Autonomic Nervous System Dysfunctions as a Basis for a Predictive Model of Risk of Neurological Disorders in Subjects with Prior History of Traumatic Brain Injury: Implications in Alzheimer’s Disease

Abstract

Autonomic dysfunction is very common in patients with dementia, and its presence might also help in differential diagnosis among dementia subtypes. Various central nervous system structures affected in Alzheimer’s disease (AD) are also implicated in the central autonomic nervous system (ANS) regulation. For example, deficits in central cholinergic function in AD could likely lead to autonomic dysfunction. We recently developed a simple, readily applicable evaluation for monitoring ANS disturbances in response to traumatic brain injury (TBI). This ability to monitor TBI allows for the possible detection and targeted prevention of long-term, detrimental brain responses caused by TBI that lead to neurodegenerative diseases such as AD. We randomly selected and extracted de-identified medical record information from subjects who have been assessed using the ANS evaluation protocol. Using machine learning strategies in the analysis of information from individual as well as a combination of ANS evaluation protocol components, we identified a novel prediction model that is effective in correctly segregating between cases with or without a documented history of TBI exposure. Results from our study support the hypothesis that trauma-induced ANS dysfunctions may contribute to clinical TBI features. Because autonomic dysfunction is very common in AD patients it is possible that TBI may also contribute to AD and/or other forms of dementia through these novel mechanisms. This study provides a novel prediction model to physiologically assess the likelihood of subjects with prior history of TBI to develop clinical TBI complications, such as AD.

INTRODUCTION

Various central nervous system structures affected in Alzheimer’s disease (AD) are also implicated in autonomic nervous system regulation, and it has been hypothesized that the deficit in central cholinergic function observed in AD could likely lead to autonomic dysfunction [1–3]. Several feasible tests can be used in clinical practice for the assessment of parasympathetic and sympathetic functions, particularly in terms of cardiovascular autonomic modulation. Dysfunctions of the sympathetic nervous system (SNS) as well as an increased risk of developing neurodegenerative diseases, such as AD, post-trauma have both been implicated as risk factors in response to traumatic brain injury (TBI) [4–8].

Because evidence suggests a definite presence of autonomic dysfunction in dementia at various degrees, a method to assess TBI by monitoring of the autonomic nervous system (ANS) would allow for the detection and possible development of interventions, should TBI be identified to be part of the clinical history, of any subject at risk for developing AD. The SNS, together with the parasympathetic nervous system (PNS), form the ANS which is charged with unconsciously regulating internal organs and bodily functions. The terms paroxysmal sympathetic storm and sympathetic storming have been used to describe the presence of overly active sympathetic signaling after a severe TBI [5]. This sympathetic hyperactivity has been associated with pupil dilation, tachycardia, and hypertension [5]. On top of these symptoms, several other physical symptoms are known to present in an individual who has experienced a TBI [9]. Physical symptoms include, but are not limited to, those listed in Table 1.

Table 1.

TBI physical symptoms

Incontinence	Ataxia	Dementia
Headaches	Vomiting	Confusion
Drowsiness	Ipsilateral Pupil Dilation	Ipsilateral Hemiparesis
Contralateral Hemiparesis	Pain	Decreased Range of Motion
Peripheral nerve compression	Lymphedema	Hypertension
Vein Thrombosis	Anosmia	Vertigo
Nystagmus	Blurred vision	Spasticity

In addition to these symptoms, an increase in sympathetic firing also leads to an increase in muscular tonicity, and therefore disproportionate muscular development and body positioning. This phenomenon is linked to the excitatory: inhibitory ratio (EIR) model, where the brain stem EIR is overwhelmed by excessive brain activity causing the spinal EIR to become overwhelmed as well and lose its protective effect, making it unable to balance and/or inhibit the increased catecholamine circulated to peripheral organs. This manifests in end-organ dysautonomia of various end organs, including smooth muscle (rigidity) or cardiac muscle (tachycardia) [10]. These phenomena affect how the body looks during a physical examination.

In cases where a TBI has occurred that only affects one side of the head and brain, a hemispheric presentation can occur where a change in physical input and output develop ipsilaterally [11]. This ipsilateral hemisphericity can present as a change in muscle tone and limb angulation on only one side of the body. In addition to head trauma through contact, TBI can occur through non-impact trauma such as a blast or explosion. The blast waves that impact the brain can cause physical changes as well. The presentations of this type of TBI include symptoms that have been addressed above, such as vertigo and motion intolerance [12].

There are many theories regarding the underlying molecular mechanisms behind sympathetic storming. Nonetheless, the physical manifestation appears to result from an exaggeration of the SNS with a diminished PNS response or a disruption in communication within the two ANS components [13]. It is likely that such TBI-mediated ANS dysfunction may contribute to some of the TBI clinical phenotypes over shorter or longer terms. These long-term effects of trauma-induced brain damage are of increasing concern to a large portion of the population, from athletes, military personnel, to younger members of society who are physically active and at risk for head injury. Therefore, novel approaches that can identify TBI patients would also allow for an increased opportunity for future preventative therapeutic interventions for AD and other forms of dementia, which those with a history of TBI may be at risk of developing. In this study we therefore present a novel predictive model to assess patients’ risk of developing clinical complications in the future, based off of prior evidence that TBI may lead to changes in the ANS.

The protocol presented here provides a first step in the investigation of potential noninvasive TBI monitoring methods that can identify subjects who would be in need of more in-depth examinations in the long-term to monitor their risk of developing future neurological complications such as AD and other forms of dementia.

MATERIALS AND METHODS

We randomly selected and pulled de-identified medical records from 50 individuals seeking chiropractic therapy at PATCH Chiropractic offices, a large private chiropractic clinic that is located in Philadelphia, PA, who had been evaluated using the simple ANS evaluation protocol. From each of the cases, we extracted only information that was gathered by using the ANS evaluation protocol, as well as information on self-reported history of TBI exposure. Among the 50 randomly selected cases, self-reported TBI history was documented for 18 of the cases and was not documented for the remaining 32 cases. As we describe in more detail below, our ANS evaluation protocol is comprised of multiple components (Table 2). Using machine learning strategies, we assess whether information gathered from individual components (or a combination of components) of the ANS evaluation protocol may correctly segregate cases with or without a documented self-reported history of TBI exposure. Studies were approved by the Ethical Committee and Institutional Review Board at the Icahn School of Medicine at Mount Sinai.

Table 2.

Summary of major categories and subcategories that were addressed by the simple autonomic nervous system evaluation protocol. Shown are the 12 major categories assessed by the protocol. The table also shows the number of minor categories that were monitored for each of the major categories. Collectively, the autonomic nervous system evaluation protocol assesses a total of 102 subcategories

Major Categories	# Subcategories (Independent variables)
Posture	7
Face-eyes	5
Pupil reaction	13
Vitals	11
Breakdown of pursuits	8
Optokinetic testing	7
Vestibuloocular reflex testing	4
Pain sensation	13
Muscle tone-sympathetics	4
Abdominal	4
Standing	12
Deep tendon reflex	14

ANS evaluation protocol

As indicated above, our ANS evaluation protocol is comprised of a simple set of self-report questionnaires and a simple set of physical examinations that are designed to address 12 general categories: posture, face-eye features, pupil reaction, vital measurements, breakdown of pursuits, optokinetic measures, vestibuloocular reflexes, pain sensation, sympathetic muscle tone, abdominal characteristics, balance while standing, and deep tendon reflexes (Table 2). For each of the general categories, multiple subcategories were evaluated (Table 2). For example, under the general category of “posture”, subcategory assessments included characteristics pertaining to gait/sway, head tilt, heal perception, angulation of limbs, arm swing, and head rotations. In addition to these questionnaires, individual participants were subjected to simple physical examinations, including manual examination of joints/posture/range of motion/temperature, manual muscle examination for joint/motion/grade, objective observations of orthopedic characteristics, and deep tendon reflexes. Each of the physical examinations included multiple subcategories. As part of the examination for deep tendon reflexes in particular, the following were assessed: reflexes for biceps, triceps, thenar eminence, patellar, achilles, and blink reflex (BR). For example, the thenar test is when the thenar eminence is stroked briskly with a thin stick using moderate pressure, from proximal to distal locations. A positive response is considered if there is a single visible twitch of the ipsilateral mentalis muscle. Collectively, our ANS evaluation protocol encompasses 102 subcategories (Table 2), which we used as independent variables for segregating cases based on presence or absence of a documented history of TBI. For example, the discrete independent variable of convergence is tested when you have the patient follow an object, like the tip of a pen, with their eyes (Table 3). The pen is held about 1–2 feet away from the patient, placing the pen between the eyes, and then moved closer to the patient as they are instructed to maintain eye contact on the object. As the person follows the pen tip, it would make them move their eyes in a way that is similar to them crossing their eyes. The pen is moved all of the way in until both eyes stop moving inward and it is noted whether one eye keeps moving inward and the other one either stops or starts to act differently. The feature E/N (Table 3) is used to state if there are any abnormalities with how one eye moves compared to the other in this test, such as shaking or twitching. Upon review, E/N is used to indicate abnormal findings when testing independent variables, such as convergence, while on the other hand, blank or negative features are denoted to show when no abnormalities were found (Table 3).

Table 3.

Features that were defined based on discrete independent variables. Shown are 62 features that were defined as independent variables in machine learning strategies to segregate 49 cases based on the presence/absence of documented traumatic brain injury (TBI) history.

Gait-Sway = L	V:A L = 2 or 3	Crude Hearing = E/N	BR-L = absent
Gait-Sway = R	Abdominal = intact	Crude Hearing = blank	Thenar-R = 2 (in a 0–4 scale where 2 = normal)
Head Tilt = L	Abdominal = blank	Gillets+ = neg.	Thenar-R = blank
Head Tilt = R	Absent Bowel Sounds = present	Gillets+ = LorR	Thenar-L = 2 (in a 0–4 scale where 2 = normal)
Head Perception = ACC measurement	Absent Bowel Sounds = blank	Biceps-R = 2	Thenar-L = blank
Head Perception = blank	Tympany = none	Biceps-R = blank	Patellar-R = 2 (in a 0–4 scale where 2 = normal)
Angulation of Limbs = E/N	Tympany = blank	Biceps-L = 2 (in a 0–4 scale where 2 = normal)	Patellar-R = 1 (in a 0–4 scale where 2 = normal)
Angulation of Limbs = not E/N	Sway Eyes Closed = none	Biceps-L = blank	Patellar-L = 2 (in a 0–4 scale
Arm Swing = L	Sway Eyes Closed = L or R	Triceps-R = 2 (in a 0–4 scale where 2 = normal)	Patellar-L = 1 (in a 0–4 scale where 2 = normal)
Arm Swing = R	Neck = E/N	Triceps-R = blank	Achilles-R==2 (in a 0–4 scale where 2 = normal)
Head Rotation = E/N	Neck = absent of carotid bruit	Triceps-L = 2 (in a 0–4 scale where 2 = normal)	Achilles-R = 1 (in a 0–4 scale where 2 = normal)
Head Rotation = L or R	Temples = E/N	Triceps-L = blank	Achilles-L = 2 (in a 0–4 scale where 2 = normal)
V:A R=1	Temples = blank	BR-R = normal	Achilles-L = 1 (in a 0–4 scale where 2 = normal)
V:A R = 2	Convergence = E/N	BR-R = blank
V:A L = 1	Convergence = blank	BR-L = normal

ACC, accelerometer; L, left; R, right; BR, blink reflex; E/N, abnormality observed; neg. or blank, no abnormal observation; V:A R (or L) = 1 (or 2, or 3), width of the vein to width of the artery ratio in the right (or left) eye is 1:1 (or 1:2, or 1:3)

Data pre-processing

Information for some of the ANS evaluation subcategories was missing among the individual de-identified medical records we randomly retrieved. We therefore identified and removed subcategories for which information was missing among ≥80% (≥40) of the 50 cases. Following these pre-processing steps, the dataset contained only 31 remaining ANS evaluation subcategories which we decided to use as independent variables to segregate cases based on the presence or absence of documented history of TBI. Moreover, we removed one case (a case with a self-reported TBI history) from the study entirely because information was missing for 84 of the 102 subcategories. Thus, we ended up using 31 ANS evaluation subcategories as independent variables in machine learning classification of 49 cases (18 cases with documented history of TBI and 31 cases without).

Case segregation using machine learning strategies

We tested three classification methods using de-identified medical record information generated by the ANS evaluation protocol to classify the 50 cases into two groups: the presence or absence of a documented history of TBI exposure. These three classification methods are (1) support vector machine (SVM), (2) random forest (RF), and (3) naive Bayes (NB), which we will discuss in more detail below.

Support vector machine

SVM [14] is a distinguished family of classification methods that focus on defining an optimal hyperplane to separate the samples of different classes. It adopts the “kernel trick” in hyperplane definition, and introduces the “margin concept” as an effective strategy to counter the overfitting problems suffered from by many earlier classification methods. In this study, we used the SVM methods implemented in the e1071 Vector Machine Support Package [15] in R, and selected the radial basis function (RBF) kernel, which has been shown to produce outstanding grouping performances in previous publications [16]. The RBF kernel function could be defined as:

$$\text{K}(\text{x},\text{y})={\text{e}}^{|-\text{y}||\text{x}-\text{y}{|}^{\text{Z}}}$$

Here, x and y are samples with gene expression and γ is the kernel parameter. In this study, the kernel parameter γ and SVM penalty parameter C were optimized by nested cross-validation over the γ value (2⁻¹⁵, 2⁻¹³, … , 2¹) and C value (2⁻⁵, 2⁻³, … , 2¹⁵).

Random forest

RF [17] is a classification algorithm that uses an ensemble of unpruned decision trees, each of which is built on a bootstrap sample of the training data using a randomly selected subset of variables. RF uses both bagging (bootstrap aggregation) for combining unstable learners and random variable selection for tree building. RF has been successfully applied in classification problems of biological data, and has demonstrated excellent predicting performance even when feature variables were highly noisy [18]. We employed the RF implementation in the R package randomForest [19] in this study. There are three important parameters whose values need to be determined: mtry (the number of input variables tried at each split), ntree (the number of trees to grow for each forest), and node size (the minimum size of the terminal nodes). We considered different parameter configurations for their values of mtry = (0.5, 1, 2), ntree = (500, 1000, 2000), and node size = 1, as recommended in a previous study [20]. The best-performing parameters were selected by a 2-layer cross-validation scheme, which will be described in a later section.

Naive Bayes

NB [21] is a rule generator based on Bayes’ rule of conditional probability. It uses all attributes and allows them to contribute to decision-making as if they were all equally important and independent of one another. We used the following conditional probability formula in which P(H/E) denotes the probability of event H (conditional on event E), and En denotes the nth attribute of the instance:

$$\text{P}(\text{H}\mid \text{E})=\frac{\text{P}(\text{H}1\mid \text{H}).\text{P}(\text{H}2\mid \text{H})\dots \text{P}(\text{En}\mid \text{H})}{\text{P}(\text{E})}$$

The NB implementation provided in the e1071 package in R was used in this analysis, in which the parameter “Laplace” (positive double controlling Laplace smoothing) was set by default (disables Laplace smoothing).

Selection of most efficacious ANS evaluation features for classifying cases using machine learning strategies

Feature selection is a domain of active research in machine learning. Feature selection methods are developed with the observation that not all features are contributing to the predicting performance of the classification method. A properly designed feature selection method intelligently identifies a small subset of the features, which, in conjunction with a proper classification method, are more effective in distinguishing the samples with differing class labels. Generally, established feature selection methods can be broadly categorized into three types: (1) filter methods, (2) wrapper methods, and (3) embedded methods [22]. Filter methods usually rank the features based on univariate measures and select the most significant feature genes. Wrapper methods search for a subset that is found to be optimal with respect to a subset evaluator such as a classifier. Embedded methods incorporate the search algorithm into the classifier. In this study, we attempted three feature selection methods: the Chi-squared method, the correlation-based feature selection method, and the SVM-based recursive feature elimination method.

Chi-squared method

The Chi-squared method is a filter method. It evaluates features individually by measuring their chi-squared statistic with respect to the classes, and ranks all features. The gene expression values are first discretized into several intervals using an entropy-based discretization method, and the chi-square statistic of each feature is defined as follows:

$${\text{X}}^2=\sum _{\text{i}=1}^{\text{m}}\sum _{\text{j}=1}^{\text{K}}\frac{{\left({\text{A}}_{\mathrm{i}}\cdot \frac{{\text{R}}_{\mathrm{i}}\cdot {\text{c}}_{\mathrm{j}}}{\text{N}}\right)}^2}{\frac{{\text{R}}_{\mathrm{i}}\cdot {\text{c}}_{\mathrm{j}}}{\text{N}}}$$

where m denotes the number of intervals; k is the number of classes; N is the total number of patterns; R_i represents the number of patterns in the ith interval; C_j is the number of patterns in the jth class; and A_i denotes the number of patterns in the ith class.

Correlation-based feature selection method (CFS)

The CFS method [23] is also a filter method. This method adopts a heuristic to measure the correlation between attributes and rewards that features subsets in which each feature is highly correlated with the class and uncorrelated with other subset features. Given a feature subset S consisting of k features, the merit of the subset is defined as:

$${\text{ Merit }}_{{\text{S}}_{\mathrm{k}}}=\frac{\text{k}\overline{{\mathrm{r}}_{\text{cf}}}}{\sqrt{\text{K}+\text{k}(\text{k}-1){\overline{\text{r}}}_{\text{ff}}}}$$

where $\overline{{\mathrm{r}}_{\text{cf}}}$ is the average value of all feature-classification correlations, and $\overline{{\mathrm{r}}_{\text{ff}}}$ is the average value of all feature-feature correlations. The CFS criterion is defined as follows:

$$\text{CSF}=\frac{{\text{r}}_{\text{cf}1}+{\text{r}}_{\text{cf}2}+\dots +{\text{r}}_{\text{cfk}}}{\sqrt{\text{K}+2\left({\text{r}}_{\text{f}1/\text{f}2}+\dots +{\text{r}}_{\text{fi}\text{fj}}+{\text{r}}_{\text{fk}\text{fl}}\right)}}$$

where the r_cfi and r_fifj are referred to as correlations.

SVM-based recursive feature elimination method (SVM-RFE)

SVM-RFE [18] (abbreviated as RFE in the remainder of this manuscript) is a wrapper method. It is an iterative algorithm that fits SVM classification models by discarding genes with small impacts on classification and selecting the smallest subset of genes that participate in the best performing classification model. The iterative procedure of this method is:

1. Train the SVM classifier.

2. Compute the ranking criterion for all features.

3. Remove the feature with the smallest ranking criterion.

Features were ranked by evaluating how well an individual feature contributes to the separation.

Ranking selected features using a two-layer cross-validation strategy

A five-fold cross-validation was used to estimate the performance of the classification algorithms (outer-loop). In order to find the best parameters, another ‘nested’ leave-one-out cross-validation was performed on the five original training sets (inner-loop). For each combination of the classifier parameters, we first obtained cross-validation performances and selected the best performing parameters inside this inner loop of cross-validation. Then we built a classification model with the best parameters on the original training set and applied this model to the original testing set.

This two-layer, nested cross-validation scheme was first proposed by Liu et al. [16] and it is an effective procedure that curbs the overfitting problem while objectively evaluating the performance of the classification models for unobserved data. The two-layer cross-validation implemented in this study was carried out with a locally developed R script.

Classification performance evaluation metrics

We used four classification performance metrics: accuracy, sensitivity, specificity, and Mathew’s correlation coefficient (MCC). The formulas used to derive these values are:

$$\text{ Accurancy }=\frac{(\text{TP}+\text{TN})}{(\text{TP}+\text{FN}+\text{FP}+\text{TN})}$$

$$\text{ Sensitivity }=\frac{\text{TP}}{(\text{TP}+\text{FN})}$$

$$\text{ Specificity }=\frac{\text{TN}}{(\text{FP}+\text{TN})}$$

$$\text{MCC}=\frac{\text{TP}\times \text{TN}-\text{TP}\times \text{FN}}{\sqrt{(\text{TP}+\text{FP})(\text{TP}+\text{FN})(\text{TN}+\text{FP})(\text{TN}+\text{FN})}}$$

where TP stands for true positive, TN for true negative, FP for false positive, and FN for false negative. Of these metrics, MCC is not used frequently outside the machine learning field. Comparing with accuracy, MCC is a more objective measure of the predicting performance when the sample sizes between classes are very unbalanced.

RESULTS

Our ANS evaluation protocol encompasses 102 subcategories (Table 2). However, information for some of the ANS evaluation subcategories was missing among individual’s de-identified medical records we randomly retrieved. As we described in the Materials and Methods, we identified and removed subcategories for which information was missing among ≥80% of the cases. This resulted in 31 ANS evaluation subcategories as independent variables in machine learning classification of 49 cases (18 cases with documented history of TBI and 31 cases without).

Among the selected 31 ANS evaluation subcategories, 2 were continuous variables, and 29 were discrete variables. For the discrete variables, a commonly applied “sparse coding” scheme [16] was employed to define 2 features from each variable. For instance, for the discrete variable “Gait-Sway”, which could take one of two possible values, “L” and “R”, we defined two features, namely “Gait-Sway = L” and “Gait-Sway = R”. The values for these two features were equal to 1 and 0, respectively, for participants whose “Gait-Sway” value was “L”, and were 0 and 1, respectively, for participants whose “Gait-Sway” value was “R”. A total of 58 features were defined based on these 29 discrete independent variables (Table 3). For the two continuous independent variables, “BP L” and “BP R” (which represent two sets of blood pressure readings, both high and low, taken from the left and right arms), we treated the four readings—high/L, low/L, high/R, and low/R—separately. For each reading, we scaled the values across all participants into the scale of [0,1]. An important consequence of this scaling is that the values of features, those generated from discrete variables and those generated from continuous variables, were precisely in the same range. Blank values were imputed using the mean value for all other participants. This procedure resulted in the definition of 4 additional features. In total, 62 features defined for the 49 remaining cases (18 cases with and 31 cases without a documented history of TBI exposure) were used for machine learning classification.

Performance prediction of three classification methods

The task of constructing a prediction model for a group classification problem and objectively evaluating its predicting performance for future untested samples, while still controlling for an overfitting problem, can conceptually be considered to comprise of the following components: (1) selecting a proper classification method, (2) selecting a proper feature selection method, (3) determining the proper number of features to include (for those feature selection methods that do not choose the “optimal” feature numbers automatically), (4) selecting proper model parameter values, and (5) evaluating prediction model performance. In practice, we carried out this task in a streamlined manner, where we examined a total of 9 classification methods: feature selection method combinations (3 classification methods and 3 feature selection methods)—integrating components (1) and (2) in a single step. In addition, we integrated components (4) and (5) together using a two-layer cross-validation scheme established by Liu et al. [16] and accomplished this (3) by monitoring changes of the model performance as the feature number decreased.

One of the three feature selection methods we included in this study, CFS, selects the “optimal” feature numbers automatically. The testing of model performance with decreased feature numbers was conducted for the other two feature selection methods, RFE and Chi-squared. We expected that as the number of features decreased, the model performance would show a biphasic characteristic. It would first increase, as the feature selection method eliminates those features that do not “help” the model performance. They are the features that add noise rather than contribute to the predicting accuracy of the model, thus elimination of these features would lead to improvement of the model performance. After all the “bad” features have been eliminated, the model performance then enters the second phase, where further removing features would result in degradation of the model performance. This is because at this time, all remaining features are contributing to the model performance, and removal of some of them would “hurt” the model. As shown in Fig. 1, despite considerable noise levels (which could be expected due to the relatively small number of samples included in this study), the biphasic characteristic was clearly visible for some of the classification methods, such as feature selection method combinations, including SVM-RFE (Fig. 1A), SVM:Chi-squared (Fig. 1B), and RF:Chi-squared (Fig. 1C). Some other classification methods, involving feature selection method combinations, in particular (NB:Chi-squared, Fig. 1D) showed a small downfall phase at the end of the curve, but lacked the uprising phase. Other classification methods of feature selection method combinations (NB:RFE, Fig. 1E, and RF:RFE, Fig. 1F) showed a complete absence of the biphasic characteristic, suggesting that the feature selection method did not work effectively in these cases.

Fig. 1.

Model performance changes with number of features remaining when using a combination of classification methods and feature selection methods. A–C) A biphasic characteristic is observed for some of the classification methods: feature selection method combinations including SVM-RFE (A), SVM:Chi-Squared (B), and RF:Chi-Squared (C). D) A classification:feature selection method combination using NB:Chi-Squared showed a small downfall phase at the end of the curve, but lacked the uprising phase. E,F) A classification:feature selection method combination using NB:RFE (E) and RF:RFE (F) showed a complete absence of the biphasic characteristic. For each of the classification:feature selection method combinations shown, the top panel depicts model accuracy and the bottom panel depicts Mathew’s correlation coefficient. SVM, support vector machine; RFE, SVM-based recursive feature elimination method; RF, random forest; NB, naive Bayes.

Best predicting performance achieved

Table 4 summarizes the best predicting performance achieved for each of the 3 classification methods without feature selection, as well as in conjunction with each of the 3 feature selection methods. Among the combination of classification models and diverse feature selection approaches that were tested, SVM-RFE achieved the most impressive predicting performance: at 27 remaining features, it achieved an accuracy of 0.98, and MCC of 0.96, i.e., of the 49 participants, only 1 was misclassified in the cross-validation based testing (Table 4). A detailed list of the 27 features that were used in SVM-RFE is shown in Table 5.

Table 4.

Best model performance achieved by each classification method with each corresponding feature selection method.

Classifier	Feature selection	Sensitivity	Specificity	Accuracy	MCC	# Featu
SVM	All Features	0.78	0.90	0.86	0.61	62
	RFE	1.00	0.97	0.98	0.96	27
	Chi-Squared	0.94	0.94	0.94	0.84	57
	CFS	0.78	0.77	0.78	0.49	4
RF	All Features	0.50	0.94	0.78	0.37	62
	RFE	0.83	0.90	0.88	0.67	17
	Chi-Squared	0.78	0.94	0.88	0.64	43
	CFS	0.56	0.90	0.78	0.39	4
NB	All Features	0.83	0.84	0.84	0.61	62
	RFE	0.94	0.90	0.92	0.81	10
	Chi-Squared	0.89	0.94	0.92	0.77	10
	CFS	0.72	0.90	0.84	0.55	4

CFS, correlation-based feature selection method; MCC, Mathew’s correlation coefficient; NB, naive Bayes; RF, random forest; RFE, SVM-based recursive feature elimination method; SVM, support vector machine

Table 5.

A listing of 27 features that supported the best predicting performance. Among all models, support vector machine-based recursive feature elimination method achieved the best prediction performance using these 27 features.

Gait-Sway = L	Tympany = none	Head Perception = blank
Head Tilt = R	Tympany = blank	Triceps-L = 2
V:A L =2 or 3	Sway Eyes Closed = none	Abdominal = intact
Head Rotation = E/N	Absent Bowel Sounds = present	Biceps-L = blank
Head rotation = L or R	Absent Bowel Sounds = blank	Patellar-L = 2
Gillets = L or R	BR=L=2	BP R/high
Convergence = not E/N	BR-L = blank	Head Perception = ACC measurement
Head Tilt = L	Triceps-L = blank
V:A R = 2	Biceps-L = 2
Sway Eyes Closed = L or R	Abdominal = blank

ACC, accelerometer; L, left; R, right; BR, blink reflex; E/N, abnormality observed; blank, no abnormal observation; V:A R (or L) = 1 (or 2, or 3), width of the vein to width of the artery ratio in the right (or left) eye is 1:1 (or 1:2, or 1:3); triceps (or biceps, or patellar)-L = 2 (normal in a scale of 0–4); BP R/high, blood pressure from the right arm/high readings

Model validation

The data used in this study are slightly unbalanced, i.e., among the 49 participants, 18 were documented for a history of TBI exposure and 31 were not. For unbalanced datasets like this, MCC, instead of accuracy, would be a more proper performance metric to examine. However, for ease of explanation, we choose to focus on accuracy in the following discussion because of its simple and intuitive physical interpretation; the accuracy is essentially the proportion of cases whose group labels are correctly predicted. For instance, the best SVM-RFE model achieved an accuracy of 0.98, which means that of the 49 participants, 0.98 × 49 = 48 were correctly classified. One way to understand how good a constructed predicting model is, is by putting it into the context of comparison with a “random guessing” model; the latter can be regarded as equivalent to a “negative control” for researchers more familiar with experimental research. For an unbalanced dataset that consists of a 31:18 splitting of participants between groups, the best “random guessing” model would produce an accuracy of 31/(31+18) = 0.63. Thus, a model of 0.98 accuracy represents a substantial improvement in predicting performance as compared to the “random guessing” model.

DISCUSSION

There is large evidence associating ANS perturbations to AD as well as ANS perturbations to TBI [1, 8, 24–29]. However, despite the large body of literature suggesting that TBI might pose a risk for accelerating the onset and possibly the progression of AD dementia [6, 7, 30, 31], little is known about ANS dysfunctions as a potential link associating the two conditions. Our present study was designed to investigate a readily-available ANS evaluation protocol as a way to identify and eventually monitor subjects at risk for TBI complications.

Initially, when a subject presents with short-term physical complications post-TBI, there are different presentations which can be bilateral or ipsilateral in nature. If a subject presents with greater sympathetic muscular tone on one side of their body, or with bad balance on the contralateral side of their body due to nerves crossing at the cerebellum, further investigation would be performed on the side of the muscular imbalances. If there are random presentations on both sides of the body, it is customary to then search for a more “centralized” area of TBI. For any subject presenting with an abnormality in sympathetic tone, a battery of tests would be conducted in order to rule in/out the likelihood of TBI on either side of the brain. The more positive the results that are returned from an examination, the more “weight” would be put into a TBI diagnosis, subsequently signaling the need for further testing at the very least. The protocol presented here therefore provides a simple first step in the investigation of potential TBI identification methods and, as such, will be an invaluable tool in ascertaining subjects who would be in need of more in-depth examinations in the long-term, to monitor their risk of developing future cognitive complications.

Evidence from our study supports the feasibility that our simple ANS evaluation protocol may reflect a history of TBI exposure, thereby identifying those at risk of developing long-term, detrimental brain responses later in life caused by head injury, such as AD. In particular, we identified a prediction model, established on the application of the SVM-based recursive feature elimination (SVM-RFE) method in the analysis of a subset of 27 select features from the ANS evaluation, which is effective in correctly segregating cases with or without a documented history of TBI exposure with an accuracy of 0.98, and an MCC of 0.96. With the implications of TBI in the etiopathology of neurodegenerative disorders such as AD [6, 32, 33], the long-term effects of trauma-induced brain damage is of increasing concern to a large portion of the population. Therefore, the ability of this novel prediction model to identify subjects with prior TBI history will allow for an increased opportunity for future preventative therapeutic interventions in neurodegenerative disorders and age-related cognitive decline.

We note that the model performance evaluations made in this study were conducted in a 2-layer cross-validation setting, thus they are objective evaluations of the model performance when they are tested on unknown future samples, as long as those future unknown samples are assumed to be taken from the same population of samples as were the ones used in our model construction work—the weakest assumption that can be made in this type of study.

The performance measurements obtained in this study should be understood in lieu of the high noise levels in the plots presented in Fig. 1 (which was expected because of the small sample number included in this study). Although at 27 remaining features, an accuracy of 0.98 was achieved. It is an overstatement that when tested using future samples, the model is expected to consistently achieve 0.98 accuracy. Following the smoothing of the curves, it appears that accuracy around the level of 0.90 would be more realistically achievable for future samples possessing the same population characteristics as the samples used in model construction.

Collectively, results from our present study demonstrating the efficacy of our prediction model, based on SVM-RFE analysis of select features from the ANS evaluation in correctly differentiating between cases with or without a documented TBI history, are consistent with the hypothesis that trauma-induced ANS dysfunctions may contribute to clinical TBI features. Our ANS evaluation protocol is designed as a simple assessment approach for potential disturbances in the ANS to identify and monitor those with a prior record of TBI. The study will therefore provide a novel prediction model as a readily applicable physiological assessment of the likelihood of experiencing potentially negative impacts of prior TBI injury in subjects who may also be at risk of developing AD and other forms of dementia.