Detecting hypoglycemia-induced electrocardiogram changes in a rodent model of type 1 diabetes using shape-based clustering

Fig 1. Summary of analytical workflow. The data collection, various preprocessing techniques, unsupervised machine learning algorithm, clustering evaluation, and analysis by experimental group are detailed.

Fig 2. Internal evaluation metrics for senator and cluster identification. Calinski Harabasz (blue) and Davies Bouldin scores (red) for relevant unsupervised clustering models. Black dots represent the lowest Davies Bouldin score and highest Calinski Harabasz score. Dashed black line represents the senator and cluster identified as the optimal model. (A) Internal evaluation metrics for all models developed using 10 clusters and no times. 400 senators model was identified as having the highest Calinski Harabasz and lowest Davies Bouldin score using the elbow method. (B) Internal evaluation metrics for all models developed using 400 senators and no times. 10 clusters model was identified as having the highest Calinski Harabasz and lowest Davies Bouldin score using the elbow method.

Fig 3. Cluster centers of the optimal model. Cluster centers representing ECG heartbeats for the unsupervised clustering model with 10 clusters, 400 senators, and no times is presented. Each of the 10 cluster centers are depicted (A–J).

Fig 4. Cluster phenotypic features of the optimal model. Radar diagram representing distribution of phenotypic features for the unsupervised clustering model with 10 clusters, 400 senators, and no times is presented. The percentage of heartbeats that originated from rats with diabetes, from rats that died, that were collected during hypoglycemia, that were collected during severe hypoglycemia, that demonstrated increased PR intervals, that demonstrated increased QT intervals, and that demonstrated increased QRS intervals are depicted for each cluster (A–J).

Fig 5. Cluster variation by experimental groups. Heatmap of heartbeat distribution across each cluster and experimental group. The color bar represents the percentage of heartbeats attributed to a given experimental group within each cluster scaled by the experimental group. Experimental groups are ordered according to similarity in hierarchical clustering demonstrated by the dendrogram at the top of the heatmap. Experimental groups were heartbeats during hypoglycemia from non-diabetic rats that lived (H-ND-L), heartbeats during severe hypoglycemia from non-diabetic rats that lived (SH-ND-L), heartbeats during hypoglycemia from non-diabetic rats that died (H-ND-X), heartbeats during severe hypoglycemia from non-diabetic rats that died (SH-ND-X), heartbeats during hypoglycemia from diabetic rats that lived (H-D-L), heartbeats during severe hypoglycemia from diabetic rats that lived (SH-D-L), heartbeats during hypoglycemia from diabetic rats that died (H-D-X), and heartbeats during severe hypoglycemia from diabetic rats that died (SH-D-X).