AI tools may help doctors predict walking ability shifts in PAH

Artificial intelligence (AI) tools using machine learning may help doctors predict how a pulmonary arterial hypertension (PAH) patient’s ability to walk might change over the next year, a study from Germany found.

The researchers said more study is needed before the machine learning model, which uses clinical data available at the start of treatment, can be routinely used in the clinic.

“These findings align with the growing interest in precision medicine approaches in PAH,” the pair of researchers wrote in the study, “Machine learning–based prediction of 12-month change in 6MWD in pulmonary arterial hypertension: a retrospective prediction model development study,” which was published in Scientific Reports.

PAH is a type of pulmonary hypertension in which high blood pressure develops when the blood vessels carrying blood from the heart to the lungs narrow, making it more difficult for blood to flow through the lungs and pick up oxygen. This can limit a patient’s ability to perform physical activities, including walking.

Recommended Reading

April 18, 2023 Columns by Mike Naple

The 6-minute walk test brings me a sense of familiarity and comfort

Algorithm identifies potential changes

Doctors often use the six-minute walk distance (6MWD) to diagnose PAH and assess its severity. During this test, a patient walks for six minutes on a flat, hard surface, and the total distance is recorded. It is a simple way to measure how much physical activity can be performed without exerting maximum effort. A shorter distance generally indicates more severe disease.

However, predicting how much walking ability will worsen over time remains challenging. The researchers asked whether machine learning could predict the 12-month change in a patient’s 6MWD. Machine learning uses computer algorithms to identify patterns in data and use those patterns to make predictions. The prediction was based solely on clinical information available at the start of treatment, known as the baseline.

The study involved 181 adults (117 women and 64 men) with a diagnosis of PAH who were evaluated at a single medical center between 2010 and 2022. About two-thirds (68%) had idiopathic PAH, meaning the cause of the disease was unknown. The researchers collected 92 measurements from each patient’s baseline evaluation. These included age, other diseases, PAH-specific medications, and pressure in the blood vessels of the lungs.

The main goal was to watch for relative changes in 6MWD after 12 months, calculated by subtracting the baseline distance from the 12-month distance and then dividing that difference by the baseline distance. The mean relative change was 0.25, although individual changes varied considerably, ranging from a decrease of 38% to an increase of more than 11 times from the baseline.

Six machine learning models were tested. They use different mathematical approaches to find relationships between the 92 baseline variables and later changes in walking distance. The data were divided into a training group consisting of information on 144 patients and a separate test set of 37 patients.

Within the training group, the researchers used fivefold cross-validation and grid search. Cross-validation repeatedly divides the training data into smaller groups to assess how well a model performs on data it did not use to learn. Grid search tests different model settings, called hyperparameters, to find useful combinations. The final models were then tested on the group of 37 patients.

The random forest model performed best overall. It produced a Pearson correlation — a measure of how closely predicted and actual changes move together —of 0.76,  The model accounted for about 58% of the observed variation in the test data. These findings suggest that clinical information available when a patient is first evaluated may help doctors predict how that patient’s ability to walk may change over the next year.

However, the study was small, and the data were from a single medical center. The researchers also used only baseline information; incorporating changes in measurements over time may improve predictions. The team said larger studies at multiple centers would be needed to confirm the model and determine whether it improves clinical care before it could be used in routine care.