Blog Post

Led by Sriram Chandrasekaran, Associate Professor of Biomedical Engineering, and Awanti Sambarey, a postdoctoral fellow, a group of experts from the University of Michigan made a multimodal AI model that can guess how well TB patients will react to treatment.

The goal of this study is to improve TB treatment outcomes and clinical care, especially for drug-resistant cases, by looking at a lot of multimodal tuberculosis (TB) patient data from the NIAID TB Portals database. Researchers looked at data from 5,060 TB cases in 10 countries with a lot of multidrug-resistant TB (MDR-TB). They used clinical, radiological, microbiological, and genomic data, among other sources.

Findings and methods that are important:

Multimodal machine learning model: Using random forests, the researchers developed a combined machine learning model that outperformed models built on separate types of data. The area under the curve (AUC) for this multimodal model was 0.84, which means it was 83.2% accurate at predicting how treatment would go. Cross-validation and external validation on new patient data were used to make sure the model was correct. It had 203 features from different types of data.

2. Drug regimen analysis: The study found certain drug combos that are linked to successful or unsuccessful treatment for different types of drug-resistant TB. The treatment plan of “bedaquiline, clofazimine, cycloserine, levofloxacin, and linezolid” was strongly linked to successful TB treatment for MDR non-XDR TB. Alternatively, the mix of “bedaquiline, clofazimine, linezolid, and moxifloxacin” was connected to the same type of TB not responding to treatment.

Scores for drug interactions: The INDIGO-MTB tool was used by the researchers to figure out scores for drug interactions for different combos. They discovered that treatment results were better when drugs were mixed in ways that worked well together (lower interaction scores). This result shows how important it is to think about drug combinations when making effective TB treatment plans.

4. Imaging features: Looking at chest X-rays and CT scans showed some imaging features that were strongly linked to treatment results and drug resistance. A lot of these were abnormal lung volumes, bronchial obstructions, pleuritis, and certain patterns of nodules and holes in different parts of the lungs.

5. Sociodemographic and clinical factors: The study found that a patient’s BMI, employment status, education level, age at onset, and the presence of comorbidities like HIV and anemia were highly linked to how well they responded to treatment.

6. Genetic aspects of the pathogen: The researchers looked at changes in Mtb genes that are linked to drug resistance. They discovered that changes in genes such as gyrA, inhA, katG, rpoB, and rpsL that happened together were strongly connected to bad treatment results.

Uses in real life of the research:

1. Individualized treatment planning: The multimodal prediction model created in this study can be used to figure out how likely it is that a treatment will not work for each patient. Clinicians can get a more accurate prediction and make treatment plans based on it by entering a patient’s clinical, radiological, and pathogen data. This tailored method might lead to better results, especially in cases of drug-resistant TB that are hard to treat.

2. Finding the best drug combinations: Figuring out which drug combinations are linked to treatment success or failure is very helpful for doctors when they are looking for the best drug combinations for MDR-TB and XDR-TB patients. For instance, the study shows that for treating MDR-TB, levofloxacin-based regimens are better than moxifloxacin-based ones.

3. Using data on drug interactions: The discovery that using drugs that work well together leads to better results can help create new treatment plans. When making treatment plans, doctors and academics can use tools like INDIGO-MTB to guess how drugs will interact with each other and choose combinations that work best together.

Radiologists and clinicians can better understand chest X-rays and CT scans of TB patients thanks to the study’s results on imaging features linked to treatment outcomes and drug resistance. This could lead to more accurate early readings of how bad the disease is and how resistant it is to drugs, making it easier to start the right treatment at the right time.

Five. Risk stratification: The societal, demographic, and clinical factors that have been linked to bad results can be used to make tools for assessing risk. These tools might help doctors find patients who are at high risk and might need more close tracking or extra care while they are being treated.

Targeted public health interventions: The study’s results on how important BMI, work, and education are to treatment outcomes can be used to help plan public health programs. To improve the general success rates of TB treatment, money can be put toward addressing these social factors that affect health.

10. Genomics monitoring: Looking at Mtb genomic data can help make surveillance of drug-resistant TB strains better. This data can be used to keep track of how resistant strains spread and help shape plans for controlling tuberculosis at the regional and global levels.

8. Systems for helping doctors make decisions: The multimodal prediction model could be built into systems for keeping track of health records to help doctors make decisions. Clinicians would be able to use real-time, data-driven information to make choices about treatment.

Research Prioritization: The results of the study can help direct future work on TB. For example, the unexpected link between some drug combos and failed treatment calls for more research and could lead to the creation of better treatment plans.

10. Managing TB around the world: Because the study looked at statistics from several countries with a lot of MDR-TB, its results can be used in many ways to help control TB around the world. The methods and insights can be changed and used in many different types of healthcare situations around the world.

Finally, this in-depth multimodal study of TB patient data gives us a lot of useful information for making TB treatment work better. By combining different kinds of patient and pathogen data, the study gives us a more complete picture of what makes a medicine work or not work. This study has real-world uses, from helping individual patients to making global public health plans. It could make a big difference in the World Health Organization’s goal of ending the tuberculosis epidemic by 2035.