A novel few shot learning derived architecture for long-term HbA1c prediction

Abstract

Regular monitoring of glycated hemoglobin (HbA1c) levels is important for the proper management of diabetes. Studies demonstrated that lower levels of HbA1c play an essential role in reducing or delaying microvascular difficulties that arise from diabetes. In addition, there is an association between elevated HbA1c levels and the development of diabetes-related comorbidities. The advanced prediction of HbA1c enables patients and physicians to make changes to treatment plans and lifestyle to avoid elevated HbA1c levels, which can consequently lead to irreversible health complications. Despite the impact of such prediction capabilities, no work in the literature or industry has investigated the futuristic prediction of HbA1c using current blood glucose (BG) measurements. For the first time in the literature, this work proposes a novel FSL-derived algorithm for the long-term prediction of clinical HbA1c measures. More importantly, the study specifically targeted the pediatric Type-1 diabetic population, as an early prediction of elevated HbA1c levels could help avert severe life-threatening complications in these young children. Short-term CGM time-series data are processed using both novel image transformation approaches, as well as using conventional signal processing methods. The derived images are then fed into a convolutional neural network (CNN) adapted from a few-shot learning (FSL) model for feature extraction, and all the derived features are fused together. A novel normalized FSL-distance (FSLD) metric is proposed for accurately separating the features of different HbA1c levels. Finally, a K-nearest neighbor (KNN) model with majority voting is implemented for the final classification task. The proposed FSL-derived algorithm provides a prediction accuracy of 93.2%.

Figure 1

High-level overview of proposed HbA1c prediction methodology.

Figure 2

(a) Illustration of BG trend in time-series format (b) Concatenation of 7 days of CGM data.

Figure 3

Schematic diagram illustrating the procedure of estimating missing CGM data points using the nearest neighbors method.

Figure 4

The training and validation error recorded while training the neural layer used to estimate missing BG data points.

Figure 5

Model architecture for few-shot learning-based feature extraction and fusion for HbA1c prediction.

Figure 6

The PSD features comparison among six HbA1c classes.

Figure 7

Illustration of the transformation of CGM time-series data into a binary image.

Figure 8

A histogram image that reflects the frequency-distribution of each of 14 days of CGM data.

Figure 9

The feature extractor takes an image as input and produces a feature vector of a specified size. In this work, we specify the size of binary and histogram feature vectors to be 4096 features.

Figure 10

The binary and histogram image separators are trained to separate images representing the CGM time series of patients of different HbA1c classes.

Figure 11

The ten classes of CIFAR10 dataset.