T2 Inflammation-Informed Class

• T2 inflammation-informed class is defined as a latent phenotype identified using Bayesian and machine learning methods to cluster patients based on molecular, cellular, and imaging markers.
• It integrates high-dimensional biomarkers such as blood eosinophil counts, allergy tests, and T2 relaxometry measures to enable precise stratification in heterogeneous inflammatory diseases.
• Empirical findings demonstrate that probabilistic classification enhances phenotyping accuracy and supports precision medicine initiatives in conditions like asthma and multiple sclerosis.

A T2 inflammation-informed class denotes a patient subgroup, defined via statistical or machine learning frameworks, characterized by molecular, cellular, or imaging correlates of type 2 (T2) immune-mediated inflammation. This class serves as a clinically meaningful phenotype for heterogeneous diseases such as asthma and multiple sclerosis (MS), where T2 inflammation underpins pathophysiology and therapeutic response. Assignments to this class are often probabilistic and leverage high-dimensional input—including blood eosinophil levels, allergy markers, healthcare utilization data, and quantitative imaging features—by deploying Bayesian latent class models or advanced machine learning classifiers. Key advances enable the robust discovery, quantification, and downstream application of a T2 inflammation-informed class in both clinical and imaging-derived datasets.

1. Model-Based Definition and Identification

The T2 inflammation-informed class is commonly operationalized as a latent construct within a generative statistical model. In large-scale EHR phenotyping, Bayesian latent class analysis is employed, incorporating prior clinical knowledge to encourage clustering of patients with elevated T2 markers.

Let $D_i \in \{0, 1\}$ denote the latent indicator of T2-class membership for patient $i$; $D_i=1$ indicates assignment to the T2 inflammation-informed class. The observed variables include baseline demographics $X_i$, continuous biomarkers $Y_{ij}$ (e.g., standardized log eosinophil counts), binary clinical features $W_{ik}$ (e.g., positive allergy skin test), and associated missingness indicators $R_{ij}$. The fully generative model is: $$p(D_i, R_i, Y_i, W_i \mid X_i, \Theta) = P(D_i \mid X_i; B_D) \prod_{j=1}^J P(R_{ij} \mid D_i, X_i; B_{R,j}) \prod_{j=1}^J \left[ P(Y_{ij} \mid D_i, X_i; B_{Y,j}, \sigma_{Y,j}^2) \right]^{R_{ij}} \prod_{k=1}^K P(W_{ik} \mid D_i, X_i; B_{W,k})$$ with Gaussian and Bernoulli conditionals and informative priors on key T2-related coefficients.

In imaging contexts, voxel-wise classifiers are constructed via decision-tree ensembles or neural networks, which assign each imaging location a probability of pathological tissue class, informed by inflammation-sensitive modalities such as T2 relaxometry and myelin water imaging.

2. Domain Knowledge Integration and Prior Specification

Embedding prior biological or clinical knowledge at the modeling stage is critical for specifying a T2 inflammation-informed class that is both interpretable and clinically meaningful. Informative priors are imposed on parameters connecting latent class to canonical T2 features:

• For standardized log eosinophil count $Y_{EOS}$:
  • $\beta_{EOS,0} \sim \mathrm{Normal}(-0.68,\,0.05)$
  • $$\beta_{EOS,1} \sim \mathrm{Normal}(0.70-\beta_{EOS,0},\,0.05)$$
  • Ensures $P(YEOS < 0.15|D=0)\approx 0.24$ and $P(YEOS \ge 0.15|D=1)\approx 0.75$ (specificity 76%, sensitivity 75%).
• For binary allergy tests and codes, priors reflect observed sensitivity and specificity in the population.
• All other predictors receive weakly informative $\mathrm{Normal}(0, 1)$ priors, allowing data-driven discovery for utilization, exacerbation rates, and pharmacotherapy.

In multi-modal MRI-based classifiers, knowledge of T2 compartment physiology informs the modeling of the signal decay spectrum (e.g., pooling T2 components in physiological ranges for myelin, intra/extracellular, and free water fractions).

3. Methodological Frameworks: Bayesian Latent Class and Machine Learning Approaches

The Bayesian latent class model with informative priors exploits both observed and missing-not-at-random (MNAR) data patterns, producing patient-level posterior probabilities for class membership and allowing formal quantification of assignment uncertainty. Marginalizing over $D_i$ yields the observed-data likelihood, and inference is performed via Hamiltonian Monte Carlo (NUTS sampling, e.g., using Stan v2.32.2).

For imaging, advanced classifiers integrate voxel-level imaging descriptors:

• Regularized non-negative least squares (NNLS) for T2 component estimation.
• Extraction of water fraction biomarkers, e.g., myelin water fraction (MWF), fractional intra/extra-neurite signals (from Spherical Mean Technique diffusion), and free water/CSF fraction.
• Ensemble classifiers such as AdaBoost (SAMME) with decision stumps as weak learners, trained directly on physiological fraction features, achieving robust tissue class assignment with explicit control over class impurity and boosting update rules.

Neural network–based T2 distribution estimation, as in the physically-primed $P_2T_2$ model, incorporates both measured MRI signal sequences and acquisition parameters as DNN inputs, with custom loss functions enforcing Wasserstein distance–based physical consistency and compartmental relevance for downstream inflammation classification.

4. Key Features and Biomarker Extraction

Whether derived from EHR or imaging, the T2 inflammation-informed class is anchored by quantitative biomarkers that have established mechanistic ties to T2 inflammation:

• Blood eosinophil count (log + standardized): posterior mean for class: $\hat{\mu}_{EOS|D=1} \approx +0.22~(0.21,~0.23)$.
• Total IgE (standardized): $\hat{\mu}_{IgE|D=1} \approx +0.24~(0.23,~0.25)$.
• Allergic ICD code prevalence: posterior probability $\approx 62.6\%$.
• Health care utilization: annual asthma encounter rate $= 4.06~(3.95,~4.17)$, ED visit rate $= 0.18$.
• Medication: ICS/LABA ever prescribed $= 85.9\%$.

For imaging classifiers:

• Estimated features: myelin water fraction $f_M$, intra-/extra-cellular pool fraction $f_{IE}$, free water $f_{CSF}$, SMT-based intra-/extra-neurite signal fractions.
• Classifier input: $$F_\mathrm{comb} = [SMT_{f_I},~SMT_{f_E},~f_M,~f_{IE},~f_{CSF}]^{\top}$$ without explicit normalization, reflecting physiological distributions.

Enhancements in the fidelity of MWF and edema fractions are directly translatable to robust detection of T2 inflammation in imaging datasets.

5. Empirical Findings and Class Properties

Posterior class probabilities in EHR-based analyses present stark bimodality (Mayer et al., 3 Nov 2025): the T2-high class comprises approximately 36–39% of the asthma EHR cohort, with minimal intermediate assignment. T2-high members distinctly exhibit raised eosinophils, allergy markers, high health care utilization, and intensive medication use—even when utilization variables receive only weakly informative priors.

In imaging-based contexts (Fischi-Gomez et al., 2021), the combination of T2 relaxometry with diffusion metrics in boosting-based classifiers substantially improves lesion–normal accuracy in MS, with the three-class lesion—normal-appearing white matter—healthy control accuracy rising to 59.9% (compared to 50.5–52.8% for either modality alone). Lesion vs. NAWM and lesion vs. control binary accuracies reach 84.7% and 85.1%, representing 6–20 percentage point improvements over single-modal approaches.

Neural network–based T2 estimation ($P_2T_2$ (Ben-Atya et al., 2022)) achieves lower mean square errors (by 21–53% for MWF at low SNR), robustly outperforms traditional methods under echo time shifts (reducing $T_2$ MSE by $\sim$35%), and yields high-fidelity compartment maps suitable for downstream inflammation detection.

6. Practicalities, Limitations, and Clinical Implications

Implementing a T2 inflammation-informed class approach requires attention to:

• Data quality (rigorous denoising, registration, correction for imaging; appropriate covariate and missingness modeling for EHR).
• Automated regularization (e.g., L-curve for NNLS lambda in MRI).
• Addressing limited sample sizes, potential overfitting, and class imbalance by cross-validation and cost-sensitive learning.
• Interpreting multi-channel classifier probability maps requires careful thresholding and external validation, ideally against histopathology or longitudinal outcomes.
• In EHR data, explicit modeling of MNAR reveals that omission of tests (e.g., IgE) is highly informative for disease severity.

A plausible implication is that T2 inflammation-informed classes, as defined by these frameworks, not only facilitate hypothesis generation but also enable precise cohort identification for precision medicine, clinical trial design, and mechanistically targeted therapeutics—particularly in complex, heterogeneous diseases lacking universally accepted phenotype definitions.

7. Future Directions and Research Opportunities

Continued progress is anticipated in the joint modeling of imaging and EHR-derived features, integration of novel T2 compartment–sensitive neural architectures, and external validation across institutions and diseases. Promising avenues include:

• End-to-end training of deep learning models harmonizing T2 distribution estimation with downstream disease classifiers.
• Incorporation of spatial priors (e.g., CRF or U-Net–style constraints) for stability in parametric mapping.
• Use of the T2 inflammation-informed class as a stratification variable for outcome studies and intervention trials across a spectrum of T2-dominant inflammatory diseases.

Persistent challenges include standardization of input data, harmonization of acquisition protocols, and clinical validation of probabilistic class membership for decision support.