📄 infectious diseases

Time-resolved predictability of end-of-therapy outcome and relapse after cure in Phase 3 tuberculosis trials

This study utilizing data from two Phase 3 tuberculosis trials demonstrates that while routine clinical data collected during treatment significantly improves the prediction of end-of-therapy outcomes, it offers only limited value for forecasting post-treatment relapse, highlighting the urgent need for specific relapse biomarkers.

Original authors: Garger, D., Burks, J. H., Paton, N. I., Sloan, D. J., Maher-Edwards, G., Theis, F., Menden, M. P., Casale, F. P.

Published 2026-09-21

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Garger, D., Burks, J. H., Paton, N. I., Sloan, D. J., Maher-Edwards, G., Theis, F., Menden, M. P., Casale, F. P.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Tuberculosis is an ancient bacterial infection that attacks the lungs, and for decades, the medical world has relied on a long course of antibiotics to clear it. When a patient finishes this treatment, doctors face a difficult uncertainty: has the disease truly vanished, or is it merely hiding, waiting to return? This question of relapse is central to the fight against the disease. While a cure is often declared when a patient's sputum tests negative for bacteria at the end of therapy, the risk of the infection coming back does not disappear instantly. Understanding how this risk changes day by day, and whether we can spot the signs of a future return before treatment even ends, remains one of the most stubborn challenges in infectious disease.

A recent study set out to map this uncertainty by looking at the data from nearly three thousand people who participated in major clinical trials for tuberculosis. The researchers wanted to know if they could predict two things: whether a patient would finish their treatment successfully, and whether they would later relapse after being declared cured. To do this, they did not just look at a single snapshot of a patient's health. Instead, they gathered information collected every single month from the start of treatment until the very end. They treated this stream of data like a timeline, training computer models to learn from the monthly updates on symptoms, age, sex, and the amount of bacteria visible in the patient's sputum.

The results drew a sharp distinction between predicting the end of treatment and predicting a return of the disease. When the goal was to forecast whether a patient would have a successful outcome at the end of their therapy, the models became quite accurate after the third month. This improvement was driven by two specific factors: whether the bacteria could still be seen under a microscope in the sputum, and whether they could be grown in a culture dish. These standard tests provided a clear signal that the treatment was working. However, the story changed completely when the researchers tried to predict relapse among those who had already finished treatment successfully. Here, the predictive power was much weaker. Even as the months passed and more data accumulated, the models struggled to foresee who would get sick again, with their accuracy remaining low and even dipping slightly as time went on.

The study also tested a newer, more flexible way of analyzing the same information. Instead of feeding the computer raw numbers and simple categories, the researchers used a method that converts all the patient details into complex digital patterns, allowing the system to find subtle connections that standard methods might miss. This approach performed just as well as the traditional method for predicting relapse, but it did offer a distinct advantage for predicting the end-of-treatment outcome. At months three and four, this advanced method was noticeably better at identifying who would finish therapy successfully. Yet, even with this sophisticated tool, the ability to foresee a future relapse remained limited. The most reliable factors for predicting a return of the disease were not the monthly test results, but rather the patient's age, sex, initial symptoms, and the bacterial burden at the very start of treatment.

Perhaps the most significant finding was that adding data collected after the first month of treatment did offer some real value, but only in a specific way. When the researchers compared a model that used only the initial baseline data against one that included the monthly updates, the updated model was better at sorting patients into low-risk and high-risk groups for relapse. For instance, four months after finishing treatment, the low-risk group identified by the updated model had a survival rate without relapse of nearly ninety-five percent, compared to about ninety-one percent for the group identified by baseline data alone. This suggests that watching a patient's progress over time does refine our understanding of their risk, but the improvement is modest. The study concludes that while routine monthly checks help, they are not enough to solve the problem of predicting relapse. The limited success of these models underscores a clear need for new, specific biological markers that can signal a return of the disease long before it becomes clinically apparent.

Technical Summary: Time-resolved predictability of end-of-therapy outcome and relapse after cure in Phase 3 tuberculosis trials

Problem Statement
Despite the apparent success of tuberculosis (TB) therapy, relapse remains a significant clinical challenge that is difficult to predict. A critical gap in current understanding is how the risk of relapse evolves dynamically throughout the treatment course. While baseline factors are known to influence outcomes, it remains unclear whether routine clinical data collected at monthly intervals during therapy can effectively stratify patients for post-treatment relapse risk or improve the prediction of end-of-therapy (EOT) outcomes.

Methodology
The study leveraged harmonized clinical data from two Phase 3 TB trials, encompassing 2,918 participants. The researchers employed a time-resolved prediction framework, training machine learning models on tabular data at monthly intervals from baseline through the end of therapy.

The methodology involved three primary modeling approaches:

Tabular Models: Standard models trained on structured clinical variables (e.g., sputum smear, culture, demographics, symptoms) updated monthly.
Large Language Model (LLM) Embeddings: A more flexible representation of the same variables, where data was converted into embeddings to capture complex, non-linear relationships.
Longitudinal and Sparse Modeling: Advanced techniques incorporating temporal dependencies and sparse variable inclusion to assess if modeling the trajectory of data points offered advantages over static monthly snapshots.

The models were evaluated for two distinct tasks: predicting EOT outcomes (cure vs. failure) and predicting post-treatment relapse among participants who achieved favorable EOT outcomes and completed follow-up. Performance was measured using ROC-AUC, with comparisons made against baseline-only predictors (specifically cavitation and sputum smear).

Key Results

End-of-Therapy (EOT) Prediction: The predictability of EOT outcomes improved significantly after month 3, achieving an ROC-AUC of up to 0.84. This improvement was primarily driven by the inclusion of sputum-smear and solid culture data.
Relapse Prediction: In contrast, the prediction of relapse among cured patients showed only modest improvement through month 3 (ROC-AUC ranging from 0.58 to 0.63) before performance declined. The strongest predictors for relapse remained age, sex, clinical symptoms, and bacterial burden.
Model Architecture Comparison:
- LLM-derived embeddings matched the performance of tabular models for relapse prediction throughout the therapy duration.
- For EOT outcome prediction, LLM embeddings outperformed tabular models at months 3 and 4, showing a $\Delta$ ROC-AUC of 0.12 and 0.14, respectively.
- Longitudinal modeling and sparse variable inclusion provided incremental improvements in relapse prediction specifically at months 4, 5, and 6 ( $\Delta$ ROC-AUC: 0.09, 0.19, and 0.12), though this came at the cost of reduced interpretability.
Risk Stratification Utility: Models incorporating data collected after baseline demonstrated incremental improvements in stratifying post-treatment relapse risk compared to using baseline cavitation and sputum smear alone. Specifically, at 4-months post-treatment, the model-derived low/high risk groups showed a relapse-free survival of 94.8%/78.5%, compared to 91.4%/82.9% for groups stratified by baseline "easy-to-treat" vs. "hard-to-treat" status.

Significance and Claims
The authors conclude that while routine clinical data collected during the treatment course can offer incremental improvements in post-treatment risk stratification, its overall predictive value for relapse remains limited. The modest performance of relapse models, even with longitudinal data and advanced embeddings, underscores a fundamental limitation in current clinical markers. Consequently, the paper posits that the field requires the development of relapse-specific biomarkers to achieve meaningful predictive accuracy, rather than relying solely on the temporal accumulation of standard clinical variables.

Technical Summary: Time-resolved predictability of end-of-therapy outcome and relapse after cure in Phase 3 tuberculosis trials

More like this