best performance in predicting 24-hour survival outcomes, with an AUC of 0.87, indicating good accuracy.	The study concluded that machine learning models, particularly the RF model, can effectively predict 24-hour survival outcomes in TBI patients, with improved recall rates and AUC after resampling.

Study characteristics

This scoping review included seven studies that analyzed the application of AI in diagnosing and treating TBI. The study characteristics can be seen in Table 1. Four of the studies pointed out that integrating AI with imaging leads to improved TBI diagnosis and outcomes [4, 7–9]. The remaining three studies concentrated on the usage of AI to foresee the outcomes of TBI patients [10–12]. Each study used different AI techniques such as deep learning, machine learning (ML), and mechanical simulations to demonstrate AI’s potential in impacting TBI diagnostics and treatment planning. All seven studies involved the use of AI as a diagnostic tool in clinical practice and as an aid for decisions relating to prognosis and treatment of TBI.

Outcomes

Of the studies concerning AI and imaging, Brossard et al. explored a tool that categorized and segmented brain injuries to predict outcomes and identify new biomarkers within TBI patients. The article highlights that the incorporation of blood-based biomarkers such as glial fibrillary acidic protein (GFAP) and ubiquitin C-terminal hydrolase-L1 (UCH-L1) may aid to better understand the pathophysiology of acute TBI and improve the performance of clinical decision rules. This study indicated that the use of AI leads to a higher level of CT scan analysis accuracy thus making it possible to foresee the outcomes and classify injuries with a higher level of precision. However, research also noted the existence of several obstacles that need to be addressed, including the standardization of imaging and the procurement of large datasets for training the tool [4].

A similar study by Hibi et al. confirms these findings and challenges. It focuses on creating a multilayered forecast model based on machine learning, with the model combining the patient’s clinical information with data from CT scans to estimate a forecast for TBI patients [7]. Testing showed this model to be more precise in its predictions than traditional methods. The tool integrates the patient’s medical details and CT score systems, offering accurate and timely forecasts, but does not eliminate the need for manual CT exams. By pairing this AI CT analysis with relevant clinical data, said models enhance prediction accuracy, streamline the diagnostic process, and improve clinical decision making.^[1] This new advance highlights the potential of machine learning to standardize forecast evaluations and incorporate multiple forms of data, leading to better patient care and more predictable results and treatment plans in TBI management.

Mitra et al. highlighted the application of machine learning to the mapping of white matter connections in the patients’ cortical and subcortical regions. The artificial intelligence (AI) method was successful in showing connectivity pattern changes with a precision of 68.16%, thus offering early diagnosis and treatment of TBI patients [8]. However, the study did not claim that AI eliminates the need for manual CT exams. Cao et al. utilized a similar method of analysis using AI on pediatric functional magnetic resonance imaging (fMRI). The deep learning program was fed fMRIs of TBI versus non-TBI afflicted children and correctly diagnosed them with an accuracy of 82.86% [9]. It also noted function in the left frontal and postcentral areas as being a predicting factor for attention deficit secondary to TBI in children [9]. This capability of picking up on small brain connectivity alterations exemplifies how AI can significantly contribute to diagnosing TBI of different severities, going beyond the computational limitations of a physician or radiologist in definitively identifying patterns.

Other research investigated the potential of machine learning to forecast TBI outcomes from diverse sources. In one study by Schroder et al., the degree of brain damage following head injuries was predicted and simulated using machine learning. The model considered an object’s velocity, location, angle, and form before causing head trauma. The system consistently predicted outcomes and accurately identified disruptions in resting state networks, including the default mode network (DMN), in the initial imaging incidents [10].

Amorim et al. conducted a study in São Paulo on a machine learning application that estimated both the death rate, and the duration of hospital stay for TBI patients using data from the Glasgow Coma Scale (GCS) along with other clinical variables to inform the artificial intelligence (AI) system [11]. These models enhance clinical decision-making, thereby filling resource gaps and benefiting primarily low-resource areas.^[1] The study demonstrated high prediction performance, with the most significant predictors being the GCS at admission, age, and pupil reaction. This approach has the potential to improve treatment decisions and counsel family members in low- and middle-income countries (LMICs) [13].

A study by Yang et al. focused on enhancing pre-existing predictive models using the Synthetic Minority Over-sampling Technique (SMOTE) to predict 24-hour survival outcomes in traumatic brain injury (TBI) patients. The study utilized random forest, k-nearest neighbors, logistic regression, and deep neural networks. Logistic regression, a traditional statistical method rather than an artificial intelligence (AI) technique, was used in this study to predict patient survival, achieving the highest accuracy and outperforming other models after resampling. While logistic regression is a widely used tool in predictive modeling, it is important to note that it is not considered a form of AI. AI methods, such as deep learning and machine learning algorithms, represent a more advanced approach to prediction and modeling [12].

Performance metrics

CRASH and IMPACT are traditional prognostic tools that have been extensively validated for outcome prediction in traumatic brain injury (TBI), typically achieving area under the curve (AUC) values near 0.80 for mortality. The CRASH model was developed using data from a large international cohort and focuses on predicting 14-day mortality and 6-month unfavorable outcomes. It incorporates variables such as age, Glasgow Coma Scale (GCS) score, pupil reactivity, and the presence of major extracranial injury. The model has been validated in various settings and has shown good discrimination and calibration, although it tends to overpredict mortality and unfavorable outcomes in some contemporary cohorts.

The IMPACT model includes three versions: core, extended, and laboratory. The core model uses age, motor score, and pupillary reactivity; the extended model adds CT findings and secondary insults; and the laboratory model includes additional laboratory parameters. The IMPACT model predicts 6-month mortality and unfavorable outcomes. It has been validated extensively and generally shows good discrimination and calibration, though it also tends to overpredict mortality in some settings. Table 2 is a summary of the articles in comparison to the traditional models of CRASH and IMPACT. Performance metrics from the studies reviewed were summarized where available. The key metrics reported included accuracy, sensitivity, specificity, and the area under the receiver operating characteristic curve (AUC). Accuracy is defined as the proportion of true results (both true positives and true negatives) out of the total number of cases examined. It reflects how often the model correctly identifies or excludes a condition. AUC, derived from the receiver operating characteristic (ROC) curve, measures the model’s ability to distinguish between different classes, such as the presence or absence of a condition. The AUC value ranges from 0 to 1, with higher values indicating superior discriminatory performance. An AUC of 0.5 indicates no ability to distinguish between classes (similar to random chance), while an AUC of 1.0 indicates perfect discrimination. Overall, the performance of the models varied, with some showing promising predictive ability, dependent on the algorithm employed and the complexity of the dataset.

Table 2.

Performance metrics summary

Study	Algorithm	Accuracy (%) & AUC	Comparison to CRASH or IMPACT
Statistical machine learning to identify traumatic brain injury (TBI) from structural disconnections of white matter networks. Mitra et al. (2016) [9]	ML Classification The algorithm involved mapping white matter connections using higher-order diffusion models and probabilistic tractography to generate connectivity matrices based on mean fractional anisotropy (FA) measures. These matrices underwent network-based statistical analysis to identify differences between TBI patients and controls, followed by principal component analysis and classification using a random forest model.	The study primarily focused on the classification accuracy and sensitivity of the machine learning model used to identify TBI patients based on structural connectivity patterns. The mean classification accuracy achieved was 68.16% ± 1.81%. Sensitivity: 80.0% ± 2.36% Specificity and AUC: Not explicitly reported in the article	The machine learning approach by Mitra et al. shows promise, particularly in sensitivity, but lacks the comprehensive validation and higher accuracy metrics demonstrated by the IMPACT and CRASH models.
Prediction of early TBI mortality using a Machine Learning approach in a LMIC population Amorim et al. (2019) [12]	Naive Bayes, Decision Trees This algorithm is a probabilistic classifier based on Bayes’ theorem with the assumption of independence between predictors.	Accuracy, sensitivity and specificity are not explicitly reported. However, the Naive Bayes model achieved the highest prediction performance for overall mortality with an AUC of 0.906	This model demonstrated a high predictive performance with an AUC of 0.906, outperforming traditional models like the CRASH and IMPACT models, which have AUCs of 0.876 and 0.821, respectively, in similar settings.1, 2This suggests that machine learning models, particularly those tailored to local data, can provide more accurate predictions than older technologies, potentially improving clinical decision-making in LMICs.
Development of a Multimodal Machine Learning-Based Prognostication Model for Traumatic Brain Injury Using Clinical Data and Computed Tomography Scans: A CENTER-TBI and CINTER-TBI Study Hibi et al. (2020) [12]	ML on CT + Clinical The algorithm used was a binary classifier based on machine learning techniques that integrated clinical data and CT scans to predict 6-month post-injury outcomes, demonstrating superior performance compared to traditional models based on clinical variables and established CT scoring systems.	Higher accuracy than traditional The machine learning model achieved an AUC of 0.846 (95% CI: 0.843–0.849) in internal validation. This performance was superior to models based on clinical variables (AUC = 0.817, 95% CI: 0.814–0.820) and established CT scoring systems (Marshall AUC = 0.829, 95% CI: 0.826–0.832; IMPACT AUC = 0.838, 95% CI: 0.835–0.841). External validation showed an AUC of 0.859 (95% CI: 0.857–0.862), outperforming the clinical variables model (AUC = 0.809, 95% CI: 0.798–0.820)	The machine learning-based prognostication model developed in the study demonstrated superior performance with an AUC of 0.846 in internal validation and 0.859 in external validation, compared to the IMPACT model (AUC = 0.838) and the CRASH model (AUC not specified in the provided references). This indicates that the new model offers more accurate and reliable predictions for TBI outcomes, outperforming traditional models that rely on clinical variables and manual CT assessments.
Machine Learning Enhanced Mechanistic Simulation Framework for Functional Deficit Prediction in TBI Schroder et al. (2021) [11]	ML Simulation The algorithm combines a library of pre-calculated head impact cases with a machine learning layer to efficiently predict resting state network disruption based on head velocity, location, angle of impact, and impactor shape.	Consistently predictive The research article does not report specific quantitative metrics such as the AUC, sensitivity, or specificity for their predictive model. The study primarily describes the development of a machine learning-enhanced framework and its initial validation against rsfMRI data from TBI patients. The framework’s predictions were shown to closely match full simulation results in a test case, indicating potential for accurate prediction of functional deficits, but further clinical validation is needed to establish detailed performance metrics.	This model offers a more personalized and precise approach by integrating mechanistic simulations with machine learning to predict resting state network disruptions based on specific head impact parameters. The traditional models of CRASH and IMPACT primarily rely on clinical and demographic data, and while they have been validated and show good discrimination and calibration, they do not incorporate the detailed mechanistic and imaging data used in the newer framework.
Abnormal structural and functional network topological properties associated with left prefrontal, parietal, and occipital cortices significantly predict childhood TBI-related attention deficits: A semi-supervised deep learning study Cao et al. (2023) [10]	Semi-supervised Deep Learning A semi-supervised deep learning study” utilized a semi-supervised autoencoder algorithm. This deep learning model was designed to investigate topological alterations in both structural and functional brain networks in children with TBI and their predictive power for post-TBI attention deficits.	The article reports that the semi-supervised autoencoder model achieved an average accuracy of 82.86% in differentiating children with TBI from controls. Additionally, the study identified that topological properties associated with the left frontal, inferior temporal, postcentral, and medial occipitotemporal regions were the most significant features for accurate classification. The AUC of 0.89 is 0.89 which indicates a high level of accuracy in distinguishing children with TBI from controls based on the identified neuroimaging features.	The semi-supervised deep learning model developed by Cao et al. for predicting childhood traumatic brain injury (TBI)-related attention deficits demonstrated an average accuracy of 82.86% and an AUC of 0.89. In comparison, the IMPACT and CRASH models, which are well-established prognostic tools for TBI, have shown AUCs ranging from 0.65 to 0.92 in various validation studies, indicating that while the deep learning model shows promise, it is essential to consider the context and specific application of each model.
Contribution of CT-Scan Analysis by Artificial Intelligence to the Clinical Care of TBI Patients Brossard et al. (2021) [4]	Deep Learning/ML: The algorithm leverages deep learning techniques, particularly convolutional neural network (CNNs) with transfer learning, to improve the segmentation and prediction capabilities for TBI management based on CT scan data.	High accuracy but not explicitly stated The AI algorithms achieve an AUC of 0.89 ± 0.17 for predicting the therapeutic intensity level (TIL) when using manual quantification of volumes of 7 lesions and their spatial locations. Automatic 4-class segmentation without transfer learning had an AUC of 0.60 ± 0.23. Transfer learning significantly improved the automatic 4-class segmentation, achieving an AUC of 0.83.	The AI algorithm demonstrates superior performance when compared to older technologies like CRASH and IMPACT with an AUC of 0.89 ± 0.17 for predicting the therapeutic intensity level (TIL), which is higher than the AUCs reported, which can range from 0.66–1.00.66.00 and 0.65 to 0.90 respectively in various trials.
Prediction analysis of TBI 24-h survival outcome based on machine learning Yang et al. (2024) [13]	Logistic Regression LR model incorporated several clinical predictors, including age, gender, blood pressure, and the MGAP score (Mechanism, Glasgow Coma Scale, Age, and Arterial Pressure). To address the imbalance in the dataset, the Synthetic Minority Over-sampling Technique (SMOTE) was used for oversampling the training set.	Highest accuracy among models with AUC of 0.87 and a sensitivity recall rate of 0.67, indicating good accuracy and improved recall rates after resampling.	The machine learning models used in the study by Y. Yang et al. for predicting 24-hour survival outcomes in traumatic brain injury (TBI) patients demonstrated an AUC of 0.87 with the Random Forest (RF) model, which is comparable to or slightly better than the performance of traditional models like the IMPACT and CRASH models. For instance, the CRASH model has shown AUCs ranging from 0.82 to 0.88, and the IMPACT model has shown AUCs ranging from 0.77 to 0.85 in various studies, indicating that machine learning models can offer similar or improved predictive accuracy.

Discussion

The purpose of this scoping review was to assess AI’s use in the different types of TBI diagnosis, prognosis, and treatment. Given the relative novelty of AI in healthcare, an aim was also to survey in what ways AI was being implemented into TBI care, with what level of efficacy, and how it might be used in the future. Four studies identified AI as a powerful tool in the analysis of CT, MRI, and FMRI imaging [4, 7–9]. The other three studies focused on AI’s capacity as a tool for modeling and simulating outcomes [10–12].

AI’s usage with imaging diagnostics for TBI has been a significant and critical development,

improving the accuracy and speed of scan analysis which in turn provides timely care and relieves physician workloads [4]. Despite the GCS being an important method for grading a TBI, current research suggests the first priority is considering neuroimaging given that GCS scores fail at directly relating to the risk of CT abnormalities [13]. AI is able to address this new demand twofold: it can analyze the increased number of scans that researchers suggest are needed and would otherwise fall to the physician alone, and it can use GCS scores for patient outcome prediction despite the scale’s lack of success in being a detailed triage tool [4, 9–11, 14]. Similarly, Schroder et al. illustrated the substantial gains in the accuracy of diagnosing by developing hybrid ML models that use both clinical data and MRI metadata [10]. In both cases, AI based models are shown to be holistic tools that can provide both diagnostic and prognostic evaluations using diverse clinical datasets and evidence.

Comparison to traditional predictive models

As mentioned previously, traditional prognostic models, such as IMPACT and CRASH, have been widely validated for predicting outcomes in TBI, showing AUCs around 0.80 for mortality prediction. In contrast, the AI models reviewed in this study demonstrate accuracies ranging from 68% to over 80%, with some models exhibiting comparable or even improved predictive performance compared to IMPACT and CRASH. A comparison is shown in Table 3. For example, models using deep learning or machine learning techniques have shown promising results in mortality prediction, but many lack external validation across diverse populations, which limits their clinical applicability. Future studies should focus on direct comparisons between AI models and traditional models like IMPACT and CRASH to better understand the added value of AI in clinical practice. Benchmarking AI algorithms against these well-established tools will provide clearer insights into their potential and limitations.

Table 3.

Comparison of IMPACT, CRASH, and AI models for TBI prognosis *AUC (Area under the Curve): A measure of predictive accuracy. Values close to 1 indicate better accuracy, while values closer to 0.5 suggest performance equivalent to random guessing

Model	AUC*(Mortality Prediction)	Key Features	Strengths	Limitations
IMPACT	~ 0.80	Uses clinical and CT data for mortality prediction	Widely validated in diverse populations	Limited to clinical and imaging data; no AI techniques
CRASH	~ 0.80	Focuses on predicting mortality, long-term outcomes	Effective across various severity levels	Lacks real-time predictive power; limited by traditional methods
AI Models	~ 0.80–0.96 Variable depending on model used	Uses deep learning/machine learning algorithms on multimodal data	Can process large datasets and detect subtle abnormalities; higher accuracy in some cases	Lack of external validation; data variability; interpretability issues

Potential benefits of AI in TBI management

Physician uncertainty around the efficacy of treatment around TBI treatment is unique and a result of little research available to delineate higher risk cases. One study showed that 5% of mTBI patients return to the ED within 72 h, and of those some were critical. The current protocol for addressing this is advising physicians to identify at risk patients preemptively and to offer anticipatory guidance, however these are both still difficult given a lack of research and effective means of production [15]. AI modeling has the potential to reduce hospital readmissions for mTBI by appropriately delineating higher risk patients through the use of predictive models. While datasets still need to be improved, AI may have the ability to create models that have more predictive power than the current methods, using the same data. This streamlining of care allows for a lower burden on providers as well as more effective and continuous care.

Future direction and implications

AI has potential to assist in the diagnosis of traumatic brain injury (TBI) by improving accuracy, speed, and accessibility. In Medical Imaging Analysis, AI powered tools can analyze CT scans, MRIs, and PET scans to detect subtle signs of TBI, such as microbleeds, swelling, or structural damage. Machine learning algorithms can help radiologists identify abnormalities faster and more accurately, reducing human error. As seen in the studies summarized in this review, deep learning models are being developed to classify TBI severity based on imaging data. AI can process large datasets, including patient history, symptoms, and biomarkers (such as blood based or cerebrospinal fluid markers), to predict TBI outcomes and tailor treatments. AI driven pattern recognition can also help identify early indicators of long-term complications, such as chronic traumatic encephalopathy or post-concussive syndrome [16]. A third area where AI can be powerful is in the analysis of electronic health records to extract relevant information from clinical notes, which has the potential to improve diagnostic accuracy.

Limitations

There are several limitations to the current study. First, the literature search was limited to PubMed, potentially missing important studies. The narrow search scope and small number of eligible studies potentially pose key constraints on generalizability. Thus, these data and evidence should be regarded as preliminary.

While artificial intelligence offers significant promise in enhancing the diagnosis and prognosis of traumatic brain injury, several limitations must be addressed before its full clinical adoption. AI models require large, high-quality datasets for development, but variability in imaging protocols, clinical documentation, and patient populations across institutions can limit their generalizability. Additionally, bias in training data—such as the underrepresentation of certain age groups, racial populations, or injury severities—can result in AI models that perform poorly on non-representative groups. Furthermore, model interpretability remains a challenge; many AI systems operate as ‘black boxes’, which makes it difficult for clinicians to understand the rationale behind predictions, thereby limiting their trust in these systems. Multimodal imaging applications, involving a combination of CT, MRI, and fMRI data, present additional difficulties. High-dimensional data increases the risk of overfitting when training datasets are small or heterogeneous, and handling such data without losing critical information remains a technical challenge. Finally, ethical issues, such as patient privacy, data security, and equitable access to AI technologies, must be carefully addressed to ensure that AI-driven tools are used responsibly and fairly across all patient populations.

Addressing these limitations will require ongoing research focused on developing interpretable, externally validated, and generalizable models, ideally through multicenter collaborations that reflect diverse patient populations and clinical environments.

Conclusion

Artificial intelligence has the potential to enhance the diagnosis and prognosis of traumatic brain injury by improving the speed, accuracy, and precision of clinical assessments. Through advanced imaging analysis and predictive modeling, AI tools can assist clinicians in differentiating injury types, identifying subtle brain abnormalities, and uncovering new biomarkers, ultimately enabling earlier and more targeted interventions. Machine learning models have also demonstrated capabilities in forecasting clinical outcomes such as mortality and hospital length of stay, providing support for clinical decision-making and resource allocation.

Despite these advancements, challenges remain before AI can be fully integrated into routine clinical practice. Variability in imaging techniques, data quality, and patient populations, along with the need for large, standardized datasets, present barriers to generalizability. Ethical considerations, including patient privacy, algorithmic transparency, and equitable access, must also be addressed to ensure responsible use. Ongoing research, collaborative validation efforts, and engagement among researchers, clinicians, and policymakers will be essential to realize the potential of AI in improving TBI management and patient outcomes.

Authors’ contributions

The study was designed by CM. CM, MG, and SN managed data capture and performed the analysis. SN supervised the conduct of the research. CM and SN drafted the initial manuscript, MG and SN reviewed and edited the initial manuscript, and all authors approved the final version of the manuscript.

Funding

No funding was received for this study.

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethics approval and consent to participate

The study was performed in accordance with the ethical standards as laid down in the 1964 Declaration of Helsinki and its later amendments or comparable ethical standards.

Consent for publication

N/A.

Competing interests

The authors declare no competing interests.