Background:
The heart and kidneys have vital functions in the human body that reciprocally influence each other physiologically. Pathological changes in 1 organ can damage the other. Epidemiologic studies show that greater than 50% of patients with heart failure (HF) have preserved ejection fraction (HFpEF). Additionally, 1 in 6 patients identified as having chronic kidney disease (CKD) also has HF. Thus, it is important to be able to predict and identify the cardiorenal relationship between HFpEF and CKD.

Objective:
Creating an electrocardiogram (ECG)-enabled model that stratifies suspected patients with HFpEF would help identify CKD-enriched HFpEF clusters and phenogroups. Simultaneously, a minimal set of significant ECG features derived from the stratification model would aid precision medicine and practical diagnoses due to being more accessible and widely readable than a large set of clinical inputs. Furthermore, the validation of the existing cardiorenal relationship using this ECG-enabled model may lead to better biological understanding.

Methods:
Using unsupervised clustering on all extractable ECG features from FinnGen, patients with an indication of HFpEF (filtered by left ventricular ejection fraction [LVEF] values ≥50% and N-terminal pro B-type natriuretic peptide [NT-proBNP] values >450 pg/mL) were categorized into different phenogroups and analyzed for CKD risk. After isolating significant predictive ECG features, unsupervised clustering and risk analysis were performed again to demonstrate the efficacy of using a minimal set of features for phenogrouping. These clusters were then compared to clusters formed using dynamic time warping (DTW) on raw ECG time series electrical signals. Afterward, these clusters were analyzed for CKD enrichment.

Results:
The PR interval and QRS duration stood out as significant features and were used as the minimal feature set. After generating and comparing clusters (k-means with all extracted ECG features, k-means with a minimal feature set, and DTW with full lead II ECG waveform), the DTW-generated clusters were most stable. ANOVA analysis also showed that several HFpEF clusters exhibited a deviation of CKD risk from baseline, allowing for further trajectory analysis. Specifically, the creatinine levels (a proxy for CKD) of several DTW-created clusters showed significant differences from the average. Based on the Jaccard score, the DTW clusters also showed the greatest alignment to baseline comparison clusters created by clustering on creatinine. In comparison, the other 2 sets of clusters (created by all extracted ECG features and the minimal set) performed similarly.

Conclusions:
This project validates both the known cardiorenal relationship between HFpEF and CKD and the importance of the PR interval and QRS duration. After exploring the use of ECG data for patient clustering and stratification, DTW clustering with lead II waveforms resulted in the most clinically meaningful clusters in the context of HFpEF and CKD. This methodology may prove useful in exploring ECG clustering applications outside of HFpEF as well.