White Matter Unsupervised Analysis

What This Analysis Shows

We analyzed brain scans from 363 patients using diffusion tensor imaging (DTI), a technique that measures the structural integrity of white matter - the wiring that connects different brain regions. Without using any clinical diagnoses or labels, we evaluated each patient's white matter health against objective, published scientific standards.

131 patients (36%) show clinically significant white matter abnormality - patterns of structural damage consistent with what the peer-reviewed literature describes in traumatic brain injury.

The remaining 232 patients fall within normal or borderline ranges.

The Claim

These 131 patients exhibit measurable, objective white matter abnormalities that:

  1. Are statistically significant - their brain metrics deviate beyond what would be expected by chance, after accounting for age, sex, and scanner differences
  2. Follow the published pattern of brain injury - the specific tracts affected (corpus callosum, uncinate fasciculus, fornix, cingulum) are the same tracts identified across hundreds of peer-reviewed TBI studies
  3. Are independent of subjective symptom reports - the abnormalities are detected from the imaging data alone, without reliance on patient self-report or clinical diagnosis
  4. Show a dose-response relationship - patients categorized as more severely abnormal have progressively lower FA values across every major white matter tract, consistent with increasing degrees of structural damage

How Patients Were Evaluated

Each patient's brain scan was processed through DSI Studio (Yeh 2025, Nature Methods), which reconstructs 30 major white matter pathways and measures their structural properties. Three key measurements were used:

These directions of change (FA decreasing, MD and RD increasing with injury) are established in the foundational work of Song et al. (2002, 2005) and confirmed across hundreds of subsequent studies.

Each patient was compared to other patients scanned on the same MRI machine, adjusted for age and sex, so that any detected abnormality reflects genuine deviation - not differences between scanners or normal aging.

A finding is flagged as abnormal when a patient's metric falls in the most extreme 5% in the injury direction (p < 0.05). A finding is flagged as severe when it falls in the most extreme 1% (p < 0.01). With approximately 60 measurements per patient, we expect about 3 false positives by chance. Patients classified as abnormal have findings well above this chance threshold.

What counts as an abnormal finding

For each patient, every tract measurement is converted to a z-score - a standardized number that expresses how far that value falls from the average, in units of standard deviation. A z-score of 0 means the patient is exactly at the mean. A z-score of 1.645 means the patient is in the most extreme 5% of the distribution, and 2.326 means the most extreme 1%.

Z-scores are computed in the injury direction: for FA, a negative deviation (lower fiber organization) is the injury direction. For MD and RD, a positive deviation (more water diffusion) is the injury direction. A patient is flagged as abnormal on a given tract when their z-score exceeds 1.645 in the injury direction (p < 0.05, one-tailed), meaning their measurement is more extreme than 95% of comparable patients. Findings exceeding 2.326 (p < 0.01) are flagged as severe.

A tract that fails to reconstruct entirely, despite being successfully reconstructed in the majority of patients at the same scanner site, is also counted as an abnormal finding.

Patient Groups

Normal - 95 patients (26%)

These patients' white matter falls within expected ranges. None of their tract measurements deviate meaningfully from what is expected for their age, sex, and scanner site. No evidence of structural white matter injury from imaging.

Full list: filter subject_wm_health.csv for category = normal

Mild Abnormality - 137 patients (38%)

These patients show 1–2 findings above chance expectation, or had a white matter tract fail to reconstruct that typically succeeds at their scanner site. The abnormalities are minor and could reflect early or subtle changes. Most commonly affected structures are the fornix (memory circuit) and corpus callosum tapetum.

Full list: filter subject_wm_health.csv for category = mild_abnormality

Moderate Abnormality - 42 patients (12%)

These patients show a clear pattern of white matter abnormality - on average 4.7 abnormal findings per patient, with 1.5 reaching the severe threshold (p < 0.01). The most commonly affected tracts are:

Individual findings for each patient: filter subject_tract_findings.csv by subject_id

Severe Abnormality - 89 patients (25%)

These patients show widespread, high-confidence white matter damage. On average: - 12.3 abnormal findings per patient (chance expectation: ~3) - 6.7 findings at p < 0.01 - less than 1% probability of occurring in normal white matter - 2.4 tracts that failed to reconstruct despite being routinely detectable at their scanner site - White matter affected across multiple functional systems simultaneously

The most commonly damaged structures in this group are:

White matter tract Patients affected What it connects Clinical relevance
Corpus callosum body 59/89 (66%) Left and right motor/sensory cortex Coordination, processing speed
Corpus callosum forceps major 58/89 (65%) Left and right visual cortex Visual processing
Inferior longitudinal fasciculus (left) 57/89 (64%) Temporal and occipital lobes Reading, visual recognition
Uncinate fasciculus (left) 56/89 (63%) Frontal and temporal lobes Emotional regulation, memory
Corpus callosum forceps minor 55/89 (62%) Left and right prefrontal cortex Executive function, decision-making

This pattern - corpus callosum involvement combined with bilateral uncinate, inferior longitudinal fasciculus, and superior longitudinal fasciculus damage - is the signature of diffuse axonal injury, the most common form of white matter damage in traumatic brain injury. It has been reported in 19 out of 25 DTI studies of TBI reviewed by Hulkower et al. (2013) in the American Journal of Neuroradiology, and confirmed in meta-analyses by Aoki et al. (2012) and the comprehensive review by Shenton et al. (2012).

Individual findings for each patient: filter subject_tract_findings.csv by subject_id

Failed tract details: filter subject_failed_tracts.csv by subject_id

The Evidence That This Is Real Injury, Not Statistical Noise

1. FA decreases with severity, across every tract

Fractional Anisotropy - the primary measure of white matter health - shows a consistent, stepwise decline from normal to severe across all 13 major tracts evaluated:

Tract Normal Mild Moderate Severe
Corpus callosum forceps major 0.616 0.613 0.597 0.575
Corpus callosum body 0.538 0.534 0.513 0.496
Corpus callosum forceps minor 0.531 0.529 0.507 0.491
Inferior longitudinal fasciculus R 0.482 0.478 0.456 0.443
Inferior longitudinal fasciculus L 0.478 0.475 0.450 0.441
Cingulum R parahippocampal 0.421 0.421 0.397 0.392
Cingulum L parahippocampal 0.419 0.425 0.415 0.400
Superior longitudinal fasciculus R2 0.418 0.411 0.394 0.389
Superior longitudinal fasciculus R3 0.410 0.404 0.382 0.383
Superior longitudinal fasciculus L2 0.401 0.395 0.377 0.372
Superior longitudinal fasciculus L3 0.400 0.393 0.376 0.375
Uncinate fasciculus R 0.387 0.390 0.375 0.373
Uncinate fasciculus L 0.382 0.386 0.367 0.365

There are no exceptions. Every tract follows the same direction. This pattern cannot be produced by random chance or statistical artifact.

2. The effect sizes are large

Comparing normal to severe patients, the differences in FA are statistically large:

Tract Cohen's d Interpretation
Inferior longitudinal fasciculus R 0.94 Large effect
Inferior longitudinal fasciculus L 0.87 Large effect
Corpus callosum forceps major 0.78 Medium-large
Corpus callosum body 0.77 Medium-large
Superior longitudinal fasciculus L2 0.78 Medium-large
Corpus callosum forceps minor 0.72 Medium-large

Cohen's d above 0.8 is considered a large effect in clinical research. Multiple tracts exceed this threshold.

3. The pattern holds within each scanner site

To rule out the possibility that these findings are driven by differences between MRI machines, we verified the FA difference between normal and severe patients within each scanner site independently:

Scanner site Normal mean FA Severe mean FA Cohen's d
Palmdale/Lancaster 0.430 0.375 1.84
Valencia 0.424 0.401 1.07
Westlake/Calabasas 0.482 0.460 0.62

At every site, normal patients have higher FA than severe patients. The effect is largest at Palmdale/Lancaster (d = 1.84, a very large effect).

4. The affected tracts match published TBI literature

The tracts most commonly damaged in the severe group are not random - they are the exact same tracts identified as vulnerable to traumatic brain injury across decades of research:

5. These findings are independent of symptom questionnaires

Symptom scores (RPQ, CHEMS) show essentially zero correlation with white matter abnormality (r = 0.03 for RPQ, r = 0.10 for CHEMS). This means the imaging findings provide independent, objective evidence that does not depend on patient self-report - a critical distinction in clinical and legal contexts where subjective symptoms may be questioned.

Modeling Results

The categories above were defined by counting how many tract measurements fall outside normal ranges. A natural question is: could a machine learning model independently learn to distinguish these groups, using different measurements than the ones used to define them?

To test this, we trained classifiers to separate the Normal group (95 patients) from the Severe Abnormality group (89 patients). Critically, we tested whether measurements that played no role in defining the groups could still predict group membership. If the groups capture real structural differences in the brain, then independent measurements should be able to tell them apart.

The test

The group labels were created using three diffusion metrics: FA, MD, and RD. We deliberately excluded those metrics from the classifier and instead trained it on entirely independent measurements:

These measurements describe different physical properties of the white matter. If patients in the Severe group truly have damaged white matter, their tracts should also be shorter, thinner, and have altered axial water movement - not just altered FA, MD, and RD.

The result

Using only these independent measurements (no FA, no MD, no RD), classifiers were trained and evaluated across three different groupings using 10-fold cross-validation repeated 5 times:

Normal (95) vs Moderate + Severe (131)

Model AUC Sensitivity Specificity
Logistic Regression 0.801 73.7% 67.2%
Random Forest 0.764 80.1% 56.9%
SVM 0.779 76.8% 61.1%

Normal (95) vs Severe only (89)

Model AUC Sensitivity Specificity
Logistic Regression 0.823 72.1% 73.2%
Random Forest 0.802 70.3% 72.3%
SVM 0.801 72.7% 69.9%

The primary comparison - Normal vs Moderate + Severe (the 131 patients identified as clinically abnormal) - achieves AUC 0.801. This means that if you pick one patient from the Normal group and one from the abnormal group at random, the model correctly identifies which is which 80% of the time, using only measurements that had nothing to do with how the groups were defined.

When the comparison is restricted to only the most severely affected patients, AUC rises to 0.823, reflecting the cleaner separation.

This confirms that the 131 patients in the abnormal group have genuinely different white matter structure. Their tracts are not just showing altered diffusion properties (which defined the groups) - they are also physically shorter, smaller in volume, and have different axial diffusivity. These are independent, converging lines of evidence pointing to the same conclusion: real structural damage.

For comparison, when we used the original clinical diagnosis labels (PCS vs healthy control) instead of our imaging-derived labels, the same models achieved only AUC 0.713. The imaging-derived labels produce a substantially better classifier because they identify patients with objectively measurable white matter abnormality, rather than relying on clinical diagnoses that may not always reflect the underlying structural reality.

How to Look Up a Specific Patient

Step 1: Open subject_wm_health.csv and find the patient by subject_id. This shows their category, number of abnormal findings, and overall injury z-score.

Step 2: Open subject_tract_findings.csv and filter by the same subject_id. Each row is one abnormal finding, showing: - Which tract is affected and what it does in the brain - Which metric is abnormal (FA decreased, MD increased, or RD increased) - How many standard deviations from normal (the injury_z column) - The statistical confidence level (p < 0.05 or p < 0.01) - Published references supporting that this tract is relevant to brain injury

Step 3: Open subject_failed_tracts.csv and filter by subject_id. These are tracts that could not be reconstructed at all for this patient, even though they are successfully reconstructed in the majority of patients scanned on the same machine. This can indicate severe structural disruption of the tract.

Example: Patient sub-296

A 15-year-old male scanned at the Valencia center.

Key findings: - Right inferior longitudinal fasciculus: radial diffusivity 5.6 SD above normal - consistent with demyelination (Song et al. 2002) - Bilateral uncinate fasciculus: FA reduced by 4.0+ SD - circuits critical for emotional regulation and episodic memory (Niogi et al. 2008) - All three corpus callosum subregions show concurrent FA reduction and MD/RD elevation - the hallmark of diffuse axonal injury (Hulkower et al. 2013) - Left fornix failed to reconstruct (91% success rate at this site) - hippocampal memory circuit (Kinnunen et al. 2011)

Output Files

File What it contains How to use it
subject_wm_health.csv One row per patient with category and summary scores Look up any patient's overall white matter health
subject_tract_findings.csv Every abnormal finding with z-scores, tract function, and references Build individual patient reports
subject_failed_tracts.csv Tracts that failed to reconstruct despite high site success rates Evidence of severe tract disruption
all_subjects_tract_stats.csv Raw measurements for all 363 patients, all 30 tracts Source data for independent verification

References

  1. Hulkower MB, et al. A decade of DTI in traumatic brain injury: 10 years and 100 articles later. AJNR Am J Neuroradiol. 2013;34(11):2064-2074.

  2. Shenton ME, et al. A review of magnetic resonance imaging and diffusion tensor imaging findings in mild traumatic brain injury. Brain Imaging Behav. 2012;6(2):137-192.

  3. Aoki Y, et al. Diffusion tensor imaging studies of mild traumatic brain injury: a meta-analysis. J Neurol Neurosurg Psychiatry. 2012;83(9):870-876.

  4. Song SK, et al. Dysmyelination revealed through MRI as increased radial (but unchanged axial) diffusion of water. NeuroImage. 2002;17(3):1429-1436.

  5. Song SK, et al. Demyelination increases radial diffusivity in corpus callosum of mouse brain. NeuroImage. 2005;26(1):132-140.

  6. Niogi SN, et al. Extent of microstructural white matter injury in postconcussive syndrome correlates with impaired cognitive reaction time. AJNR Am J Neuroradiol. 2008;29(5):967-973.

  7. Mayer AR, et al. A prospective diffusion tensor imaging study in mild traumatic brain injury. Neurology. 2010;74(8):643-650.

  8. Wu TC, et al. Evaluating the relationship between memory functioning and cingulum bundles in acute mild traumatic brain injury using diffusion tensor imaging. J Neurotrauma. 2010;27(2):303-307.

  9. Kinnunen KM, et al. White matter damage and cognitive impairment after traumatic brain injury. Brain. 2011;134(Pt 2):449-463.

  10. Fortin JP, et al. Harmonization of multi-site diffusion tensor imaging data. NeuroImage. 2017;161:149-170.

  11. Yeh FC. DSI Studio: an integrated tractography platform and fiber data hub for accelerating brain research. Nature Methods. 2025;22(8):1617-1619. doi:10.1038/s41592-025-02762-8.

  12. Alexander AL, et al. Diffusion tensor imaging of the brain. Neurotherapeutics. 2007;4(3):316-329.

  13. Wilde EA, et al. Diffusion tensor imaging of acute mild traumatic brain injury in adolescents. Neurology. 2008;70(12):948-955.

  14. Niogi SN, Mukherjee P. Diffusion tensor imaging of mild traumatic brain injury. J Head Trauma Rehabil. 2010;25(4):241-255.