Using the Power of Machine Learning to Hone in on HF Subtypes

The study covered in this summary was published in medRxiv.org as a preprint and has not yet been peer-reviewed.

Key Takeaways

Machine learning identified five subtypes in incident heart failure (HF) with good predictive accuracy for 1-year all-cause mortality.
Future trials and clinical practice might be able to use the electronic health record (EHR) and machine learning to identify HF subtypes to aid in management and prognosis.

Why This Matters

This is the first study to define and validate HF subtypes across multiple machine-learning methods, nationally representative datasets, and multiple validation methods.
The study's structured framework of internal, external, prognostic, and genetic validation could extend acceptability and generalizability of machine learning to clinical practice and is transferable to other diseases.

Study Design

The study involved 313,062 patients 30 years and older with incident HF and at least 1 year of follow-up in The Health Improvement Network (THIN) and Clinical Practice Research Datalink (CPRD).
EHRs from the two primary care populations were linked with hospital admissions and a death registry in the United Kingdom.
UK Biobank data from 9573 patients were used for genetic validation.
HF subtypes were identified using four unsupervised machine-learning methods with 87 (from 645) factors (demography, history, examination, blood laboratory values, and medication) included in the algorithm.
Validation was done internally, externally, prognostically, and genetically.
A HF clustering model and open-access app were developed for use in routine practice.

Key Results

Five clusters were identified and labeled as subtypes: early-onset, late-onset, atrial fibrillation (AF)-related, metabolic, and cardiometabolic.
The distribution of subtypes 1 to 5 was similar across THIN (16.5%, 30.9%, 8.9%, 14.0%, and 29.7%) and CPRD (12.6%, 35.6%, 9.3%, 13.8%, 28.7%).
Age varied across subtypes, with the oldest patients having late-onset HF and the youngest having early-onset HF.
There were also differences by sex, with the most women in the metabolic subtype and the least in the cardiometabolic group.
Cardiovascular risk factors and diseases — such as hypertension (72.9%), obesity (34.3%), diabetes (41.1%), and atherosclerosis (59.2%) — were highest in the cardiometabolic subtype.
In CPRD, 1-year mortality was 2%, 46%, 6%, 11%, and 37% for subtypes 1 to 5, respectively, with C-statistic of 0.68, 0.62, 0.57, 0.71, and 0.68, respectively.
Differences in mortality between THIN and CPRD were found for clusters 2 and 5 only.
Hypertension, myocardial infarction, stroke, and peripheral vascular disease generally occurred before HF diagnosis in the cardiometabolic subtype and after HF diagnosis in the AF-related and early-onset subtypes.
Polygenic risk scores for atrial arrhythmias, diabetes, hypertension, myocardial infarction, obesity, and stable and unstable angina were all associated with one or more HF subtypes.
Eight single-nucleotide polymorphisms were nominally associated with predicted HF subtype (P = .05), of which four were associated with the AF-related subtype.
Five clinicians questioned in the study stated that the results had clinical utility and that the app would be useful.

Limitations

The investigators used EHR phenotypes of HF, which do not have complete biochemical and imaging profiles and, therefore, some classifications could not be done.
The risk factor phenotypes relied on the timing and accuracy of the clinicians' notes in the EHR.
Both datasets used are from the United Kingdom, which might not be representative of HF in other places.
The polygenic risk scores were only for 11 traits and 12 single-nucleotide polymorphisms.

Disclosures

Amitava Banerjee is supported by research funding from the National Institute for Health Research (NIHR), British Medical Association, AstraZeneca, and UK Research and Innovation.
Benoit Tyl and Tomasz Dyszynski are employees of Bayer. All other authors declared no competing interests.

This is a summary of a preprint research study, Identifying subtypes of heart failure with machine learning: external, prognostic and genetic validation in three electronic health record sources with 320,863 individuals, written by Amitava Banerjee, BMBCh, DPhil, from University College London, and colleagues, on MedRxiv provided to you by Medscape. This study has not yet been peer-reviewed. The full text of the study can be found on medRxiv.org.

Comments

Commenting is limited to medical professionals. To comment please Log-in.

Comments on Medscape are moderated and should be professional in tone and on topic. You must declare any conflicts of interest related to your comments and responses. Please see our Commenting Guide for further information. We reserve the right to remove posts at our sole discretion.