An Integrated Data-Driven Model for Clinical Phenotyping of Tuberculosis Disease Severity

Samantha Malatesta, Karen R. Jacobson, C. Robert Horsburgh, Maha Farhat, Tara Carney, Krista J. Gile, Eric D. Kolaczyk, Laura WhitemedRxiv 2025 · DOI

TL;DR

Researchers have developed TB-STATIS, a statistical framework that moves beyond binary "mild vs. severe" tuberculosis classifications. By applying event-based modeling to routine diagnostic data, the tool reconstructs a nuanced spectrum of disease severity that accurately predicts how patients will respond to treatment.

What Problem Does This Paper Address?

Tuberculosis (TB) treatment is notoriously long—typically six months of multi-drug therapy. While recent trials suggest some patients can be cured in just four months, clinicians lack a precise way to identify who qualifies for these shorter regimens.

Currently, disease severity is often treated as a binary: "minimal" or "advanced." This oversimplification ignores the biological reality that TB is a continuous spectrum of bacterial burden and inflammation. The challenge is that we usually only see a patient at one point in time (diagnosis). Without longitudinal data, it is difficult to map where a patient sits on the "timeline" of disease progression. This paper aims to fill that gap by using cross-sectional data to infer a population-level disease sequence.

Methods

The authors introduce TB-STATIS (Tuberculosis SeveriTy Assessment Tool for Informed Stratification). The method is rooted in Event-Based Modeling (EBM), a technique originally used to track the progression of neurodegenerative diseases like Alzheimer’s.

Key technical features include:

  • Sequence Inference: It assumes that as disease worsens, biomarkers (like smear grade or cough) transition from "normal" to "abnormal" in a predictable order. By looking at a "snapshot" of a large population, the model identifies the most likely order of these transitions.
  • Categorical/Ordinal Integration: Unlike standard EBMs that require continuous data, TB-STATIS is designed for the categorical and ordinal data common in TB clinics (e.g., smear grades of 1+, 2+, 3+).
  • No "Healthy" Controls Required: Most progression models need a reference group of healthy individuals to define "normal." TB-STATIS bypasses this by using validated diagnostic cutoffs, making it more practical for real-world clinical datasets.
  • Maximum Likelihood Estimation: The model uses an iterative search algorithm to find the sequence of clinical states that best explains the observed data across the entire cohort.

Datasets

The model was validated using three distinct sources:

  1. Simulation Studies: 1,000 synthetic datasets were used to test the model's robustness against varying sample sizes and levels of data "noise" (diagnostic uncertainty).
  2. REMoxTB Trial (N=1,351): Data from a global Phase 3 clinical trial testing 4-month vs. 6-month regimens. This provided high-quality clinical and laboratory measures.
  3. TRUST Study (N=351): An observational cohort from a clinic in South Africa, representing a "real-world" population with high HIV prevalence.

Key Insights and Ideas

  • Reconstructing Time from a Snapshot: The most powerful aspect of TB-STATIS is its ability to infer a temporal progression (a timeline) using only data collected at a single visit. It leverages the variation across a population to "fill in" the stages of the disease.
  • Clinical Meaningfulness: The severity classes generated by the model aren't just statistical clusters; they correlate with biological markers of treatment response, such as culture conversion (the point when a patient's sputum no longer grows TB bacteria).
  • Flexibility: The model is "plug-and-play." It doesn't require a fixed set of variables, meaning researchers can incorporate new diagnostics—like gene expression markers or advanced imaging—as they become available.

Key Results

  • High Accuracy in Simulation: In simulated environments with 4 to 8 clinical states, TB-STATIS recovered the "ground truth" disease sequence in over 90% of cases when diagnostic accuracy was $\ge 0.85$.
  • Stratification in REMoxTB: The model identified six distinct severity classes. Patients in higher (more severe) classes had significantly higher odds of remaining culture-positive at 8 weeks—a key proxy for treatment failure.
  • Validation in TRUST: In the South African cohort, the model identified seven classes. Higher severity classes were associated not only with slower culture conversion but also with known risk factors like lower BMI and history of drug use.

Limitations and Future Work

The authors acknowledge several important caveats:

  • Linearity Assumption: The model assumes a single, universal sequence of disease progression. In reality, subgroups (e.g., people living with HIV) might follow different biological trajectories. Future iterations could incorporate "subtype" modeling to address this.
  • Independence of Measures: The current model assumes that different clinical measures (like X-ray findings and symptoms) are independent, which may not always hold true.
  • Input Sensitivity: The model requires users to specify the "accuracy" of each diagnostic measure. While the study showed the model is robust to slight misspecifications, this remains a manual step.

Ultimately, TB-STATIS represents a significant step toward precision medicine for TB, providing a data-driven way to tailor treatment duration to the individual's specific stage of disease.

Summary updated

Abstract

A common approach to describing tuberculosis (TB) disease severity is to use a binary classification such as "advanced" and "minimal or early disease," though this may not fully capture the range of clinical presentations. As individuals transition through stages of disease, we expect to observe increased bacterial burden and inflammation which corresponds to worsening disease severity and increased risk of a negative outcome. We develop a new method, tuberculosis SeveriTy Assessment Tool for Informed Stratification (TB-STATIS), to understand the various disease severity phenotypes that exist at time of clinical presentation. Our method integrates data from multiple sources (i.e. smear microscopy, chest x-ray findings, symptoms, etc.) to identify a set of disease severity classes and obtain a predicted disease class for each individual given their observed data. Our approach is motivated by the statistical framework used in event-based modeling, a type of data-driven disease progression modeling. We show in simulation TB-STATIS can correctly identify the true set of disease classes with various sample sizes, data sources to integrate, and levels of uncertainty in the observed data. We apply TB-STATIS to two data sets, data from an observational TB cohort in South Africa and data from a global phase 3 clinical trial that tested the non-inferiority of two 4-month regimens compared to the standard 6-month regimen for the treatment of TB. We observe disease classes generated from TB-STATIS correlate with culture conversion, a proxy for TB treatment response. We demonstrate our approach to classifying TB disease severity generates clinically meaningful strata.
BibTeX
@misc{malatesta2025integrated,
  author = {Samantha Malatesta and Karen R. Jacobson and C. Robert Horsburgh and Maha Farhat and Tara Carney and Krista J. Gile and Eric D. Kolaczyk and Laura White},
  title = {{An Integrated Data-Driven Model for Clinical Phenotyping of Tuberculosis Disease Severity}},
  year = {2025},
  doi = {10.1101/2025.11.25.25341018},
  howpublished = {medRxiv},
}
Figure