Artificial intelligence in traumatic brain injury: Brain imaging analysis and outcome prediction: A mini review

Abstract

Integration of artificial intelligence increases in all aspects of human life, particularly in healthcare systems. Traumatic brain injury is a significant cause of mortality and long-term disability, with an important impact on the socio-economic system of healthcare. The role of artificial intelligence in imaging and outcome prediction for traumatic brain injury patients is reviewed with a particular emphasis to the characteristics of machine and deep learning methods. Evidence of potential improvement in the clinical practice in discussed.

Key Words: Artificial intelligence; Traumatic brain injury; Outcome prediction; Brain imaging

Core Tip: Several reviews in the literature contributed to the role of Artificial Intelligence (AI) in medicine. However, this mini-review focuses of the potential role of AI in imaging analysis and outcome prediction of traumatic brain injury patients. Evidence of the latest results obtained with different AI-based technologies, in particular Machine learning and Deep learning algorithms, are provided.

INTRODUCTION

Traumatic brain injury (TBI) is a significant cause of mortality with a prevalence between 1/3 and 1/2 of all injury-related deaths[1]. Moreover, TBI accounts as the leading cause of disability among individuals under the age of 40[2]. In low- and middle-income countries, the high incidence, the significant mortality rates and long-term disabilities represents a significant burden, in particular for socio-economic aspects and for the consequences on life expectancy and quality of life[3].

Therapeutic interventions for TBI are strictly related to a reliable patient’s diagnosis, in particular to establish if a neurosurgical approach is advisable and potentially successful[4]. The assessment of TBI severity is mainly based on early brain images obtained by computed tomography (CT) and scores based on the radiological reading finding as, i.e. the Marshall, Rotterdam, Stockholm or Helsinki scores[5-8]. Nevertheless, significant gaps and challenges are still apparent in TBI care. In particular, precise diagnosis and prognosis are affected by a marked heterogeneity in signs and symptoms that can additionally be confounded by pre-existent neurological or mental health conditions. Moreover, further factors that can impact the clinical presentation include patient age, comorbidity, current therapy and/or the consumption of illegal or toxic substances[9]. The TBI severity classification is also characterized by a significant heterogeneity in types, location and extension of brain lesions[10]. Primary injuries occur at the same time of the traumatic events and include epidural or subdural hematomas, microvascular, cortical and axonal damage[11]. Secondary injuries occur after hours/days since the trauma and brain damage presents as cerebral edema and increased intracranial pressure as the end effect of complex biochemical events[12].

Artificial intelligence (AI) is facing an increasingly important role in the field of radiological images analysis, particularly in the diagnosis and outcome prediction of TBI[4,13-15]. The integration of AI technologies- in particular machine learning based algorithms- into radiological practices is potentially of high impact to improve accuracy, efficiency, and prognostication of TBI patients[16,17]. Among the key aspects of AI in radiological field, image analysis and traumatic injuries detection, quantitative assessment of damage, prediction of early complications and functional outcomes, integration with clinical data are the most relevant[18].

As technology continues to evolve, collaborative efforts between clinicians, radiologists, and AI developers will be crucial for realizing the full benefits of these advancements in managing TBI patients.

A literature search of PubMed, Goggle and Cochrane was performed using the following search terms: “Artificial intelligence”, “neural network”, “deep learning”, “machine learning”, “MRI OR magnetic resonance”, “CT OR computed tomography”, “brain damage”, “brain trauma”, “traumatic brain injury”. The group of selected key words was extended by screening the references of the included studies to find possible synonyms.

Aim of this mini review is to report latest evidence related to the role of AI in the brain imaging analysis and outcome prediction in TBI patients. The most frequently AI algorithms adopted to analyze CT or magnetic resonance (MR) images are reviewed with a particular emphasis to the performance achieved and to the present limits. Rehabilitation strategies and their integration with AI technologies are not included in the present review.

ARTIFICIAL INTELLIGENCE METHODS IN TBI PATIENTS

The AI methods encompass different technologies aimed at developing smart algorithms able to self-modifying without external interventions. Deep learning algorithms correspond to a class of artificial neural networks that improve their performance by repeating an assigned task[17]. The software architecture is based on layers resembling the neural structure of the human brain with nodes that correspond to neurons with multiple dendrites as input and an axon as the output[19].

The AI algorithms performance can be established by statistical performance metrics that include accuracy, sensitivity, specificity. The algorithm’s accuracy represents the percentage of good predictions as the ratio between the sum of true positive and negative predictions (TP, TN) over the sum of all the predictions- true and false positive and negative (TP+TN+FP+FN). Accordingly, sensitivity represents TP/(TP+FN) and specificity is TN/(TN+FP). Furthermore, the receiver-operator curve and the measure of the corresponding the area under the curve (AUC) represents the combined performances of the classification algorithm and is a metric widely used to evaluate classification algorithms[4]. The performance of the model is usually compared to that of one or more expert radiologist readings[20,21].

The machine learning (ML) algorithms are focused on the extraction of useful information from data without being explicitly programmed and are based on procedures that automatically learn and improve the knowledge on a particular topic from previous trained experience[22]. The ultimate goal of ML is to identify patterns, postulate predictions, or propose actions[17].

Several ML algorithms have been used for images segmentation in TBI[14,18]. The most commonly adopted include the K-Nearest Neighbors, the support vector machines (SVM), the random forest (RF), the K-Means clustering algorithm and the Gaussian Mixture Model[17,18], Table 1.

Table 1 Characteristics of the main machine learning algorithms.

Traditional machine learning algorithms	Main characteristics
K-nearest neighbors	Pixels are based on their similarity to neighbors. Adopted in early TBI studies for lesion segmentation
Support vector machines	Regions of interest are distinguished using hyperplanes. Adopted for classifying brain tissue types in TBI
Random forest	Learning method that uses decision trees for pixel classification. Adopted in multimodal MRI segmentation
K-means clustering	Unsupervised algorithm that groups pixels based on intensity values. Adopted for quick but coarse lesion detection
Gaussian mixture model	Probabilistic model that assumes pixel intensities follow a Gaussian distribution. Adopted in segmentation tasks

TBI: Traumatic brain injury; MRI: Magnetic resonance image.

Deep learning is a subfield of ML developed to obtain high level of abstractions through learning algorithms designed on hierarchical representations of data that has significantly advanced TBI image segmentation due to its ability to obtain high-level representations[19,23]. The most commonly adopted include the convolutional neural networks, the fully convolutional networks, the recurrent neural networks and long short-term memory[24], Table 2.

Table 2 Characteristics of the main deep learning algorithms.

Deep Learning algorithms	Main characteristics
CNNs	Automatic learning of hierarchical features from brain images. Widely adopted to TBI segmentation
Fully convolutional networks	Replacement of fully connected layers with convolutional layers, enabling pixel-wise classification. Adopted for segmenting lesions and hemorrhages in TBI MRI scans
Recurrent neural networks & long short-term memory	Adopted in time-series analysis for dynamic brain imaging data. LSTMs combined with CNNs useful in sequential TBI segmentation across multiple time poi

TBI: Traumatic brain injury; MRI: Magnetic resonance image; CNNs: Convolutional neural networks; LSTMs: Long short-term memory.

As an example, a study applied CNN algorithms to detect the presence or the absence of one type of lesion in a scan with an AUC from 0.92 to 0.97 when externally validated[25].

Both ML and deep learning-based algorithm are characterized by a strong dependence on the data set used in the training step of the method.

ARTIFICIAL INTELLIGENCE-BASED ANALYSIS OF BRAIN IMAGING IN TBI PATIENTS

The AI algorithms can be used to analyze radiological imaging, such as CT and MR scans and to achieve early and accurate detection of TBI including contusions, hemorrhages, and skull fractures. CT scan can identify intracranial hemorrhage, shifting of midline structures, and skull fractures. In particular, intracranial hemorrhage can be divided into 5 sub-types: Intraventricular, intraparenchymal, subarachnoid, epidural and subdural hemorrhage[26]. The intraparenchymal, epidural and subarachnoid cases can potentially be surgically treated on the basis of the volume and the thickness of the hemorrhage and so a fast and correct sizing of the lesion is mandatory[27].

Several studies proved that several important TBI-related CT findings that can automatically be identified and quantified with AI methods[28]. The results proved that deep learning network algorithms accurately detect abnormal patterns associated with TBI including intracranial hemorrhage (type and volume) and midline shift[29].

The use of MR imaging as the first approach in acute TBI patients is limited due to significant costs, availability, potential contraindication, potential scarce patient compliance or motion artifacts[30]. On the other side, MR can be crucial to evaluate the possible diffuse axonal injury and small cortical or extra axial bleedings in patients with mild TBI (mTBI)[31]. Noticeably, the literature reports few studies on the performance of AI algorithms to detect and classify mild damage in TBI patients on MR images. Recent studies evaluated the diffuse axonal injury through the detection of cerebral microbleeds adopting a deep learning approach- CNN- on MR scans[32,33]. The detection of microbleeds by a two-step method reached a sensitivity between 78% and 93%, according to the training set adopted[33]. Analogously, further case-control studies evidenced how different models of CNN (contrasted and segmented), on MR scans are able to detect microbleed with high sensitivity and with low false positive events[34]. The SVM algorithm was applied to evaluate the AI performance to discriminate mTBI from healthy controls. The analysis, based on 7 MRI parameters, led to an AUC of 0.78 with an accuracy rate of 81.1%, a specificity of 75% and sensitivity 88%[33].

ARTIFICIAL INTELLIGENCE-BASED OUTCOME PREDICTION IN TBI PATIENTS

The AI models can utilize radiological data, along with clinical and demographic information, to predict early complications and functional endings following TBI and can assist healthcare providers in estimating the likelihood of various outcomes, enabling better communication with patients and their families[13]. The AI can additionally contribute to the early detection of complications and secondary associated injuries, such as hydrocephalus or delayed hemorrhage that, when early identified, allows for prompt intervention and can improve patient outcomes[14].

Several studies provided evidence that ML-based methods can be used to effectively predict in-hospital mortality or unfavorable outcome of TBI patients and the algorithms most testes are SVM, CNN, RF[35,36].

Even if the studies present a significant heterogeneity in the selected input variables used to prediction the mortality and the unfavorable outcomes, the critical clinicopathological variables include abnormal serum glucose values, lactic acidosis, older age, lower Glasgow scale score (GCS) at admission[13]. A recent study reported an accuracy of ML algorithm to predict in-hospital mortality over 90%, with the higher values of obtained with CNN method (95.3%)[13]. Analogously, the potential utility of the tested AI algorithms is confirmed by the high sensitivity, specificity, AUC, positive and negative predictive values obtained[37]. The performance of ML method to predict unfavorable outcomes at 6 months was evaluated with several variables as vital signs, pupil reactivity, abbreviated injury severity, initial CT and/or MR scans, initial Marshall score and GCS, final extended Glasgow outcome scale. The AUC values obtained were consistently above 0.8[37].

DISCUSSION AND CONCLUSION

The present mini review evidences the potential role of AI in brain imaging analysis and outcome prediction in TBI patients. The reviewed studies proved the significant impact of AI algorithms on the identifications and quantification of CT and MR findings. The analysis of CT scans by deep learning methods proved to be effective in the damage detection and classifications, with a quantitative assessment of the brain lesion. The AI methods proved to also a powerful tool to analyze and characterize MR scans to assess diffuse axonal injury and micro-bleedings. The AI proved also to be highly promising to improve the outcome prediction of TBI patients. Both ML and deep learning algorithms evidenced high statistical performance, measure as accuracy, sensitivity, specificity and AUC values, to predict in-hospital mortality and unfavorable outcomes combing radiological, clinical and demographic information.

The possible role AI in different fields of medicine is emerging as crucial, as witnessed by the huge increase of dedicated studies. As a further example is represented by the application of AI in the managements of hearth rhythm disorder and cardiac arrest. In a scoping review, the potential role of AI to predict both in- and out-of-hospital heart arrest is depicted as a promising tool to improve the survival of patients[38].

A limit of the adoption of AI in the routinely clinical practice is related to the heterogeneity of available methods that makes the scenario more complex and the choice of the most appropriate tool more difficult. The data reported in literature are encouraging in terms of statistical measures, but future studies are necessary to identify the most relevant variables to be considered in the frame of the AI technology.

Moreover, heterogeneity in the input variables can significantly affects the algorithm performance[39]. Heterogeneity in input data can be related to different data sources, differences in population demographics or data format. Negative effects on AI performance include failing to generalize if training data set does not reflect the real clinical scenario (domain shift), bias on dominant subsets if unbalanced inputs are used (model bias), need of preprocessing or model tuning (increased complexity).

Clinical examination is still the hallmark of the acute TBI assessment, but integration of AI-based analyses with clinical data could enhance the overall understanding of a patient's conditions and lead to more comprehensive and personalized treatment plans. The utility of AI in the area of TBI assessment is particularly promising considering the prominent and decisive role of the image analysis in this context. As the comprehension and inclusion of ML methods continue to evolve, it becomes increasingly apparent a possible greatly enhance of the patient care by improving the accuracy and efficiency of diagnosis and treatment strategies. Automated imaging analysis on CT or MR scans of the brain could prove a real-time triage in emergency setting. In particular the short timescale (seconds) required to analyze images could improve the priority assessment and speed up the decision-making process[40,41].

However, integrating AI into the daily management of TBI presents, at the moment, some difficulties. In fact, ethical considerations, data privacy concerns, physician’s confidence remain significant barriers that require a careful resolution[42]. Clinical complexity, high risks decisions, patient conditions make ethical considerations critical. First issue is patient safety and AI methods can influence life-changing decisions (surgery, withdrawal of care, admission to neuro-intensive care unit). Heath inequalities can be amplified in underrepresented groups as, i.e., in pediatric patients if the model is trained mostly on adult subjects[43].

Among the different limitations of including AI in routine clinical practice, besides the heterogeneity of methods, partial data sets can introduce a bias on some demographics or injury types reducing the general applicability. In fact, AI algorithm trained on biased datasets could perform poorly on diverse populations, leading to misdiagnosis. The International Mission for Prognosis and Analysis of Clinical Trials in TBI (IMPACT) represents an important initiative that, in future studies, could be integrated with AI technology to overcome this limitation. The IMPACT project critically surveys TBI trials and investigates the application of conventional and innovative methods for design and analysis of trials in TBI, in particular to develop data sets from completed randomized controlled trials and observational studies[44,45].

CONCLUSION

Future studies should standardize datasets to improve model robustness, aim at reaching a multimodal data integration that include CT/MR scans, EEG output, laboratory and clinical values, develop a tailored prognosis and therapeutic plan. Finally, multicenter trials are advisable to validate AI tools before clinical adoption to improve generalizability, reduce bias, increase clinical trustworthiness.