Machine learning improves patient selection for CRT

23Oct - by aiuniverse - 0 - In Machine Learning


A novel machine learning algorithm improved patient selection for cardiac resynchronization therapy (CRT) in a study of nearly 1,000 heart failure patients in the U.S., representing an opportunity to optimize care and spare certain individuals from a pricey procedure that might not benefit them.

CRT is indicated in patients with medically refractory systolic HF and left ventricular dyssynchrony to improve left ventricular ejection fraction (LVEF), corresponding author Charlotta Lindvall, MD, PhD, and co-authors explained in PLOS ONE earlier this month. Improvement in LVEF after CRT implantation is associated with improved survival and reduced HF hospitalizations, but at least one-third of CRT patients don’t see that improvement by 18 months post-op.

Right now, medical guidelines suggest patient selection for CRT implantation should depend on a number of factors, including NYHA functional class, LVEF, QRS duration, type of bundle branch block, etiology of cardiomyopathy and atrial rhythm. Physicians are also urged to consider patients’ “general health status,” but consensus statements offer little guidance on how to go about evaluating these CRT candidates.

“Accurate patient selection is important to minimize morbidity and mortality related to the device, and to control healthcare costs,” Lindvall, of the Dana-Farber Cancer Institute in Boston, and colleagues said. “It will only become more important as the population of heart failure patients continues to grow. However, more than a decade of work has shown that it is not easy to identify new predictors of CRT response.”

The authors said advancements in AI and machine learning offer a new opportunity for patient selection for CRT. If integrated with natural language processing, machine learning could make use of both structured and unstructured EHR data to build a precise, usable prediction tool.

Lindvall et al. built their AI model by applying machine learning and natural language processing to EHRs of 990 patients who received CRT at two academic hospitals between 2004 and 2015. Demographics, lab values, medication status, clinical characteristics and medical history were extracted from EHRs available before the CRT procedure, and the researchers also extracted bigrams from patients’ clinical notes using natural language processing. Patients accrued, on average, 75 clinical notes.

The team built a machine learning model using 80% of the patient sample and tested the model on the remaining 20% of the study pool. The average age of participants was 72 years old, and mean baseline LVEF was 24.8%. The authors’ primary endpoint was reduced CRT benefit, defined as less than 0% improvement in LVEF at six to 18 months post-procedure or death at 18 months.

Of the 990 patients studied, 403—40.7%—saw a reduced benefit from their CRT device within a year and a half. Just over a quarter of patients saw no improvement in LVEF by 18 months, and 15.6% had died by then.

Lindvall and colleagues’ finalized model identified 26% of patients who wouldn’t benefit from CRT implantation at a positive predictive value of 79%.

“The amount of data available in the EHR is massive, and is rapidly expanding,” the authors wrote. “Analysis of these data using methods from computer science can allow for discovery of complex patterns that are clinically important, but difficult for the human mind to identify. Machine learning does not require prior assumptions about causative variables and allows for an exploration of all available data for non-linear patterns.”

The team said validated models that can run in the background of an EHR could enable recognition of many marginal risk factors which, on their own, might not seem significant. The approach could allow for more individualized risk assessment and, in time, might change our current guidelines-based approach to patient selection.

“Clinicians often encounter patients with demographic and clinical characteristics that differ from the patients who participated in the studies that formed the evidence basis for the guidelines,” Lindvall et al. wrote. “Thousands of CRT procedures are performed in the United States alone every month and so opportunities for refined decision support tools should be pursued.”

Facebook Comments