4

Clinical Decision Support from Data to Impact

___________________

¹ This list is the rapporteurs’ summary of points made by the individual speakers identified, and the statements have not been endorsed or verified by the National Academies of Sciences, Engineering, and Medicine. They are not intended to reflect a consensus among workshop participants.

The fourth session of the workshop included a firsthand account of difficulties patients with traumatic brain injury (TBI) can face in pursuing diagnosis and treatment. The session also provided an overview of a handful of innovative diagnostic technologies, including machine learning approaches, ultrahigh-performance imaging, and electrophysiological and eye-tracking devices. The objectives of the session included (1) examining emerging approaches to using large and complex datasets from electronic health records, advanced imaging, and other sources to inform clinical care, and (2) exploring trends in multimodal TBI classification to describe clinical phenotypes and inform treatment approaches. Michelle LaPlaca, professor

and director of the Neurotrauma and Translational Bioengineering Laboratory at the Georgia Institute of Technology and professor at Emory University, moderated the session. She described a need for practices that can be implemented in the clinic that better characterize a person’s injury given the heterogeneity and complexity of TBI, and the need to use such knowledge to guide care. LaPlaca remarked that consideration will also need to be given to the copious quantities of data generated by clinical decision tools and the potential for confusion that large amounts of data can create.

A LIVED EXPERIENCE PERSPECTIVE

Patricia Adell, TBI survivor and managing partner at Real Estate Solutions Group, gave an account of sustaining a TBI and the barriers she encountered in obtaining an accurate diagnosis and treatment. She described three incidents of head impact. The first time, she was thrown off her bike and her head slammed against the ground, but she was wearing a helmet and experienced no repercussions. The second head impact occurred when she was struck by a car and hit the pavement. Despite a lump that remains on her head to this day, Adell had no other repercussions from that injury. Her third accident was far less dramatic than the first two. She slipped while walking and fell onto the pavement. Upon standing, she immediately realized something was wrong and had difficulty walking home. Struggling with balance, light, and sound, Adell went to the emergency department (ED), where physicians looked into her eyes, performed a computed tomography (CT) scan, and told her that although she likely had a concussion, they lacked a definitive method of diagnosing it.

After being informed that her symptoms were consistent with concussion and that she did not have a brain bleed, she was instructed to go home and lie down. Adell called her physician, who gave her the same guidance. Searching for help, she reached out to a friend who is a sports medicine orthopedist. He made calls on Adell’s behalf that led to finding Christina Master, pediatrician and sports medicine specialist at Children’s Hospital of Philadelphia and professor at the University of Pennsylvania. Adell described connecting with Master as the beginning of her TBI education and recovery.

After diagnosing Adell, Master prescribed physical therapy. Adell recounted that over the course of 3–4 months, she began to feel like herself again. She stated that Master helped her throughout her recovery experience, which included a postconcussion brain bleed. She remarked that most of the TBI survivors she meets have not been able to find a recovery pathway with a proactive provider who recommends interventions. Adell remarked that in her experience, primary care providers (PCP) have little

to no knowledge about TBI treatment and seem unaware of current care options and appropriate referrals to care.

The majority of the TBI survivors she has communicated with have experienced the barrier of providers who did not recommend specialists or interventions. Given that TBI can lead to poor outcomes and even death, Adell stated her gratitude at finding effective treatment. She commented that in the absence of tools in the ED to diagnose and assess the severity of TBI, patients are left without knowing what is wrong, whether it can be treated, and what kind of outcome to expect. Adell remarked that she is surprised that science has been slow in this area, and she is heartened to learn that diagnostic tools are currently being developed and implemented.

STATISTICAL AND MACHINE LEARNING APPROACHES TO TBI PREDICTION

Adam Ferguson, professor and director of data science in the Brain and Spinal Injury Center at the University of California San Francisco, discussed machine learning approaches to building TBI prediction models and incorporating these into electronic health records (EHR).² He remarked that the care failures Adell experienced are emblematic of the need for better prediction models. TBI is a complex and extremely heterogeneous condition and affects a variety of biological features that evolve over time (Irvine and Clark, 2018). Furthermore, TBI increases long-term risk for dementia (Shively et al., 2012). Having validated predictive models could help guide primary care providers in improving TBI care, said Ferguson.

Validated Prediction Models

Two major TBI clinical prediction models currently have traction for the acute phase of injury, Ferguson stated. The Corticosteroid Randomization after Significant Head Injury (CRASH) model and the International Mission for Prognosis and Clinical Trials in TBI (IMPACT) model were both published in 2008 to provide the probability of 6-month mortality and risk of 6-month unfavorable outcome at 6 months after TBI.

The CRASH study was a randomized controlled trial of corticosteroid therapy in Europe involving 10,000 patients with TBI (Perel et al., 2008). The trial was unsuccessful, but researchers used the study data to develop

___________________

² Ferguson noted that his work was supported by grants from the National Institutes of Health, Department of Veterans Affairs, Department of Defense, Defense Advanced Research Projects Agency, Department of Energy, Craig H. Neilsen Foundation, and Wings for Life Foundation. He is involved in data science consulting for Santa Clara Valley Medical Center, Neuronasal Inc., and SpineX Inc., and he is involved in an industry collaboration with DataRobot (in-kind) as part of its AI for Good program.

a clinical prediction model and cross-validated it with 8,500 patients from IMPACT (Steyerberg et al., 2008). Stratified by levels of economic development, CRASH is an international model that performs differentially in high-income versus low- and middle-income countries. The IMPACT model, on the other hand, was developed from data on 8,500 patients in 11 completed clinical trials that failed from 1984 to 2007. Data from IMPACT trials were curated, placed into a database, and cross-validated on over 6,000 patients from the CRASH trial. The two models used data from each other in developing the prediction tools, and Ferguson also noted that both models are valid for TBI patients having Glasgow Coma Scale (GCS) scores of less than 12 and therefore exclude those with mild TBI.

Web-Based Prediction Calculators

Both CRASH and IMPACT offer web-based calculators to encourage their use, said Ferguson.³ Anyone with an Internet connection has access to these tools, yet many acute care providers who are not TBI specialists are unaware of them, he remarked. These models include fields for various clinical features. Once values are entered into these fields, the calculators provide a prediction of 6-month outcomes. Both calculators include optional fields for CT information, and the IMPACT tool includes optional fields for biofluid biomarker values.

The Transforming Research and Clinical Knowledge in TBI (TRACK)TBI study (McCrea et al., 2021), a large-scale observational study, validated both the IMPACT and CRASH prognostic models in a recent paper (Yue et al., 2024). This paper assessed the measures on patients enrolled in TRACK-TBI from 2014 to 2018 and found that both models adequately discriminated mortality and unfavorable outcome. However, the models overpredicted mortality in the overall cohort for patients with severe and moderate TBI when certain data fields were included; CRASH predictions that incorporated CT values and IMPACT predictions that used biomarker values were found to overpredict mortality for this group. The paper stated, “This suggests the presence of predictors in contemporary TBI care that are not accounted for by these models, which were developed using data from over 2 decades ago” (Yue et al., 2024). This raises the question of how to identify and navigate the full set of predictors for TBI outcomes beyond these established models, Ferguson said.

___________________

³ The CRASH head injury prognosis tool is available at http://www.crash2.lshtm.ac.uk/Risk%20calculator/index.html. The IMPACT prognostic calculator is available at http://www.tbi-impact.org/?p=impact/calc (both accessed June 4, 2024).

Multidimensional TBI Outcome Prediction Models Using AI and Machine Learning

Artificial intelligence (AI) and machine learning tools, which are designed to incorporate as much information as possible, are well suited to incorporating a full set of potential predictors associated with TBI, Ferguson stated. These tools have been used in drug development (Chakradhar, 2017), and early signs indicate that applying them to TBI holds potential for improving prediction measures. This type of analysis helps identify groupings of patients with similar features (TBI phenotypes) to further inform development of more tailored TBI classification and care approaches.

Jessica Nielson, one of Ferguson’s team members, conducted a topological data analysis examining numerous TBI variables, including GCS score, neurocognitive function, functional changes, molecular biomarkers, and tissue imaging (Nielson et al., 2017). Machine learning tools were used to identify linkages across data layers to establish patient groups and inform patient stratification (see Figure 4-1). Topological data analysis projects all input data onto a map that describes the disease space of TBI and clusters patients that AI deems similar. This allows comparison of an individual patient to other TBI patients, Ferguson explained.

Ferguson described the TBI map as resembling a Rorschach test that can be color-coded for specific clinical variables. For instance, as shown in Figure 4-1, TBI patients—each represented as a point within the map—can be colored according to CT scan results. Patients whose CT scan showed brain injury are in red and patients without brain lesions on CT are in blue (Nielson et al., 2017). The map was also color-coded to reflect each patient’s magnetic resonance imaging (MRI) results. Comparing the CT and MRI maps indicates that a subset of patients without lesions on CT did have injury findings on MRI. Additionally, color-coding for functional changes based on Glasgow Outcome Scale Extended measurements taken at 3 months and 6 months revealed a phenotype of mild TBI with a positive MRI finding and degeneration of function. Layering additional variables indicated that this phenotype features enhanced levels of post-traumatic stress.

A novel clinical prediction model that makes use of AI uses patient information—such as blood-based biomarker levels and absence of lesion on CT—to predict outcome, said Ferguson. In collaboration with the U.S. Department of Energy, supercomputers were used to analyze all intake features from the TRACK-TBI pilot study and all TBI outcome features (Tritt et al., 2023). This process identified 20 subgroups of patients within the dataset. After stratifying patients based on multiple dimensions of intake and outcome, the researchers developed a canonical correlation in which the full set of intake features is correlated with the full set of outcome features. From this, they generated an R-squared value (i.e., a coefficient of determination) that can be used to predict TBI outcome. Ferguson highlighted several examples using different methods from other legacy datasets, such as a machine learning algorithm based on data from the Protective Effects of Progesterone (ProTECT) III trial, and a mixture model framework (Bark et al., 2024; Kaplan et al., 2022).

Implementation of TBI Prediction Models

Ferguson stated that an obvious method of implementing prediction models in clinical care is embedding tools into EHRs. Randomized controlled trials have examined the effects of integrating clinical decision support tools into EHR systems. A meta-analysis of such randomized controlled trials found a statistically significant effect of decision support systems evident in the prevention of morbidity in every disease studied (Moja et al., 2014).

Ferguson underscored that the effect of embedding tools or health care interventions in EHRs can be significant, as evidenced in other areas of neurology. For instance, multiple sclerosis (MS) researchers collaborated with patients, caregivers, and physicians to develop the MS NeuroShare system, which contains a series of MS management dashboards. The dashboards can be included in the EHR via the MyChart system and have been imple-

mented in the Sutter Care System in California. This enables physicians, patients, and caregivers to view the prediction models in situ and in real time. Research indicates that this type of system allows sophisticated prediction activity (Bove et al., 2021). Furthermore, a study found that using enriched EHR made it possible to identify early signs of MS (prodromes) up to 5 years before the first symptoms were detected by physicians, Ferguson explained (Nelson et al., 2022).

Regarding TBI in children, Ferguson said that there has been incremental progress in implementing a pediatric clinical decision support tool intended to avoid unnecessary CT scans and associated radiation exposure on children’s developing heads. Two studies demonstrated that the implementation of such decision support tools decreases the rate of pediatric CT scans (Atabaki et al., 2017; Masterson Creber et al., 2018). Furthermore, the studies found no significant change in the rate of return visits to the ED within 7 days, and return visits were not associated with misdiagnosis. He stated that evidence indicates that embedding prediction tools into the EHR changes TBI care and may improve recovery. Ferguson remarked that implementing TBI machine learning prediction models into EHRs will be key for clinical translation and physician education in guiding decision-making through clinical decision support pathways.

ULTRAHIGH-PERFORMANCE MRI IN TBI

J. Kevin DeMarco, neuroradiologist at Walter Reed National Military Medical Center, discussed advanced MRI technologies and their potential role in TBI diagnosis.⁴ There remain substantial unmet imaging needs, he said, particularly for patients assessed as having mild TBI and who continue to experience symptoms but do not show abnormal results on current brain imaging. Even a 3 Tesla (3T) MRI—a form of MRI that features a stronger magnet and creates more detailed images than 1.5T MRI technology—does not always identify anatomical changes related to mild TBI in a significantly symptomatic patient. Research suggests that diffusion-weighted imaging may yield opportunities to identify abnormalities that do not have an anatomical correlate.

Advanced Imaging Technology in TBI

DeMarco noted that the TRACK-TBI study involved direct recruitment from Level 1 trauma centers of TBI patients with head CT scans. Among

___________________

⁴ DeMarco noted that he received government research funding, but that views expressed in his presentation are those of the authors and do not reflect the official policy of the Department of Army/Navy/Air Force, Department of Defense, or the U.S. government.

367 patients with mild TBI, 28 percent had abnormalities on CT scans, and 47 had positive findings on MRI (Palacios et al., 2022). Researchers sought to determine whether diffusion-weighted imaging would indicate abnormality after TBI, and whether this could be associated with a poorer recovery outcome.

Several types of diffusion-weighed imaging exist. Diffusion tensor imaging (DTI) is a simplified model to describe the anisotropic Brownian motion of water molecules in the brain, and it assumes the Gaussian diffusion process (i.e., no restrictions or barriers are present). The study revealed that DTI abnormalities in axonal diffusivity and mean diffusivity were independently associated with a 6-month incomplete recovery after TBI. Other studies have examined several higher-order diffusion MRI metrics, DeMarco shared. Diffusion kurtosis imaging (DKI) estimates both the Gaussian and non-Gaussian components of diffusion and is sensitive to boundaries and restrictions of white matter in addition to the direction of brain fiber tracts. DeMarco explained that DKI reveals radial diffusivity changes that begin between less than 72 hours to 7 days postinjury, at which point imaging gradually begins to normalize over time (Muftuler et al., 2020).

New Technologies in TBI Imaging

DeMarco remarked on a shift from large-bore whole-body MRI technology—a device that can scan the entire body but does not optimally image the brain—to dedicated head-only gradient 3T systems and said that businesses are recognizing the opportunity to develop ultrahigh-performance brain MRI scanners. These dedicated systems optimize diffusion-weighted imaging and enable evaluation of intra-axonal water using ultrahigh b-value diffusion-weighted imaging with a high signal-to-noise ratio, making it possible to identify abnormalities.

DeMarco provided an overview of how MRI technology generates an image. A whole-body MRI scanner contains miles of wire, and it operates by chilling this wire to extremely cold temperatures to induce superconducting properties that generate high static magnetic field strengths. A radio frequency coil transmits radio waves, which the patient absorbs and releases, creating an image. Specifying that gradient refers to linear change, he explained that gradient coils in the three orthogonal planes (anterior-posterior, right-left, superior-inferior) to the patient generate a small magnetic field that modifies the large static magnetic field, making it possible to localize the point in space where the MR signal is coming from and to generate various types of contrast, such as diffusion. DeMarco noted that MRI technology has evolved significantly in the past 20 years, but this has not extended to the whole-body gradient coil systems. The Joint Program Com-

mittee for Combat Casualty Care (JPC-6)⁵ funded a collaboration between Walter Reed National Military Medical Center, Uniformed Services University, and GE HealthCare Technologies and Innovation Center to create a novel MRI gradient coil to accommodate only the head, thereby eliminating the need to use a large whole-body gradient coil system when assessing head injury and ensuring ultrahigh performance optimized for brain imaging.

DeMarco explained attributes that make a smaller gradient system desirable. High-performance gradient coils as large as those in whole-body MRI scanners induce substantial peripheral nerve stimulation, limiting the maximum gradient amplitude that can be used. The gradient coil funded by JPC-6, dubbed Microstructure Anatomy Gradient for Neuroimaging with Ultrafast Scanning (MAGNUS), has a smaller gradient coil design than whole-body MRI scanners. MAGNUS can produce high gradient strength (G_max) with a simultaneously very high slew rate (speed at which the gradient signal can be ramped on and off) without painful peripheral nerve stimulation. A faster slew rate enables the MRI signal to be captured more quickly, resulting in decreased signal loss. This increases the signal-to-noise ratio and reduces distortion, especially near air-containing structures like the skull base. Thus, MAGNUS features all the desirable aspects for brain diffusion imaging: high G_max, high gradient field, high b-values (another factor reflecting gradient strength, timing, and durations), and substantial increase in signal to noise while minimizing artifacts.

DeMarco said that ultrahigh b-value diffusion MRI technology enables the measurement of effective axon radius maps. Diffusion-weighted MRI obtained at b = 1 ms/mm² captures a signal primarily from the extra-axonal water with some intra-axonal signal. Increasing b-values to 25 or 30 ms/mm² results in images of only the water trapped within the axons, which enables direct measurements of the axon diameter and other advanced axonal measurements. DeMarco stated that the MAGNUS technology makes such measurements possible and noted that three of the largest MRI manufacturers are exploring building a clinical product similar to the technology in the MAGNUS proof-of-concept.

DeMarco shared preliminary evidence suggesting that high-performance head-only MRI could be used to collect effective axonal radius measurements, and that these could be a specific marker of white matter abnormalities in individuals with chronic symptomatic mild TBI that researchers have previously been unable to obtain.

DeMarco and colleagues compared the results of DTI, DKI, and effective axon radius maps. Using the patient as their own internal control, researchers compared the white matter tract parcels on the right and left

___________________

⁵ Joint Program Committees consist of U.S. Department of Defense (DoD) and non-DoD medical and military technical experts who coordinate research and development efforts.

sides of the brain. DeMarco’s proof-of-concept study included four representative examples of 17 patients with chronic symptomatic mild TBI. These demonstrated lateralization differences in the effective axonal radius measurements in different white matter parcels in each chronic mild TBI patient, with these lateralization abnormalities appearing in different directions. No DTI or DKI lateralization in these parcels was seen in these same chronic mild TBI patients. This individual abnormal white matter parcel lateralization is consistent with the understanding that TBI affects different parts of the brain depending on the mechanism of injury.

Exploring Next Steps and Opportunities

DeMarco emphasized opportunities for dedicated ultrahigh-performance neuroimaging MRI scanners to contribute to advances in TBI diagnosis and care. DeMarco and colleagues plan to evaluate the benefit of combining high-performance imaging techniques to identify new imaging biomarkers in people with chronic symptoms after TBI. They also intend to conduct a similar analysis on people with acute TBI, which offers the additional advantage of collecting data on the same person over multiple time points, thereby decreasing biological variance and correlating symptomology over time. He underscored that ultrahigh b-value diffusion with effective axonal radius measurements has demonstrated sensitivity to pathology that neither DTI nor DKI have been able to achieve in the same population, and that larger studies are needed to fully characterize axon size distributions in people with mild TBI and healthy cohorts.

MULTIMODAL ELECTROPHYSIOLOGICAL BIOMARKERS

Leslie S. Prichep, chief scientific officer at BrainScope Company, described how electroencephalography (EEG) can be combined with AI and machine learning approaches to identify brain activity abnormalities, classify TBI patients into more precise pathophysiologic subtypes, and predict injury outcomes.⁶

Introduction to Quantitative Electroencephalography (qEEG)

Prichep explained that EEG, which measures brain electrical activity, is a test often used as part of diagnosing epilepsy and seizures. Quantita-

___________________

⁶ Prichep noted that she is an inventor of intellectual property licensed by BrainScope from NYU School of Medicine and that her presentation reports results of independent prospective FDA validation studies, to which BrainScope Company was blinded. The validation studies were part of DoD-funded research contracts. The views, opinions, and/or findings contained in the presentation are those of the authors and should not be construed as a position, policy, or decision of the funding sources.

tive EEG (qEEG) processes EEG waveforms to characterize the signal into features that can be used to describe brain activity and how it compares to normal, expected brain activity for different ages. In particular, qEEG allows the characterization of aspects of brain electrical activity not visible to the naked eye via visual inspection. The past decade has seen advances in the EEG landscape, including development of a handheld device that performs EEG, analyzes data, and provides feedback in real time. Advances in signal processing have enriched measures derived from the EEG and have enabled the creation of thousands of features that characterize the signal. Machine learning and AI can be used to develop algorithms to identify signal patterns indicating abnormalities that then become biomarkers.

She explained that EEG is uniquely sensitive to brain changes associated with traumatic structural and functional brain injury owing to the millisecond time base and to features that are particularly important in the algorithms developed in this area. These features include connectivity, complexity, and frequency distribution. Connectivity reflects disruption of neuronal transmission between brain regions, with measures including coherence, phase synchrony, phase lag, and asymmetries. Complexity of the signal reflects disorganization of the neural networks via fractal dimension, entropy, scale-free brain activity, and other indicators. Frequency distribution reflects changes in the neurochemistry, oxygen flow, glucose metabolism, and presence of edema (swelling) in the brain. The use of AI and machine learning approaches with qEEG features as inputs enables the development and identification of distinct profiles of pathophysiology that optimally separate groups of patients or predict outcome, Prichep stated.

TBI Classification Using qEEG Data

Multiple sources of information can be used as part of TBI diagnosis, classification, and treatment, and Prichep described EEG-based algorithms developed by BrainScope and cleared by the U.S. Food and Drug Administration (FDA) to explore the likelihood of brain bleed in patients who are CT positive for a TBI, CT negative, have a concussion, and do not have a concussion. She outlined the process of moving from EEG recordings to TBI classification, emphasizing the importance of having a high quality signal and addressing artifacts (signal features arising from the equipment itself). Otherwise, the artifacts will skew input to the algorithm, and the machine learning approach may separate patients with high levels of artifacts from those with low levels of artifacts rather than separating patients based on the intended biomarker metrics.

For signal quality, BrainScope uses a suite of eight artifact algorithms that run in real time, providing user feedback and stopping data collection in the event an artifact is present, Prichep said. After recording the EEG, the

process involves extracting qEEG features and transforming them to age-expected normal values using z-scores. This establishes the same dimensionality of standard deviation units to enable data to be combined statistically. The data reduction phase can involve extraction of 10,000 features. Prichep noted that inputting 1,000 patients and 10,000 features into a machine learning algorithm will result in excellent separation that cannot be prospectively validated. Additional modalities—such as blood biomarkers or neurocognitive findings—can also be incorporated at this stage if desired, Prichep stated. Entering the pool of features into the machine learning classification algorithm yields output in the form of a weighted combination of features for optimal separation.

Validation of the Concussion Index

The Concussion Index (CI) is a multimodal, AI-derived classifier function based on EEG features, said Prichep. The features are predominantly brain connectivity measures that characterize neuronal transmission regions as well as multimodal inputs, such as vestibular and procedural reaction time. A study of 1,577 sessions of data on 580 athlete patients from 10 clinical sites validated the CI (Bazarian et al., 2021). Establishing a threshold of concussion, the CI classifies patients with a CI score of less than or equal to 70 as likely concussed and those with CI scores above 70 as likely not concussed (see Figure 4-2). The study examined whether patients were above or below the threshold at three time points: day 0 (i.e., the first 72 hours of injury), clinical determination of return to play (RTP), and 45 days after RTP. Athletes without concussion served as controls, and they were consistently above the concussion threshold at all time points. The constancy of the control group across time enabled reliable interpretation of changes within the injured patients.

At day 0, a highly significant difference between the control and injured groups was evident. At RTP, 80 percent of injured subjects were within normal limits. Prichep underscored that this finding reveals that 20 percent of subjects who passed all RTP guidelines continued to report symptoms and show abnormal brain electrical activity. At 45 days after RTP, all participants were above the concussion threshold. She added that subjects with a rapid RTP of less than 14 days after injury had a higher CI value than those with a prolonged RTP of 14 or more days. Therefore, the CI score reflects severity and the potential to predict rate of recovery, she explained.

Prichep highlighted another study that indicated that those with rapid RTP have higher CI scores at day 5 than those with prolonged RTP, suggesting that change in CI score tracks recovery and that the slope of recovery is different between those who will have prolonged recovery and those who will have rapid recovery (Jacquin et al., 2021). Examining the relationship

between imaging (using DTI) and CI, a study found a significant relationship between several DTI measures and CI in the same patients with concussion in which more abnormal DTI measures are associated with lower CI scores, said Prichep (Wilde et al., 2019). This finding suggests that CI reflects changes in white matter integrity. Prichep noted that BrainScope’s device only collects data from the frontal and frontal temporal regions of the brain, and this study indicates that this data collection method does not limit reflection of abnormalities found using DTI.

EEG-Based Phenotypes

Prichep outlined BrainScope’s efforts to identify EEG-based TBI phenotypes. Collecting data on 771 individuals with concussion diagnoses from studies conducted by BrainScope in recent years, researchers used a machine learning approach to look for subtypes of brain injury pathophysiology among people who shared a clinical profile (Armañanzas et al., 2024). They identified five electrophysiological subtypes according to signal characteristics that provide information about gray matter, white matter, the gray–white interface, and underlying pathophysiology. Measures of power, connectivity, and complexity result in different compositions for each of the five subtypes.

Prichep noted that this is the first demonstration of electrophysiological subtypes within the heterogeneous group of patients with concussion and offers potential for advancing understanding of underlying pathophysiology. Given that these subtypes come from data taken within 72 hours of injury, the ability to personalize diagnosis carries implications for treatment recommendations and associated outcomes. She described how a search of available clinical information did not reveal clear correlates between symptoms after concussion and underlying subtypes; symptoms were often present in multiple subtypes. However, subtype predicted time required to attain RTP status with high accuracy. Thus, within 72 hours of injury, the identification of subtype is a predictor of recovery duration, said Prichep.

EEG Changes Related to Blast Exposure

The Investigating Training-Associated Blast Pathology (INVICTA) study of the effects of blast exposure involved multimodal evaluation that included EEG obtained using BrainScope devices, Prichep noted (Roy et al., 2022). Radar plots illustrate different brain regions, with spokes representing complexity of the EEG signal in each region. She described findings from EEG complexity measures of the dorsolateral prefrontal cortex, a region associated with cognitive and executive function. Higher complexity represents higher functioning. Researchers plotted complexity measures at various time points in relation to blast exposure: preexposure baseline, 6 hours postblast, 24 hours postblast, and 2 weeks postblast.

Radar plotting revealed that immediately after blast exposure, the complexity of every region of the dorsolateral prefrontal cortex decreases significantly. Slight increases in complexity occur at 24 hours postblast. At the 2-week time point, complexity returns to levels not significantly different from baseline. In unexposed control subjects, complexity remains at similar levels across time points. This signifies that a set of measures of EEG signal complexity is highly correlated with blast exposure and normalizes over time, Prichep explained, indicating the possibility of developing an EEG-based biomarker for evaluation of subconcussive blast exposure.

Reflections

Outlining conclusions from this EEG research, Prichep emphasized that qEEG is a tool to quantify brain electrical activity and that qEEG features can reflect changes in neuronal transmission, integrity of neural networks, and neurochemistry, such as edema (swelling). Electrophysiological biomarkers can serve as decision support tools and aid in initial diagnosis, prediction of rate of recovery, and quantitative tracking of change over time.

She stated that machine learning and AI models in populations of patients with mild TBI have identified patterns of qEEG features that serve as TBI biomarkers and noted that the CI has been cleared by FDA for indicating the likelihood of concussive brain injury. Stability of CI scores over time in noninjured controls allows reliable interpretation of change over time in the injured group, she said. She underscored the need to explore the power of integrating various biomarkers to determine if using multiple biomarkers improves the accuracy of predictions after TBI. Prichep described subtyping as an interesting approach to identifying phenotypes that may lead to faster, more personalized treatment planning and potentially to better outcomes.

CONCUSSION DIAGNOSTICS WITH EYE TRACKING

Uzma Samadani, founder of Oculogica and neurosurgeon at the Minneapolis Veterans Administration Medical Center, outlined use of the EyeBOX eye-tracking test and the benefits of using eye tracking as a physiologic measure of brain function to assess brain injury.⁷ She noted that eye movements have been considered a test for brain injury for approximately 3,000 years. Prior to the invention of radiology and head CT, brain injury was defined by its symptoms, and one of the most obvious symptoms was dysfunctional ocular motility.

Cleared by FDA in 2018, the EyeBOX Test measures pupil position while an individual watches a video. She noted that the video content is not relevant; patients may choose to watch music videos, sports, or any other preferred genre. A video window moves to different locations of the screen, and a small camera measures pupil size and position as the eyes follow the video window. EyeBOX takes measurements at 500 hertz (i.e., 500 times per second), and the noninvasive test collects x- and y-coordinates of the pupils and size measurements for 220 seconds. EyeBOX does not require a blood draw or centrifuge, the results are instant, and a preinjury baseline is not required because most functions of ocular motility are relatively well preserved, Samadani explained.

___________________

⁷ Samadani noted that she has intellectual property related to concussion and brain injury assessment, to assessment of dementia after brain injury, and to treatment of intracranial hemorrhage. She also indicated that she has grant funding, salary/employment, consulting fee, honoraria, or equity from Abbott Diagnostic Laboratories; Continuing Legal Education in MN and NY; Hennepin County Medical Center; Hennepin Health Foundation; Integra Corporation; Islamic Medical Association of North America; Medtronic Corp; Minnesota Brain Injury Alliance; Minnesota, Texas, Louisiana, Wisconsin, Wyoming High School Coaches Association; National Football League; National Neurotrauma Society; North American Brain Injury Society; Oculogica Inc; Steven and Alexandra Cohen Foundation for Veteran Post-Traumatic Stress and Traumatic Brain Injury; Department of Veterans Affairs; and USA Football.

Samadani outlined the benefits of using eye tracking to assess brain injury via cranial nerve function. The cranial nerves are highly sensitive to injury and have a large catchment area inside the brain. A TBI that damages areas controlling relevant cranial nerves can be monitored and measured through features such as pupil size, pupil reactivity, and pupil position. In a healthy individual, cranial nerves reflexively cause pupils to constrict in higher light levels and both eyes move together. Approximately 97 percent of people having pupils equal in size at baseline, making anisocoria (having pupils of different sizes) relatively rare. Additionally, the majority of people have eyes that move together. In cases of amblyopia in which muscle weakness prevents the eyes from moving together, test results are distinct from those associated with brain injury, she said.

As people age, ciliary muscles become weaker and pupils are not as effective at constricting, but the ability of eyes to move together without conscious effort remains relatively well preserved (supranuclear control of eye movements). Different pathways coordinate eye coordination, vertical movement up, vertical movement down, and horizontal movements. These eye movements are coordinated in the brain stem, and when a person has an impairment, it manifests itself in particular ways depending on which pathway is impacted. Examining ocular motility can thus enable identification of the affected pathway, and this was the rationale for developing EyeBOX, Samadani explained.

Eye-Tracking Research

Samadani and colleagues conducted numerous studies exploring the usefulness of eye tracking, including research demonstrating that issues with the third and sixth cranial nerves are detectable through eye tracking (Samadani et al., 2015a). Another study demonstrated sensitivity and specificity of eye-movement tracing in an ED population (Samadani et al., 2016). Research indicated that eye tracking detects disconjugate eye movements associated with structural injury from TBI (Samadani et al., 2015b). Other studies examined elevated intracranial pressure and reversible eye-tracking changes (Kolecki et al., 2018), eye-tracking ramifications on TBI diagnosis and classification (Samadani, 2016), and the ability of eye tracking to measure central nervous system physiology (Samadani, 2015). Noting the importance of unbiased evidence, Samadani emphasized that to her knowledge, Oculogica provides devices to collaborating researchers but has never made a direct payment to an investigator. Rather, Oculogica issues payment to an institution to compensate for an investigator’s time, but this is not passed on to the investigator and they do not receive compensation beyond their regular salary. She also highlighted the existence of studies conducted by researchers not affiliated with Oculogica. For example, research con-

ducted by nonaffiliates has explored the reliability of objective eye-tracking measures among healthy controls (Howell et al., 2020).

Physiologic Measures of the Brain

EyeBOX obtained FDA clearance in 2018 as a mechanism for assessing central nervous system (CNS) function, said Samadani. Research submitted to FDA in support of this clearance indicated an intended use population aged 5–67 years and included testing performed in ED and non-ED settings (Samadani et al., 2022). The study used two subsets of the Sport Concussion Assessment Tool 3, the Symptom Severity Scale Score, and the Standardized Assessment of Cognition.

Samadani described that for millennia, physical and psychological exams were the only methods of assessing CNS function. In more recent times, imaging, EEG, qEEG, and other tests became available. Angiography provides information about vascularity and the brain’s appearance. Other developments include methods of assessing the integrity of systems, advanced imaging such as 3T MRI, and the ability to assess cranial pressure and brain oxygenation. Samadani remarked that uncalibrated or non-spatially calibrated eye-movement tracking is a new type of technology most closely related to original physiology assessments. Eye tracking is not intended to replace CT or MRI imaging, Samadani noted, but is intended as an automatable, objective, nonrisky, nonradiation exposing, noninvasive physiologic method that is agnostic to language, culture, and education.

Improving Classification of TBI

Samadani discussed the value of TBI diagnostics as part of refined classification approaches and their connections to early intervention and therapeutic development. Drawing a comparison between cardiac and TBI care, Samadani highlighted how the advent of the cardiac biomarker troponin changed ED treatment of chest pain. The presence of troponin enables numerous causes of chest pain to be ruled out simultaneously and hastens performance of an electrocardiogram or cardiac catheterization. In contrast, patients with head injury generally receive a physical exam and a head CT scan before being diagnosed with TBI, while the multiple pathophysiologies that can affect a person’s clinical outcome are not fully classified, said Samadani. Differentiation is not made between axonal disruption, microhemorrhages, epidural hematoma, or subdural hematoma, she noted. Instead, pathophysiologies are grouped together, both for clinical treatment and for research toward developing therapeutics. She remarked that the Glasgow Outcome Scale (GOS) is limited to placing patients within eight categories, yet it is the most widely used TBI outcome measure available.

Samadani argued that a collective TBI community should highlight the need for better hierarchical and multimodal classification tools for brain injury, and that multiple metrics should be used to determine what is wrong with the patient, including eye tracking, algorithmic analysis of CT and 3T MRI, blood-based biomarkers, and genetics. Samadani described genetics as an elephant in the room in that it plays a role in outcome after brain injury, yet is not often accounted for. Building all these components into algorithms that classify the nature of brain injury could improve outcomes, she noted. Recognizing TBI soon after injury and classifying physiology could disrupt the undesirable consequences of undiagnosed brain injury and facilitate the ability of patients to receive early intervention, she said. She remarked that a patient with chest pain would not be sent home after a negative result on one test for heart attack, yet this trajectory can happen for patients with TBI, who may have symptoms discounted owing to negative head CT scans. Improved diagnostics, whether in the ED or at a followup appointment with a specialist, could improve outcomes for more people.

Financial Considerations

Samadani also raised issues related to reimbursement for diagnostic devices. In 2019, EyeBOX obtained a Category 3 current procedural terminology (CPT) code as a concussion diagnostic from the American Medical Association (AMA), she said. The Centers for Medicare & Medicaid Services (CMS) also approved the device. She underscored the steps involved, which included submitting papers, providing data, and demonstrating effectiveness in the intended population, and remarked that the single greatest barrier to wider incorporation of new assessment technologies such as eye tracking is reimbursement.

Samadani reflected on the role of CPT codes in the reimbursement system. In the absence of a CPT code, health providers must use an unlisted code. Payers may argue that the test is not validated for someone as complicated as the patient or that the patient’s level of functioning is too high for the test’s capability, and therefore the insurer may prefer use of a CT scan and refuse coverage for the eye tracking test. Although some large payers reimburse for eye tracking, such as Blue Cross Blue Shield and UnitedHealthcare, she said, not all insurance companies currently reimburse for this service. Insurance companies want indication that use of a diagnostic device changes disease management before they will reimburse for it on a regular basis, she commented. Furthermore, lack of reimbursement could discourage investment in the diagnostic space, she contended. When a TBI diagnostic fails, it can discourage investors from entering the market. She remarked that this has happened to some extent in the area of dementia diagnostics and could happen in TBI.

Samadani described a chicken-and-egg dynamic, in which the inability to precisely and accurately classify TBI impedes the development and availability of new therapeutics. Heterogeneity causes trials for TBI therapeutics to fail, she said, while diagnostics are needed to help identify and refine injury subtypes to better address this TBI heterogeneity. Thus, having better diagnostics fosters the development of better therapeutics, but making reimbursement for diagnostics contingent on their ability to change management creates a chicken-and-egg barrier, she said. Advocacy efforts are needed to shift insurance policy to expand reimbursement for TBI diagnostics, Samadani maintained.

DISCUSSION

Reimbursement Considerations

Leslie Wise, chief executive officer at EvidenceMatters, clarified that a CPT code is the numerical representation of a provided service and is not a reimbursement code. Professional societies generally will not support CPT codes outside of their specialties, she noted. The Medicare, Medicaid, and SCHIP Balanced Budget Refinement Act of 1999 requires that the valuation of a new code be accompanied by retiring an existing code or by having all codes within the relevant specialty revalued for a lower reimbursement.⁸ Wise described obtaining a Level 1 CPT code as a strategic process that often requires 5 years and involves five publications and an outcome study at Level 2A or higher. Organizations can assist with this time-intensive process. Wise explained that the entity obtaining the code sets the amount for which the corresponding service will be sold; CMS does not set the amount. She added that the process can be political, and that an organization’s lack of support for a new code may not be indicative of a lack of enthusiasm for the service, but rather reflective of resistance to having their existing codes revalued. Every time a new code is added within a specialty—in the absence of removing a current code—reimbursement amounts for existing codes decrease after revaluation in compliance with the Balanced Budget Act, said Wise.

Specialties such as radiation and MRI-guided CT have previously collaborated to establish a group to consider components of a particular disease state, Wise remarked. This approach could be used to systematically structure and publish an argument for needed evolution within the TBI space that addresses various pieces of the larger context, including imaging, EEG, and eye tests, she suggested. Wise stated that this approach helps pay-

___________________

⁸ Medicare, Medicaid, and SCHIP Balanced Budget Refinement Act of 1999, Public Law 106-113, 106th Cong., 1st sess. (November 29, 1999).

ers develop an understanding of what is needed within a field and has the potential to lead to coverage for new services. She argued that randomized controlled trials alone are insufficient for influencing coverage of extremely expensive technologies such as imaging machines; instead, payers respond to a framework that organizes the research and illustrates trends and possibilities of the future space. Breaking down TBI phenotypes and correlating these with types of injury or other characteristics could be effective in helping payers understand the relationships between injuries and outcomes, she contended. Wise maintained that such collaboration within the TBI field could advance the integration of emerging science into practice.

Patient Engagement

Cynthia Grossman, subject-matter expert in the science of patient engagement and real-world evidence, emphasized the heterogeneity in TBI patient populations, the substantial effects TBI can have on people, and the role of comorbidities. She remarked that although heterogeneity can pose a challenge in the patient engagement space, it also draws more stakeholders. Grossman suggested that consideration be given to engaging patient communities in the innovation journey by communicating directly about TBI innovations to illustrate stories behind the advances and their potential benefits. Corinne Peek-Asa, vice chancellor for research at the University of California San Diego, asked about the most appropriate time frame for engaging patients and how to drive advances toward addressing the symptoms that patients want treated. She commented that the ability to identify disruptions among brain connections and damage to brain white matter is exciting, for example, but that most patients are likely focused on personal ramifications such as why they have a headache, lack balance, or cannot remember where their keys are.

Samadani replied that many of the patients evaluated with eye tracking had prior interactions with the health care system for their TBI. For instance, several patients went to the ED after their injuries and were told that they did not need a head CT or that the CT was negative, and they should go home and rest. Days, weeks, or even months later, they continued experiencing problematic symptoms. She remarked that some patients break into tears after having TBI identified via eye tracking, communicating that no one had previously believed them. This experience can be validating and empowering, encouraging patients to again seek treatment. Therapeutics or interventions may be available to help address symptoms, but before a patient can access them, a provider must agree with the patient that an injury is present, said Samadani. Noting increased attention on medical gaslighting (i.e., when a provider invalidates or ignores a patient’s concerns), she described that some genders and races are associated with

medical gaslighting more frequently than others. Objective measures are a mechanism for fostering equity in health care, Samadani maintained.

Prichep recalled experiences in TBI clinics in which high school students with persistent symptoms from sports injuries—and their parents—expressed relief at being identified with a numerical value indicating that something was wrong after a CT did not reveal abnormalities. Moreover, because some students are eager to return to playing sports but their parents do not allow it, a quantitative measure that provides a numerical value and associated information would be helpful. Prichep stated that such a measure from diagnostic technology may not reflect all features of the abnormality, but that the information is generally well received by patients and caregivers. DeMarco remarked that in working with an active-duty patient population, he was frequently asked to show patients abnormalities on their images, and agreed that patients want to see external evidence of their experience. He noted that within his other research area of carotid plaque, patients are also eager for engagement and involvement.

Ferguson noted that the recent TBI classification conference convened by the National Institutes of Health (NIH) opened with people sharing their lived experience, indicating steps toward greater patient engagement.⁹ He noted that a unique aspect of clinical prediction models and machine learning is that the expertise for these tools resides outside of the medical field, including in fields such as investment banking. A subset of professionals in those fields will experience TBI, and thus increased engagement with other fields also offers a useful opportunity to drive research, he said.

Highlighting stroke care as an example of implementing advances in the field, DeMarco described that a decade of research identified how to diagnose and treat acute stroke, and these practices are now widely available in any ED. Furthermore, patients in need of more advanced care are now able to be identified and transferred to comprehensive stroke services. Providers are in need of effective instruction on diagnosing TBI, and the progress made in stroke care offers insights on advancing TBI care. Once TBI can be diagnosed in a reproducible, consistent manner across EDs, consideration should be given to methods of identifying TBI patients with more complex needs and to the creation of comprehensive care centers that offer the various modalities discussed at this workshop, said DeMarco. He contended that improvements in prognostication and enrolling patients could also lead to more effective therapies.

___________________

⁹ Video from the Traumatic Brain Injury Classification and Nomenclature Workshop, convened by NINDS in January 2024 is available from https://videocast.nih.gov/watch=54118 (accessed July 9, 2024).

Provider Education

Adell stated her excitement at learning of new TBI diagnostic methods, and she noted that athletes and military personnel seem to be the focus of much TBI research because the cause of injury is more predictable. She remarked that people like her who experience TBI from a range of unpredictable causes of injury seem to receive less attention. Highlighting the need for increased education efforts for providers in the ED and in primary care, she observed that when providers have difficulty diagnosing TBI, they tend to dismiss patient concerns as emotional or mental health issues. Recalling the physical manifestations she experienced, she described light sensitivity that prevented her from looking at a computer or watching television, noise sensitivity, and lack of balance. She was certain something was wrong with her, but without the help of a physician friend able to locate a care pathway for her, she would have had to suffer with her symptoms and assume nothing could be done to improve her outcome. Adell said that the workshop has brought the complexity of insurance and reimbursement issues to her attention.

Samadani noted that she previously was on the faculty at New York University, during which time she taught the only course offered on brain injury during the 4-year medical school program of study. Most people educated in brain injury are residents in neurosurgery or physical medicine and rehabilitation, as these fields often treat patients with brain injury, she said. Additionally, specialties such as emergency medicine, neurology, and pediatrics have rotations involving patients with brain injury. However, many medical students proceed through 4 years of medical school without attending a lecture on TBI or seeing a patient with brain injury. Samadani recalled a paper indicating that many physicians receive TBI education through the media, highlighting a gap in education and an opportunity for substantial improvement.

Introducing technologies in the U.S. Department of Veterans Affairs (VA) is a method of simultaneously gaining traction for innovations and educating providers, suggested Wise. Given that VA does not face the same insurance reimbursement issues that the general health care system contends with, technology uptake can involve fewer barriers. Because VA is able to purchase technology deemed beneficial, it can more readily use new innovations than the broader health care system. Moreover, 70 percent of the nation’s physicians train in the VA system at some point in their careers, and doctors tend to carry their training into practice, she remarked. While VA physicians become familiar with an innovation, a company can continue working through the process of attaining a CPT code and getting a reimbursement structure in place. When physicians move from the VA system into regular practice, they will continue to purchase the technol-

ogy, she maintained. While in the orthopedic industry, the company Wise worked for achieved large market share by introducing technologies to the VA system.

Maheen Mausoof Adamson, professor at the Stanford School of Medicine and senior scientist for rehabilitation services at the VA Palo Alto Healthcare System, stated that in working at a VA facility for nearly 2 decades, she has seen the role that the department can play in adopting innovation into practice. For instance, VA is collaborating with several virtual reality companies and has a simulation network. She remarked that some companies are reluctant to approach VA because of the lengthy wait times that can be involved in establishing a connection with the department. She described how VA centers of excellence feature three prongs—research, clinical practice, and education—and highlighted the importance of translating research into clinical practice that is then disseminated through continuing medical education. This model could be applied to the TBI field, she suggested, noting that education products can vary widely, from guidelines to TikTok reels.

Joel Scholten, director of Physical Medicine and Rehabilitation at the Department of Veterans Affairs, noted that the patient population at the VA differs from the private sector in that most individuals with TBI experienced the incident event at least a year prior during deployment or active duty. He remarked that the VA welcomes innovation and the integration of new technology but requires evidence that it is of benefit for long-term effects and chronic conditions. Additionally, the VA has a robust research infrastructure that can aid in the identification of an innovation’s long-term signal that pertains to individuals with a chronic condition.

Potential Drawbacks in Machine Learning and Artificial Intelligence

Darío Villarreal, head of science and technology at Toyota Way Forward Fund, highlighted that eye-tracking research explores cause and effect from the physiological perspective to understand why pupils dilate. He asked about the implications of the black box in machine learning, in which data entered into an algorithm are used to create models without humans knowing how the variables were combined to make predictions. Remarking on the difficulty of determining whether a machine learning result is correct or incorrect in the absence of understanding cause and effect, he emphasized that bias could enter the data and lead to unethical results. Villarreal asked how the results of a black box can be confirmed and determined to be free from bias against certain populations.

Ferguson replied that the danger of overfit is a classic problem in machine learning. For example, results may fit the 15 patients studied but

not fit any future patients. He noted that researchers often cross-validate machine learning models, as was the case with IMPACT and CRASH. Advances toward interpretable machine learning tools are being made (Tritt et al., 2023), such as modern attention networks and expert-augmented machine learning that solicits clinician feedback to ensure that the results do not seem improbable or harmful to patients. The clinician feedback is then used as input to the machine learning to improve the system. Ferguson remarked that guardrails for machine learning are being developed, and that caution is warranted when implementing machine learning results.

Prichep echoed that attention to black boxes is needed to protect the community into which tools and algorithms are being introduced. She described herself as extremely cautious about overtraining in machine learning and noted that data reduction, internal cross-validations, and independent validations after algorithm finalization are methods BrainScope uses to avoid incorrect results. In cases where expected accuracy cannot be demonstrated in a population new to the algorithm, more work remains before results are implemented. Prichep emphasized that the providers who will be using the results and the patients they treat deserve interpretable outputs that have been carefully considered, and the onus for interpretable results is on model developers.

A participant asked about the power of AI tools to extract several different features from the EEG signal and how to differentiate between the most realistic features versus those lacking clear rationale. Prichep stated that some methods of interacting with AI avoid treating it as a black box. These include initial steps that could be considered heuristic, such as including features demonstrated in literature to be important in the area and omitting any variables that have not been shown to be replicable within more than one EEG sample within a population. She remarked that if a variable is not replicable within a person, she is careful to omit it when developing an algorithm for a population. Developing the pool of features entered into the algorithm aids in limiting results with no rationale.

She emphasized that in using multivariate machine learning models, the single variables that reflect the most significant differences between two groups are not necessarily the best predictors for the population. Thus, including only the most significant features could preclude identification of the best predictors. Prichep remarked that the beauty of multivariate models is their ability to address heterogeneity in capturing the variety of ways that individuals may be abnormal. Ultimately, AI tools generate a summation of a weighted set of features, and the same score with the same features can be generated via different weightings. Prichep stated that researchers can build protection against results lacking reason into their methods of using AI.

Emergency Medical Services and Machine Learning

Jeremy Kinsman, emergency medical services (EMS) specialist at the National Highway Traffic Safety Administration (NHTSA), asked whether vehicle crash data or EMS prehospital care data have been incorporated into machine learning approaches. If not, Kinsman asked whether NHTSA could collect any crash or prehospital assessment and care data that would fill gaps in machine learning models, particularly as NHTSA considers clinical decision support tools in TBI classification. Ferguson noted that some of the papers referenced in his presentation included information about mechanism of injury. Remarking that the medical records in Level 1 trauma centers often collect mechanism of injury, he stated that the incorporation of data into machine learning approaches may depend on the type of vehicle crash information the medical record can receive. Biomechanically detailed information would likely be beneficial when using machine learning tools, said Ferguson.

Pathophysiologic Endophenotyping

Jeffrey Bazarian, professor at the University of Rochester, described the ability to identify and target treatment to pathophysiologic endophenotypes¹⁰ as the holy grail in TBI research. He asked to what extent technologies presented today could be used in identifying those endophenotypes and in understanding the best treatments for specific endophenotypes. Bazarian also queried whether AI has a role in this process. Ferguson replied that the identification of multidimensional clusters is a potential step toward the identification of a subgroup of individuals with unique multidimensional features that call for specific interventions. Prichep echoed that the identification of subtypes is a step toward the goal of establishing pathophysiologic endophenotypes because subtypes suggest different underlying pathophysiologies, which in turn suggest various treatment interventions. She remarked that the availability of more robust outcome data would expand the ability to identify subtypes.

BrainScope’s outcome data are related specifically to the length of time until individuals were RTP cleared and 45 days beyond. Their outcome data do not include the evolution of different clinical sequela or responses to various treatments. Ultimately, the ability to input the response to treatment into the initial model to determine its relationship with the five subtypes could aid in identifying the subtypes that respond well to specific treatments. She stated that such information could then be integrated into

___________________

¹⁰ The term endophenotype refers to a subgroup with measurable characteristics or indicators that more precisely characterize different types of TBI.

modeling in a manner that is not possible until more data are available. Currently, the first steps are in place to begin the research needed to make this possible, said Prichep. DeMarco added that current MRI data are useful but only in terms of groupwide analysis. The opportunity to work with more advanced systems carries potential to identify changes in an individual patient that could then be targeted with specific treatments, DeMarco noted.

Michel Baudry, founder and chief scientific advisor at NeurAegis and professor at the Western University of Health Sciences, asked about the use of different technologies in the identification of TBI subtypes and whether eye tracking, EEG, and high-performance MRI identify similar subtypes of patients, reflecting underlying molecular or physiological mechanisms. Samadani replied that eye-tracking technology reveals changes that are more consistent with acute TBI. Approximately two-thirds of patients with concussive brain injury will have eye-tracking abnormalities consistent with altered intracranial pressure or supratentorial mass effect; these abnormalities resolve within 4–6 weeks for 75 percent of this patient group. A small percentage will develop abnormal eye-tracking metrics that are more consistent with accommodative disorders or vergence dysfunction and may have persistent symptoms for a longer period of time. Samadani noted that a publication on characterizing subtypes is in preparation, and that Oculogica is also currently applying to FDA to expand EyeBOX’s current clearance for injury within 2 weeks to an indefinite time period. Other measures (levels of different blood biomarkers or imaging metrics, for example) may be of most use as indicators for other types or severities of TBI. Discussion did not fully clarify whether patient subtypes currently being characterized by different technologies reflect common underlying mechanisms, and more work remains.

Multimodal Testing Data Considerations

Frederick Korley, professor at the University of Michigan, commented that different diagnostic neuromonitoring technologies all have strengths and highlighted the potential value of combining them in an effort to maximize predictive value. He reflected that one challenge, to his knowledge, is that no datasets exist that measured all various inputs on the same individual. Given this data missingness, he asked whether machine learning approaches can combine these different data or whether a dataset containing complete information on the same individuals is required. He also asked if the Forum on Traumatic Brain Injury can encourage awareness and efforts toward collecting such multimodal data to assess how different modalities work together.

Ferguson replied that recent machine learning tools are capable of accommodating missingness. For example, a mixture model framework specifically measures the pattern of missingness and then creates new features from the missing data pattern (Kaplan et al., 2022). Such models harness the patchwork of missingness that occurs across a population to improve prediction. Ferguson added that any forum efforts to generate more complete datasets would be beneficial. Prichep remarked on the difficulty of securing funding to conduct testing of various modalities on the same individuals. The INVICTA study has numerous touch points on each individual soldier, but it focuses on blast exposure, not on blast TBI, and is a different subtype of the general group. She noted that one of INVICTA’s aims is to take measurements from blood biomarkers, EEG, virtual reality, and eye tracking and input them into one model. Prichep stated that the forum could consider developing a method to share data in one place for one population and proceed from there.

LaPlaca asked about current and potential data sharing, noting the challenges that are often involved in data sharing within the private sector. She also noted that reimbursement issues could pose a barrier to multimodal testing and asked how this might be addressed. Samadani remarked that data from any clinical tests could be saved and used in a post hoc analysis, a process that creates a real-world experience user database that is reviewed retrospectively. A prospective study would require substantial funding, given that research subjects could experience test fatigue, which introduces ascertainment bias, she said. Adding that participants in research studies are not always representative of the general population, Samadani highlighted the value of TRACK-TBI and its fairly representative patient population. Yet, significant ascertainment bias remains present in TRACK-TBI because its participants had the means to return for multiple follow-up visits, and significant segments of the population lacking the resources or time for return visits are not included in the study, Samadani pointed out.

Ferguson described that large movement disorder databases have been assembled from accelerometers—such as wearable devices, including smart-phones that many people currently opt to carry on their person throughout the day—that are then used in algorithms that aid in Parkinson’s disease diagnosis. Ferguson stated that wearables could be a data source to address TBI data gaps. Prichep highlighted the possibility that certain measures could potentially act as surrogates for others; should this prove to be the case, obtaining data on all five measures would be unnecessary. Individuals without a certain measure could perhaps be represented by another highly correlated measure. Prichep specified that such an approach is better suited to research than to clinical implementation, but it could serve as a starting place. Stuart Hoffman, senior health science officer for TBI for the Office

of Research and Development at the VA, stated that a law requires the VA to participate in a mental health process that includes TBI. This involves harmonization of two large datasets, both of which are ongoing, longitudinal datasets primarily focused on TBI but also containing mental health disorders related and not related to TBI. The goal of this effort is to integrate multimodal biomarkers after they have been harmonized between the two datasets, which together include longitudinal data on approximately 4,000 individuals.

LaPlaca remarked that many additional technologies are under development, and those presented today represent advances in innovation and highlight the need for multimodal assessments from the research level to clinical use. She emphasized the value of patient engagement and the early consideration of aspects related to FDA clearance and reimbursement processes.

This page intentionally left blank.