Precision Medicine and Artificial Intelligence: A Pilot Study on Deep Learning for Hypoglycemic Events Detection based on ECG

Abstract

Tracking the fluctuations in blood glucose levels is important for healthy subjects and crucial diabetic patients. Tight glucose monitoring reduces the risk of hypoglycemia, which can result in a series of complications, especially in diabetic patients, such as confusion, irritability, seizure and can even be fatal in specific conditions. Hypoglycemia affects the electrophysiology of the heart. However, due to strong inter-subject heterogeneity, previous studies based on a cohort of subjects failed to deploy electrocardiogram (ECG)-based hypoglycemic detection systems reliably. The current study used personalised medicine approach and Artificial Intelligence (AI) to automatically detect nocturnal hypoglycemia using a few heartbeats of raw ECG signal recorded with non-invasive, wearable devices, in healthy individuals, monitored 24 hours for 14 consecutive days. Additionally, we present a visualisation method enabling clinicians to visualise which part of the ECG signal (e.g., T-wave, ST-interval) is significantly associated with the hypoglycemic event in each subject, overcoming the intelligibility problem of deep-learning methods. These results advance the feasibility of a real-time, non-invasive hypoglycemia alarming system using short excerpts of ECG signal.

Conflict of interest statement

The authors declare no competing interests.

Figures

Figure 1

Proposed CNN based system illustrating the study objectives. To detect the low glucose levels using the ECG signal, we set three objectives: (OBJ 1) was to build a classifier (using a CNN network) for the low glucose levels detection task. Secondly, the chosen method for performing the classification (i.e., CNN) enables to investigate further the learned representation of the input heartbeats (OBJ 2), representation (embedding) that can be used in for data visualisation/clustering in lower-dimensional space. The method used for the nonlinear dimension reduction is t-SNE. The third objective (OBJ 3) was to investigate the important regions in the input time series (the heartbeat signal) that contribute the most to the final classification result (Grad-CAM method).

Figure 2

Proposed CNN + RNN system for low blood glucose detection in a 5-minute window of time. The individual heartbeats were firstly isolated, then grouped into 5-minute segments. Each considered 5-minutes segment was chosen if it contained at least 200 heartbeats. This condition also implies that the glucose event (low/normal) should last for at least 5 minutes, thus each 5-minute ECG segment was associated with a single label: low/normal glucose. Each heartbeat was firstly transformed into a feature representation using a CNN network, representation that was fed as input to the sequence model (RNN cells). The outputs of the final RNN are the inputs to a linear layer with a softmax producing a distribution P over the two possible outputs: normal or low glucose values.

Figure 3

Hypoglycemia detection during the night using the heartbeat majority voting in the 10-minute window of time. The black waveform represents the glucose values recorded by the CGM, considered as ground truth glucose level in this study. The grey shaded regions illustrate a ± 10% error boundary for the CGM glucose readings, as it has been reported in previous studies. The colour of the points indicates the predicted class: red for the predicted low-glucose levels and green for the predicted normal-glucose levels. Moreover, dark colours indicate more certain predictions: dark red points accounted for low-glucose predictions with the predicted probability > 0.7, while light red accounted for low-glucose prediction with predicted probability ≤ 0.7; dark green accounted for normal-glucose prediction with predicted probability > 0.7 and light green accounted for normal-glucose prediction with a probability ≤ 0.7. Images a and b present the glucose levels predictions for a sample training day, while columns c and d present the glucose predictions for the same sample test day. The missing heartbeats are due to one of the following reasons: participant removing the sensor, artifact due to movement; missing glucose data, ECG pre-processing.

Figure 4

Identification of the most relevant heartbeat segments for hypoglycemia detection using Grad-CAM method. The solid lines represent the mean of all the heartbeats that correspond to each subject in the training dataset: green during normal glucose levels, red during hypoglycemic events. The comparison among 4 different subjects highlighted the fact that each subject may have a different ECG waveform during hypoglycemic events for instance Subjects 1 and 2 present a visibly longer QT interval during hypoglycemic events, differently from subjects 3 and 4. The error bands represent the standard deviation of the considered heartbeats. The vertical bars represent the histograms of the sample points that were >0.9 in the normalised heatmaps obtained from applying Grad-CAM methods on all the training heartbeats.

Figure 5

t-SNE visualisation of the test heartbeats corresponding to subject 3 in the activation space representation. The red heartbeats correspond to a low glucose level (