Brain imaging signatures of neuropathic facial pain derived by artificial intelligence

Timur Latypov, Matthew So, Peter Shih-Ping Hung, Pascale Tsai, M. Walker, Sarasa Tohyama, Marina Tawfik, Frank Rudzicz, Mojgan HodaieScientific 2023 · DOI

Abstract

Advances in neuroimaging have permitted the non-invasive examination of the human brain in pain. However, a persisting challenge is the objective differentiation of neuropathic facial pain subtypes, as diagnosis currently relies on subjective patient descriptions. This study utilizes artificial intelligence (AI) models with neuroimaging data to distinguish subtypes of neuropathic facial pain—Classical Trigeminal Neuralgia (CTN) and Trigeminal Neuropathic Pain (TNP)—and differentiate them from healthy controls (HC). Analyzing diffusion tensor and T1-weighted imaging data from 479 adults, our models achieved up to 95% accuracy in distinguishing CTN from HC and 91% for TNP from HC. While subtype differentiation remained challenging (51% accuracy), the models identified key structural indicators in the insula and orbitofrontal cortex. This work demonstrates that AI can identify objective regional structural signatures of chronic pain from brain imaging data alone.

The Challenge of Subjective Diagnosis

Neuropathic facial pain, particularly Classical Trigeminal Neuralgia (CTN), presents a significant diagnostic hurdle. Clinicians currently depend almost exclusively on patient descriptions of symptoms—often characterized as "stabbing" or "electric" sensations. Because different pain subtypes can share similar symptoms despite having distinct underlying causes, this subjectivity frequently leads to a "trial and error" approach to treatment. While neuroimaging has established that chronic pain alters brain structure, standard statistical methods often fail to identify consistent, generalizable biomarkers. This research leverages AI to uncover objective "imaging signatures" that could revolutionize diagnostic precision.

Multimodal AI Methodology

The study employed a multimodal approach, integrating two distinct MRI modalities:

  1. T1-weighted imaging to measure gray matter thickness, surface area, and volume.
  2. Diffusion Tensor Imaging (DTI) to assess white matter microstructural integrity through metrics like Fractional Anisotropy (FA) and Mean Diffusivity.

To analyze this high-dimensional data, researchers utilized two interpretable machine learning architectures: Balanced Random Forest (RF) and Bagged Logistic Regression (LR). Unlike "black box" models, these algorithms allow for the identification of specific brain regions driving classification. Unsupervised techniques, including Principal Component Analysis (PCA) and t-SNE, were used to visualize natural data clustering.

t-SNE visualization and PCA-based kernel density estimates showing separation between healthy controls and chronic pain patients.

Data and Validation

The analysis was conducted on a dataset of 479 adults, representing one of the largest cohorts in chronic pain literature. The group included 371 pain patients (265 CTN, 106 TNP) and 108 healthy controls. Data was aggregated from local sources and the Cambridge Centre for Ageing and Neuroscience (CamCAN) repository, with rigorous statistical validation (Two-One-Sided Tests) ensuring compatibility across different data sources.

Key Results and Model Performance

The AI models demonstrated exceptional performance in identifying the presence of chronic pain. The Bagged Logistic Regression model achieved a peak accuracy of 95% (AUC 0.99) for CTN vs. HC and 91% (AUC 0.97) for TNP vs. HC.

ROC curves illustrating the high classification performance of Logistic Regression (LR) and Random Forest (RF) models in distinguishing pain patients from healthy controls.

Key findings include:

  • Modality Influence: Random Forest models prioritized white matter (DTI) metrics, whereas Logistic Regression weighed gray and white matter more equally, suggesting white matter changes are particularly robust indicators in certain algorithmic contexts.
  • Predictive Regions: The most influential regions for identifying pain included the temporal area, contralateral insula, olfactory cortex, and white matter tracts such as the fornix and corpus callosum.
  • Subtype Differentiation: Distinguishing between CTN and TNP proved more difficult, with accuracy hovering around 51%. This suggests significant macro-structural overlap between these subtypes or potential "noise" in current clinical labeling. However, the models did highlight the insula and orbitofrontal cortex (OFC)—regions involved in emotional and inhibitory pain processing—as areas of divergence.

Limitations and Future Directions

A primary limitation is the "ground truth" reliance on clinical diagnoses, which are inherently subjective. If the initial human diagnosis is imprecise, it caps the AI's potential accuracy. Furthermore, while the study focused on cortical and subcortical structures, it did not include high-resolution measurements of the trigeminal nerve itself. Future models incorporating nerve-specific diffusivity metrics alongside whole-brain data may provide the necessary resolution to distinguish between pain subtypes with higher precision, ultimately leading to more personalized and effective treatment strategies.

Summary updated

Abstract

Advances in neuroimaging have permitted the non-invasive examination of the human brain in pain. However, a persisting challenge is in the objective differentiation of neuropathic facial pain subtypes, as diagnosis is based on patients' symptom descriptions. We use artificial intelligence (AI) models with neuroimaging data to distinguish subtypes of neuropathic facial pain and differentiate them from healthy controls. We conducted a retrospective analysis of diffusion tensor and T1-weighted imaging data using random forest and logistic regression AI models on 371 adults with trigeminal pain (265 classical trigeminal neuralgia (CTN), 106 trigeminal neuropathic pain (TNP)) and 108 healthy controls (HC). These models distinguished CTN from HC with up to 95% accuracy, and TNP from HC with up to 91% accuracy. Both classifiers identified gray and white matter-based predictive metrics (gray matter thickness, surface area, and volume; white matter diffusivity metrics) that significantly differed across groups. Classification of TNP and CTN did not show significant accuracy (51%) but highlighted two structures that differed between pain groups-the insula and orbitofrontal cortex. Our work demonstrates that AI models with brain imaging data alone can differentiate neuropathic facial pain subtypes from healthy data and identify regional structural indicates of pain.
BibTeX
@inproceedings{latypov2023brain,
  author = {Timur Latypov and Matthew So and Peter Shih-Ping Hung and Pascale Tsai and M. Walker and Sarasa Tohyama and Marina Tawfik and Frank Rudzicz and Mojgan Hodaie},
  title = {{Brain imaging signatures of neuropathic facial pain derived by artificial intelligence}},
  booktitle = {Scientific},
  year = {2023},
  doi = {10.1038/s41598-023-37034-y},
}
Figure