The detection of mild traumatic brain injury in paediatrics using artificial neural networks

Highlights

• This study is the first to apply shallow and deep ANN for mTBI diagnosis in paediatrics using non-imaging data.

• This study is the first to apply shallow and deep ANN for mTBI diagnosis in paediatrics using non-imaging data.

• Deep ANN effectively diagnoses mTBI compared to the hybrid model.

Abstract

Head computed tomography (CT) is the gold standard in emergency departments (EDs) to evaluate mild traumatic brain injury (mTBI) patients, especially for paediatrics. Data-driven models for successfully classifying head CT scans that have mTBI will be valuable in terms of timeliness and cost-effectiveness for TBI diagnosis. This study applied two different machine learning (ML) models to diagnose mTBI in a paediatric population collected as part of the paediatric emergency care applied research network (PECARN) study between 2004 and 2006. The models were conducted using 15,271 patients under the age of 18 years with mTBI and had a head CT report. In the conventional model, random forest (RF) ranked the features to reduce data dimensionality and the top ranked features were used to train a shallow artificial neural network (ANN) model. In the second model, a deep ANN applied to classify positive and negative mTBI patients using the entirety of the features available. The dataset was divided into two subsets: 80% for training and 20% for testing using five-fold cross-validation. Accuracy, sensitivity, precision, and specificity were calculated by comparing the model's prediction outcome to the actual diagnosis for each patient. RF ranked ten clinical demographic features and twelve CT-findings; the hybrid RF-ANN model achieved an average specificity of 99.96%, sensitivity of 95.98%, precision of 99.25%, and accuracy of 99.74% in identifying positive mTBI from negative mTBI subjects. The deep ANN proved its ability to carry out the task efficiently with an average specificity of 99.9%, sensitivity of 99.2%, precision of 99.9%, and accuracy of 99.9%. The performance of the two proposed models demonstrated the feasibility of using ANN to diagnose mTBI in a paediatric population. This is the first study to investigate deep ANN in a paediatric cohort with mTBI using clinical and non-imaging data and diagnose mTBI with balanced sensitivity and specificity using shallow and deep ML models. This method, if validated, would have the potential to reduce the burden of TBI evaluation in EDs and aide clinicians in the decision-making process.

Introduction

Traumatic brain injury (TBI) is one of the major concerns in the world with a high percentile of morbidity and mortality in children [1]. The Glasgow coma scale (GCS) is the standard scoring system that is used to assess the level of consciousness and categorise TBI patients into mild (13–15), moderate (9–12), or severe (<9) according to injury severity [2]. Mild TBI (mTBI) forms at least 76% of all TBI cases in the United States (US). Moreover, paediatric TBI emergency department (ED) visits increased by 34.1% between 2006 and 2013 in the United States (US) [3]. Thus, the ED has witnessed an evolutionary increase in the number of TBI paediatric evaluations. Computed tomography (CT) is a vital diagnostic tool for TBI; around 3.9 million patients have head CT scans in the US EDs every year [4], of which approximately 644,679 are paediatric [5]. Among the head CT scans done annually, only 9% have positive TBI findings and 91% are normal. However, classifying CTs as positive or negative for TBI is costly, utilizes the ED resources negatively, and consumes clinician time to identify the presence of TBI on the CT scan. Thus, there is a need to develop effective ways that better identify positive mTBI on CT scans.

Numerous studies have implemented clinical rules to predict paediatric TBI outcome [[6], [7], [8]]. The paediatric emergency care applied research network (PECARN) rule is a well-validated clinical rule to recognize children at low risk of clinically important TBI following head trauma and identify the necessity of having a CT [7,9]. PECARN is a paediatric emergency medicine research network in the US, which conducts high-priority, multi-institutional research to prevent and manage acute diseases in paediatrics and young adults [10]. A number of studies have developed predictive metrics using the PECARN dataset for the early identification of recurrent seizures [11], headache [12], and paediatric GCS performance for preverbal children [13] in paediatric TBI patients. However, none have used machine learning (ML) models to predict mTBI using non-imaging data.

The application of ML models in the medical field has witnessed an exponential surge as a tool for clinical diagnosis and neurosurgery prognostication [14]. Machine learning is more accurate and efficient in processing voluminous data [[15], [16], [17]], and helps in analysing medical data to monitor patients, diagnose, prognose, and prevent diseases [18]. Artificial neural network (ANN) is reported as a powerful ML model that outperformed other models in predicting TBI outcome between 2016 and 2018 [14] and is superior to the conventional statistics because of its discriminant analysis and strong pattern recognition abilities [19].

Diagnostic and prognostic models based on ML have been booming as supportive tools in the medical field over the years [18,20]. Many studies have been conducted to predict TBI outcome using ML and have utilized various data types [[21], [22], [23], [24], [25]]. A hybrid ML model comprising decision tree (DT) and ANN was applied to predict the attributes crucial for the prediction of TBI in patients. The hybrid model outperformed the DT model alone [24]. ML has also been applied to delineate diffuse axonal injury (DAI) from structural connectivity patterns in mTBI patients on MRI and DTI images [23]. Four ML models were applied to predict mortality in 2059 isolated moderate and severe TBI patients using clinical data, and the ANN model had superior performance compared to DT, LR, support vector machine (SVM), and naive Bayes (NB) [25]. In another study, three ML models were compared to predict seizures from 49 moderate to severe TBI patients using resting-state functional MRI (fMRI) images; RF outperformed ANN and SVM in terms of specificity and accuracy [26]. Overall, ML models have shown superior and robust outcome compared to conventional methods [27] and LR model [28].

ANN is a supervised learning model that can learn highly complex and nonlinear relationships between the labelled input data. The shallow ANN model consists of an input, a hidden, and an output layer where the information moves in one direction from the input to the output layer. The input data to a shallow ANN model is the features of the labelled data and these features have a huge influence on the model's performance. Irrelevant features affect the model's accuracy negatively, so the careful selection of features has a significant effect on the desired output. Feature selection algorithms seek a subset of relevant features that have highly correlated information with the specific target [29]. Thus, feature selection techniques can improve the learning performance, reduce computational complexity, and build a robust and generalizable model [30]. Random forest (RF) is a highly efficient and popular feature selection algorithm in the area of bioinformatics due to its advantages in processing complex data structures and high-dimensional feature spaces [[31], [32], [33]].

ANN has been applied for the prediction of patient outcome following a TBI [14,25,28,34]. ANN outperformed Logistic regression (LR) models in terms of accuracy and area under the receiver operating characteristics curve (AUC) in eight different studies in predicting TBI outcomes [34], hospital mortality prognosis after TBI surgery [35], and in estimating the surgical decision for TBI patients (Y. C [36]. Moreover, ANN has been very promising in its application to the paediatric population [21,28,37]. ANN applied to a dataset of 12,902 TBI children was able to predict clinically relevant TBI (CRTBI) using clinical and radiologist-interpreted CT data with 99.73% sensitivity, 60.4% specificity, and 97.98% accuracy [21]. Chong et al. showed that predicting moderate to severe TBI in a dataset comprised of 195 children using ANN was significant compared to multivariable LR with a sensitivity of (94.9% vs. 82.1%), specificity of (94.9% vs. 82.1%), PPV (90.2% vs. 72.7%), and NPV (98.7% vs. 95.4%), and AUC of (98% vs. 93%). The ANN model was able to discover nonlinear interactions among the various clinical variables and associate them with the clinical outcome [37]. Furthermore, ANN outperformed many CT classification systems (Marshall, Helsinki, and Rotterdam) in six-months favourable (alive) and unfavourable (death) prognosis of paediatric TBI patients [28].

Deep ANN is built based on the shallow ANN architecture but uses multiple layers to allow for a deeper network structure [38,39]. Deep learning can discover and extract high-level features using representation learning to filter the features through the hidden layers and then picks out the most effective features [40] without an additional feature engineering phase to extract and select the features. Therefore, the essential differences between deep and shallow learning are the ability of deep learning to automatically discover and extract the features from the big data independently, and solve problems using an end-to-end approach [41].

To build a diagnostic model using the conventional/shallow ML approach, the data features are first extracted with the aid of a domain expert. Second, feature selection is carried out for dimensionality reduction, improving the learning process, and reducing computational complexity. Then, the selected features are fed to a shallow ML model to understand the relationship between the input (selected features) and the output (data label) through the classification process. The performance of the shallow learning is hugely affected by the feature selection process which is time consuming and depends on human expertise [41]. On the contrary, a deep learning model performs feature extraction, feature selection, and classification using the end-to-end approach without human intervention [40,41].

This work provides a model that has significant potential to be a future tool that can aid clinicians in the clinical decision-making for mTBI using the available data resources in EDs. This tool, if validated, would help to direct only the mTBI patients (9% compared to 100%) for longer term treatment. There have been numerous studies that have evidenced the potential of machine learning in the diagnosis of traumatic brain injury; however, these have been limited to 1) moderate to severe traumatic brain injury [37] and 2) all have used non-clinical data sources such as resting state functional network connectivity (rsFNC) [42] and Electroencephalography (EEG) [43]. To our knowledge, this is the first study that has applied machine learning to detect mTBI using standard clinical data resources that are readily available in EDs.

We hypothesized that 1) RF could rank the most relevant clinical predictors of mTBI in paediatrics and support the shallow ANN model in enhancing its performance to best predict a positive mTBI on CT (a hybrid model) and 2) Deep ANN would outperform the hybrid model even with no feature selection algorithm involved.