AI Identifies Novel Predictors of TB in People With HIV

An artificial intelligence (AI) model using routinely collected data predicted subsequent development of active tuberculosis (TB), Swiss researchers reported. The AI model outperformed biological tests for latent TB in identifying HIV-positive patients at high risk of developing TB.

As well as immune function and sociodemographic variables, the AI model retained several biomarkers indicative of patients’ well-being and metabolism.

In Switzerland and other countries with good access to antiretroviral therapy, TB is a rare but serious co-infection in people living with HIV, frequently linked with late HIV diagnosis. To prevent progression to active TB disease, people known to have latent TB infection can be offered preventive treatment with isoniazid and/or rifampicin.

But detection of latent TB is challenging, especially in people with HIV. In a previous Swiss analysis, a combined approach using interferon gamma release assays (IGRA) and tuberculin skin tests identified only 3src% of people who subsequently developed active TB.

“It was worse than tossing a coin,” Joahnnes Nemeth, MD, an attending physician in the department of infectious diseases and hospital epidemiology at the University of Zürich, Zürich, Switzerland, told Medscape Medical News.

The problem is that the tests rely on immune response, which may be impaired. “You interrogate the very system that is malfunctioning during HIV infection, so it’s not a surprise that the tests perform poorly,” he explained.

This led him and his colleagues to look into alternative ways to identify patients at risk. They leveraged data from the Swiss HIV Cohort Study, which includes around 7src% of people receiving HIV care in the country.

Over 23 years’ worth of data were analyzed using machine learning, a subset of AI that enables computers to learn patterns from data and make predictions without being explicitly programmed for each task. Their machine learning model employed a random forest — an algorithm which combines the outputs from multiple decision trees.

The model looked at data collected at HIV diagnosis in order to predict active TB disease that developed at least 6 months later. Rather than only considering variables which the researchers thought were potential risk factors, the model reviewed all the variables for which they had sufficient data.

“What I really liked about this machine learning approach is that we threw all the data we collect into the machine and just asked it: Can you do something with that?” Nemeth said. “I think that really paid off.”

The first iteration of the model included 48 variables and had a sensitivity of 7src.1% and a specificity of 81.src%. A streamlined second version retained 2src variables — making it computationally less demanding — while delivering a sensitivity of 57.1% and specificity of 77.8%.

Given that biologic tests had a sensitivity of 3src% and specificity of 94%, for Nemeth this “blows everything of the water.” The model doesn’t require additional data collection or have the expense of IGRA.

As might be expected, the 2src retained variables included immunological parameters, hematological markers, and sociodemographic factors, but some were more surprising: along with several variables linked with metabolism (cholesterol, high-density lipoprotein, glucose, and creatinine), body mass index, and mean arterial pressure.

The researchers noted that TB is associated with malnutrition and said that some of these markers may reflect metabolic perturbations and compromised muscle mass in people at risk for TB.

The model was first validated on a portion of the Swiss cohort which it was not trained on, and then on a cohort in Austria. Despite the many parallels between the two cohorts, initially the model performed badly in Austria.

The researchers realized the issue stemmed from different migration patterns between the countries: Most people with TB in Switzerland have moved from sub-Saharan Africa, while in Austria, most come from the former Soviet republics. Only after modifying the ethnicity and region of birth variables did the model begin to work effectively.

“This is a cautionary tale,” said Nemeth. “You go to a very similar setting with a little difference, and all this stops working. With machine learning models, we really have to be careful and test them vigorously before we rely on them.”

Emily Wong, MD, is an associate professor at the University of Alabama at Birmingham who has used AI to aid interpretation of chest radiography in South Africa, but was not involved in the new study.

The Swiss research “opens one’s eyes to the idea that with very large data sets with lots of clinical variables, you can discern meaningful and predictive patterns that predict whether someone will go on to develop TB,” she told Medscape Medical News. 

Nemeth is working on an implementation study in which physicians whose patients have never been tested for TB will be randomly allocated to either receive a reminder to test, or a risk score based on the machine learning model. A key question is whether the latter will be enough to convince physicians to take further action, such as offering preventative therapy.

Wong noted that the potential benefits and risks (including liver toxicity) of preventative therapy need to be weighed up for each patient. But a machine learning model could help clinicians to do this.

“The idea that in the future, based on key demographic and clinical information of a person, and maybe including their chest x-ray or IGRA test, or maybe not, we would have a well-functioning clinical decision making tool that would guide a health care worker to make TB prevention decisions for the patient in front of them is definitely a worthy goal,” she said.

The study was funded by the Swiss National