Neil Samuel Nyholm Graham

Large language models (LLMs) are increasingly evaluated for clinical tasks, but safe use requires reliable uncertainty estimates as well as high accuracy. We tested eight general-purpose LLMs on 100 validated neurology board-style multiple-choice questions and compared three confidence signals: self-reported confidence, critique-implied confidence from an independent model instance, and consistency-based confidence from repeated sampling. Accuracy varied from 45% (Llama 3.1 8B) to 92% (Gemini 3 Flash/Pro), but confidence was generally overestimated, particularly in lower-accuracy models. Consistency-based confidence provided the best discrimination between correct and incorrect answers (pooled AUROC 0.74), outperforming self-report, although calibration differences between methods were not significant. Calibration should therefore be reported alongside accuracy when clinical LLMs are evaluated. Short Description: We evaluated the accuracy and confidence of eight large language models on 100 neurology board-style questions. We found substantial overconfidence, and differences between confidence methods in their ability to distinguish correct from incorrect answers.

Background: Military personnel are at increased risk of traumatic brain injury (TBI). However, the late effects of injury on cognition, psychiatric symptoms and function remain unclear. The UK ADVANCE cohort, comprising 1144 male service personnel deployed to Afghanistan, of whom half were exposed to major combat-related physical trauma, was established to clarify long-term outcomes.Methods: Assessments performed on 1018/1144 participants a median of 13·7 years (IQR 12·5-15·0) post-injury included TBI ascertainment, neuropsychological testing, psychiatric and functional outcome questionnaires, and advanced brain MRI. Outcomes were related to injury severity, with voxel-based morphometry and diffusion tensor imaging quantifying brain morphometry and microstructure. Findings: Among 514 trauma-exposed participants, 306 (60%) sustained a battlefield TBI. 81 (15·8%) experienced moderate-severe (definite), 144 (28·0%) mild (probable), and 81 (15·8%) symptomatic (possible) TBI. Compared with Uninjured personnel (n=504), moderate-severe TBI was associated with multidomain cognitive deficits, including executive function (CWIT Inhibition/Switching: B=-5·44, 95% CI -8·86 to -2·07), attention (RBANS: B=-8·11, 95% CI -11·92 to -4·29), and delayed memory (RBANS: B=-4·0, 95% CI -7·42 to -0·91). Psychiatric symptoms and perceived health differed across injured groups. Unemployment risk was elevated across all injured groups and greatest after moderate-severe TBI (RR 10·74, 95% CI 5·86-19·68). Regional grey and white matter volume differed by injury severity, with modest differences in fractional anisotropy following moderate-severe TBI.Interpretation: TBI was common among combat trauma-exposed personnel. Moderate-severe TBI was associated with multidomain cognitive deficits, the largest unemployment risk, and structural brain changes persisting into the chronic phase post-injury. This highlights the need for targeted rehabilitation and long-term follow-up in service personnel.

Background: Musculoskeletal conditions are a leading global cause of disability, yet the factors influencing long-term musculoskeletal health, particularly following trauma, remain incompletely understood. Machine learning could be applied to identify previously unknown patterns in large-scale, multimodal datasets. Objective: This study aims to test the ability of a new sparse group factor analysis method to uncover hidden patterns in large-scale multimodal datasets and generate testable, clinically relevant hypotheses. Methods: This study applies sparse group factor analysis, a hierarchical unsupervised machine learning method, to the Armed Services Trauma and Rehabilitation Outcome (ADVANCE) cohort to identify latent structures in multimodal clinical data. ADVANCE is a prospective longitudinal dataset of 1145 UK military personnel and veterans who served in Afghanistan. Half the cohort sustained combat injuries, and the remainder were frequency matched on deployment, service, rank, role, age, and ethnicity. Study 1 validated the approach by rediscovering known group-level patterns between combat-injured and noninjured participants, including poorer outcomes in pain, mobility, and bone health among those with lower limb loss. Study 2 explored the injured, nonamputee subgroup without prespecified labels to identify new hypothesis-generating clusters that could subsequently be tested using standard hypothesis-testing methods. Results: The ADVANCE cohort was 34.1 (SD 5.4) years old and 8.3 (SD 2.1) years postinjury or 7.7 (SD 1.9) years since matched deployment. A subgroup of 125 individuals with worse musculoskeletal outcomes was uncovered. This group had greater body mass (mean 92.6, SD 14.7 kg vs mean 88.0, SD 13.4 kg; P =.002), higher injury severity (median 12, IQR 5-22 vs median 9, IQR 4-14; P =.002), and reduced health-related quality of life with head injury. These findings led to a novel hypothesis that head injury, including potential traumatic brain injury, is associated with long-term musculoskeletal deterioration. This hypothesis is supported by literature in both athletic and military populations and will be tested in follow-up analyses. Conclusions: Our findings demonstrate how sparse group factor analysis, combined with clinical insight, can uncover hidden patterns in large-scale datasets and generate testable, clinically relevant hypotheses that inform prevention, treatment, and rehabilitation strategies.

BACKGROUND Musculoskeletal conditions are a leading global cause of disability, yet the factors influencing long-term musculoskeletal health, particularly following trauma, remain incompletely understood. Machine learning could be applied to identify previously unknown patterns in large-scale multimodal datasets. OBJECTIVE Test the ability of a new sparse Group Factor Analysis method to uncover hidden patterns in large-scale multi-modal datasets and generate testable, clinically relevant hypotheses. METHODS This study applies sparse Group Factor Analysis, a hierarchical unsupervised machine learning method, to the ADVANCE cohort—a longitudinal dataset of 1445 UK Afghanistan War servicemen—to identify latent structures in multimodal clinical data. Study 1 validated the approach by rediscovering known group-level patterns between combat-injured and non-injured participants, including poorer outcomes in pain, mobility, and bone health among those with lower limb loss. Study 2 explored the Injured, non-amputee subgroup without prespecified labels to identify new hypothesis-generating clusters that could subsequently be tested using standard hypothesis testing methods. RESULTS A subgroup of 125 individuals with worse musculoskeletal outcomes was uncovered. This group had greater body mass, higher injury severity, and a higher prevalence of head injury. These findings led to a novel hypothesis: that head injury, including potential traumatic brain injury, is associated with long-term musculoskeletal deterioration. This hypothesis is supported by literature in both athletic and military populations and will be tested in follow-up analyses. CONCLUSIONS Our findings demonstrate how sparse Group Factor Analysis, combined with clinical insight, can uncover hidden patterns in large-scale datasets and generate testable, clinically relevant hypotheses that inform prevention, treatment, and rehabilitation strategies.

There is widespread concern among former athletes about the link between head injury and dementia. Neurologists are increasingly assessing ex-contact sports athletes with cognitive and behavioural issues following repetitive head impacts and traumatic brain injury. Their assessment and management can be challenging due to the broad differential diagnosis, including psychiatric issues, trauma-related impairment and, in some cases, neurodegeneration. There may be a range of pathologies present after trauma exposure, including Alzheimer’s disease and chronic traumatic encephalopathy. Currently, we have only limited understanding of specific clinical phenotypes for distinct types of post-traumatic dementia, nor are there in vivo tests for many of the pathologies. Informed by our experience running a midlife brain health clinic for retired elite contact sport athletes, we describe a practical framework for the workup of athletes with cognitive concerns, highlighting key clinical features, an approach to investigation including neuroimaging and advanced fluid biomarkers, symptomatic management strategies and research directions.

Background: Contact sports, including rugby union, are associated with higher rates of neurodegenerative dementia, due to various underlying pathologies such as Alzheimer’s disease (AD) and chronic traumatic encephalopathy (CTE). New ultrasensitive multiplexed immunoassays may clarify disease mechanisms after repetitive head impacts (RHI) and traumatic brain injury, potentially aiding risk-stratification, early diagnosis and dementia treatment. Methods: Midlife participants in the ABHC cohort underwent plasma biomarker quantification (NULISA - NUcleic acid Linked Immuno-Sandwich Assay; n=124 markers), 3T MRI, trauma exposure ascertainment and phenotyping. Regressions quantified exposure-specific protein expression, relationship to trauma (including position) and brain atrophy, using cluster analysis to test correlates of traumatic encephalopathy syndrome (TES). Results: 197 former elite rugby players and 33 controls were assessed. 24 (12.2%) met criteria for TES but none had dementia. Ex-players returned reduced plasma glial fibrillary acidic protein (GFAP), kallikrein-6 (KLK6) and synaptosomal-associated protein 25 (SNAP25). Ex-forwards specifically showed reduced plasma beta-site amyloid precursor protein cleaving enzyme 1 (BACE1), amyloid beta-38 (Aβ38), and increased phospho-tau ₁₈₁ (p-tau ₁₈₁ ). KLK6 was lower in ex-backs than controls. No biomarkers related to career duration, concussion load or regional brain volume, nor did clustering relate to TES. Conclusions: Ex-players showed distinctive plasma biomarker changes, more prominently in ex-forwards, possibly reflecting greater RHI exposure. Plasma KLK6, an endothelial serine protease, was reduced across the ex-player group, with potential diagnostic or prognostic utility in future. Reduced GFAP and SNAP25 in ex-forwards has an uncertain basis, while elevated p-tau- ₁₈₁ more so than p-tau ₂₁₇ points towards non-AD tau pathology. Our findings motivate longitudinal characterisation, including comparison with other neurodegenerative diseases.

Epidemiological studies have shown that elite rugby players are at greater risk of neurodegenerative disease in later life, with post-mortem studies conducted in ex-players demonstrating the presence of neuropathologies related to repetitive head impacts, such as chronic traumatic encephalopathy. However, detailed prospective data establishing the clinical presentation of former rugby players with brain health concerns are lacking. In particular, the rates of traumatic encephalopathy syndrome, the clinical correlate of chronic traumatic encephalopathy, and the relationship between clinical outcomes and repetitive head impacts are unknown. Two hundred former elite rugby players with brain health concerns and 33 matched healthy control subjects were assessed. Self-reported concussion history, career duration, player position, self-rated scales of depression, anxiety, sleep quality, post-concussion symptoms and quality of life, self and informant ratings of neuropsychiatric symptoms and executive function behaviours, were obtained. Formal cognitive testing, traumatic encephalopathy syndrome classification and 3 T MRI were performed. Former players had a median age of 44 years (90.5% male, median career length = 10.5 years, median self-reported career concussions = 7); 63% were forwards and 37% were backs. Ex-players had elevated scores compared to controls on all symptom scales except sleep quality. Despite frequent subjective memory complaints, performance on cognitive testing did not significantly differ from controls. No players fulfilled criteria for dementia. Twenty-four former players fulfilled research criteria for traumatic encephalopathy syndrome (seven with cognitive impairment, 12 with neurobehavioral dysregulation, five with both). Provisional levels of certainty for chronic traumatic encephalopathy were relatively low (21 ‘suggestive’, three ‘possible’, zero ‘probable/definite’). Forwards and those with higher self-reported concussions were more likely to be classified as having traumatic encephalopathy syndrome based on neurobehavioral disturbance. Symptom burden (depressive and anxiety symptoms, post-concussion symptoms, behaviour rating of executive dysfunction, and neuropsychiatric symptom severity) was higher in those with higher self-reported concussions but was unrelated to years of play or position played. Cavum septum pellucidum was visible on structural imaging in 24% of players (12% in controls) and was more common in the high compared to low concussion group (32% versus 16%). In summary, former elite rugby players in mid-life had significant symptom burden, especially those self-reporting more concussions. In contrast, objective cognitive impairments and traumatic encephalopathy syndrome were relatively uncommon and there was no evidence of dementia. These results provide insights into the clinical presentations of former elite rugby players with brain health concerns during mid-life and highlight the complex relationship between symptoms, cognition and head impact exposure.

Repetitive head impacts and traumatic brain injuries in contact sports, such as rugby union, are associated with increased risk of neurodegenerative diseases such as Alzheimer's disease and chronic traumatic encephalopathy. Advances in fluid and imaging biomarkers are transforming dementia diagnosis but have not been systematically applied to individuals previously exposed to head impacts during rugby participation. We used biomarkers, including those with sensitivity and specificity for early Alzheimer's pathology to explore neurodegenerative risk in mid-life elite retired rugby players with significant repetitive head impact exposure. Plasma neurofilament light, glial fibrillary acid protein, amyloid-β (Aβ)42, Aβ40 and phospho-tau217 were quantified using ultrasensitive single molecule array digital enzyme-linked immunosorbent assay (SiMoA) in former elite rugby players as well as age/sex-matched unexposed controls. 3 T MRI and neuropsychology assessments were performed, with National Institute for Neurological Disorders and Stroke criteria used to ascertain the presence of traumatic encephalopathy syndrome. Regression models were used to relate plasma/imaging biomarkers to clinical phenotype. Individual-levels analyses were performed for fluid and imaging metrics, based on control biomarker distributions. Biomarker data from two aligned un-exposed Alzheimer's cohorts were included to contextualize our findings. Two hundred ex-rugby players (median age 44 years, 90% males) and 33 unexposed controls were assessed. Twenty-four (12%) ex-players fulfilled criteria for traumatic encephalopathy syndrome but none had dementia. Plasma phospho-tau217 concentrations were 17.6% higher in ex-rugby players than controls (95% confidence interval 3.7–33.3, P = 0.047). A total of 46 (23.1%) ex-players had elevated phospho-tau217 at the individual level; as did 18 (9.0%) players in relation to raised plasma neurofilament light. Ex-players’ concentrations were lower than in unexposed adults with late onset Alzheimer's disease (n = 69). Ex-players also showed significantly reduced volumes in the frontal/cingulate cortex on voxel-based morphometry at the group level; with reduced white matter and lower hippocampal volume associated with longer career durations within ex-players. Trauma-associated white matter changes measured with diffusion tensor imaging were uncommon in ex-players (4.6%). Traumatic encephalopathy syndrome was significantly more common in ex-players with elevated phospho-tau217, while those with raised plasma neurofilament light had significantly more anxiety and depressive symptoms. Frontal brain volumes correlated negatively with neurofilament light (r = −0.21, P = 0.010), and hippocampal volumes correlated negatively with phospho-tau217 (r = −0.19, P = 0.024). Elite rugby participation is associated with abnormal fluid and neuroimaging neurodegeneration biomarkers in mid-life. These include elevated phospho-tau217, which may indicate amyloid-dependent tau pathology. The results provide support for using state-of-the-art neurodegenerative biomarkers in the evaluation of long-term effects of sports head impact exposure.

Background: Traumatic brain injury (TBI) is common in military campaigns and is a risk factor for dementia.ArmeDSerVices TrAuma and RehabilitatioNOutComE-TBI (ADVANCE-TBI) aims to ascertain neurological outcomes in UK military personnel with major battlefield trauma, leveraging advances in quantification of axonal breakdown markers like neurofilament light (NfL), and astroglial marker glial fibrillar acidic protein (GFAP) in blood. We aimed to describe the causes, prevalence and consequences of TBI, and its fluid biomarker associations.Methods: TBI history was ascertained in 1145 servicemen and veterans, of whom 579 had been exposed to major trauma. Functional and mental health assessments were administered, and blood samples were collected approximately 8 years postinjury, with plasma biomarkers quantified (n=1125) for NfL, GFAP, total tau, phospho-tau₁₈₁, amyloid-β 42 and 40. Outcomes were related to neurotrauma exposure.Results: TBI was present in 16.9% (n=98) of exposed participants, with 46.9% classified as mild-probable and 53.1% classified as moderate to severe. Depression (β=1.65, 95% CI (1.33 to 2.03)), anxiety (β=1.65 (1.34 to 2.03)) and post-traumatic stress disorder (β=1.30 (1.19 to 1.41)) symptoms were more common after TBI, alongside poorer 6 minute walk distance (β=0.79 (0.74 to 0.84)) and quality of life (β=1.27 (1.19 to 1.36), all p<0.001). Plasma GFAP was 11% (95% CI 2 to 21) higher post-TBI (p=0.013), with greater concentrations in moderate-to-severe injuries (47% higher than mild-probable (95% CI 20% to 82%, p<0.001). Unemployment was more common among those with elevated GFAP levels post-TBI, showing a 1.14-fold increase (95% CI 1.03 to 1.27, p<0.001) for every doubling in GFAP concentration.Conclusions: TBI affected nearly a fifth of trauma-exposed personnel, related to worse mental health, motor and functional outcomes, as well as elevated plasma GFAP levels 8 years post-injury. This was absent after extracranial trauma, and showed a dose-response relationship with the severity of the injury.

Pathophysiology and outcomes after Traumatic Brain Injury (TBI) are complex and heterogenous. Current classifications are uninformative about pathophysiology. Proteomic approaches with fluid-based biomarkers are ideal for exploring complex disease mechanisms, as they enable sensitive assessment of an expansive range of processes potentially relevant to TBI pathophysiology. We used novel high-dimensional, multiplex proteomic assays to assess altered plasma protein expression in acute TBI. We analysed samples from 88 participants from the BIO-AX-TBI cohort (n=38 moderate-severe TBI [Mayo Criteria], n=22 non-TBI trauma, n=28 non-injured controls) on two platforms: Alamar NULISA™ CNS Diseases and OLINK^® Target 96 Inflammation. Patient participants were enrolled after hospital admission, and samples taken at a single timepoint up to 10 days post-injury. Participants also had neurofilament light, GFAP, total tau, UCH-L1 (all Simoa^®) and S100B (Millipore) data. The Alamar panel assesses 120 proteins, most of which were previously unexplored in TBI, plus proteins with known TBI-specificity, such as GFAP. A subset (n=29 TBI, n=24 non-injured controls) also had subacute (10 days to 6 weeks post-injury) 3T MRI measures of lesion volume and white matter injury (fractional anisotropy). Differential Expression analysis identified 16 proteins with TBI-specific significantly different plasma expression. These were neuronal markers (calbindin2, UCH-L1, visinin-like protein1), astroglial markers (S100B, GFAP), neurodegenerative disease proteins (total tau, pTau231, PSEN1, amyloid-beta-42, 14-3-3γ), inflammatory cytokines (IL16, CCL2, ficolin2), cell signalling (SFRP1), cell metabolism (MDH1) and autophagy related (sequestome1) proteins. Acute plasma levels of UCH-L1, PSEN1, total tau and pTau231 correlated with subacute lesion volume. Sequestome1 was positively correlated, whilst CLL2 was inversely correlated, with white matter fractional anisotropy. Neuronal, astroglial, tau and neurodegenerative proteins correlated with each other, IL16, MDH1 and sequestome1. Exploratory clustering (k means) by acute protein expression identified 3 TBI subgroups that differed in injury patterns, but not age or outcome. One TBI cluster had significantly lower white matter fractional anisotropy than control-predominant clusters, but had significantly lower lesion subacute lesions volumes than another TBI cluster. Proteins that overlapped on two platforms had excellent (r>0.8) correlations between values. We identified TBI-specific changes in acute plasma levels of proteins involved in neurodegenerative disease, inflammatory and cellular processes. These changes were related to patterns of injury, thus demonstrating that processes previously only studied in animal models are also relevant in human TBI pathophysiology. Our study highlights how proteomic approaches might improve classification and understanding of TBI pathophysiology, with implications for prognostication and treatment development.