Researchers deploy AI tool to predict disease progression in PAH
Model helps identify patients at low, medium, or high risk of clinical worsening
Written by |
Using artificial intelligence (AI) analyses of clinical data from hundreds of patients, scientists have created a new system to calculate the risk of future clinical worsening in people with pulmonary arterial hypertension (PAH).
“To our knowledge, this is the first study to use machine learning to leverage existing [electronic health record] data to predict the risk of clinical worsening events among patients with PAH,” the scientists noted.
They said this study “adds to the growing body of literature that describes ways of applying machine-learning and artificial intelligence approaches to utilize real-world, [electronic health record], and [insurance] claims-based data to predict clinical outcomes.”
The study, “Use of Machine-Learning Models to Identify Clinical Features Associated With A Future Clinical Worsening Event in Patients With Pulmonary Arterial Hypertension,” was published in Pulmonary Circulation. The work was funded by Johnson & Johnson, and most of the study co-authors are employees of J&J or the AI company Nference.
Model predicts which PAH patients will experience clinical worsening
PAH is a chronic disorder marked by abnormally high pressure in the vessels that carry blood through the lungs. One of the tricky things about clinical care for people with PAH is that it can be hard to predict which individuals with the disease are at greatest risk of major health issues in the near future.
Machine learning is a form of AI that basically works by feeding a computer a large set of data, along with mathematical algorithms that the computer uses to identify patterns in the data. Over the last few years, researchers all over the globe have been exploring how this type of cutting-edge computer analysis may help predict future clinical outcomes for people with a range of chronic disorders.
Here, scientists applied this strategy to identify factors associated with clinical worsening events in people with PAH. For purposes of the analysis, a clinical worsening event was defined as a hospitalization due to any cause, worsening symptoms, initiation of certain types of treatment, lung or heart-lung transplant, or death.
The analysis included data on 455 people with PAH who received care through the Mayo Clinic system. Most of these patients were white, female, and diagnosed with PAH when they were older than 55. Over a median follow-up time of slightly less than one year, just over half of these patients had a documented clinical worsening event, most commonly hospitalization due to any cause.
These findings demonstrate the feasibility of using routinely collected clinical data to explore predictors of short-term disease progression.
The best-performing machine learning model predicted clinical worsening events with a sensitivity of 77.9% and specificity of 58.5%. In other words, the algorithm accurately identified more than three-quarters of patients who experienced such an event, and just over half of those who did not.
Statistical analyses of the machine learning model indicated that the strongest predictors of clinical worsening events were body mass index (BMI) and red cell distribution width (RDW). BMI is a ratio of weight to height, whereas RDW assesses the variability in size of red blood cells (the cells that carry oxygen through the bloodstream). Patients with more clinical visits within a one-month window also had higher rates of clinical worsening events.
Building off their model, the researchers devised a simply calculated risk prediction score to divide PAH patients into three groups: those with low, medium, or high risk of future events. Most (74%) patients in the low-risk group did not have any worsening events, whereas almost all (82%) of those in the high-risk group did. In the medium-risk group, roughly half of the patients experienced worsening events.
“These findings demonstrate the feasibility of using routinely collected clinical data to explore predictors of short-term disease progression,” the researchers concluded.
The scientists noted that although their model shows promise, the study has noteworthy limitations, including the relatively homogenous demographics of patients and the fact that data were based on records from just one healthcare system. As such, they stressed a need for further work to validate and expand upon their findings.
“Automated risk prediction using [electronic health record] data may represent a clinically useful tool in the future; however, further validation in independent cohorts and additional implementation research are necessary before such approaches can be incorporated into routine PAH management,” the scientists wrote.
