An essential part of this work is to provide a data-driven model for predicting blood glucose levels that will help to warn the person with type 1 diabetes about a potential hypo- or hyperglycemic event in an easy-to-manage and discreet way. In this work, we apply a convolutional recurrent neural network on a real dataset of 6 contributors, provided by the University of Ohio [5]. Our model is capable of predicting glucose levels with high precision with a 30minute horizon (RMSE = 17.45 [mg/dL] and MAE = 11.22 [mg/dL]), and RMSE = 33.67 [mg/dL] and MAE = 23.25 [mg/dL] for the 60minute horizon. We believe this precision can greatly impact the long-term health condition as well as the daily management of people with type 1 diabetes.
Type 1 diabetes is a disease in which the cells responsible for insulin production are destroyed. Because insulin is the hormone that triggers absorption of glucose within the cells, people with diabetes need to monitor their glucose concentration in the blood and readjust it by frequent insulin injections, following a well-defined medical protocol (e.g., once during the day and once before each meal, to keep blood sugar levels within the normal range). The main challenge in handling diabetes is the optimization of insulin injections in order to avoid hypoglycemia and hyperglycemia. This is complicated by the fact that besides insulin intake and diet, glucose levels are also affected by several other factors such as physical activity, lifestyle, mental state, stress, etc. Despite the various accomplishments made in continuous diabetes monitoring (a.k.a. continuous glucose monitoring, CGM), such methods remain invasive. Furthermore, they are only able to provide the glycemic state at a given time, when the insulin level may already be unacceptable (too high or too low). A proactive detection could therefore dramatically improve the daily handling of diabetes by the patients themselves.
This work presents an approach based on deep learning algorithms
for predicting glucose levels in the future (30-minute and 60-minute
horizons). Our work is based on the architecture of a recurrent
neural network (CRNN) from [
The goal of using a CRNN architecture is twofold. (1) Convolutional layers act as filters and automatically learn to detect the features of interest for prediction. They are also particularly convenient for analyzing time series with little signal processing required. And (2), recurrent neural network are well-known for the capacity to learn long-term relationships between the different values. For instance, it is necessary for the network to be able to capture a correlation between the ingestion of carbohydrates now and a change in the blood glucose level in a near future.
This paper is organized as follows: Section 2 presents related work for glucose prediction, Section 3 formulates the problem of glucose prediction and our contribution, Section 4 details our methodology and discuss experimental results. Finally, in Section 5 we summarize the importance of our contribution and suggest some future work. 2 Predicting blood glucose levels for diabetes (type 1 or type 2) using machine learning has gained a lot of attention and has resulted in several methods and applications being proposed recently. They are however either based on solely measuring the glucose levels or the resulting prediction accuracy is not yet high enough to be considered as a reliable predictor of a potential critical glycemic condition. Several types of regression algorithms can be used, including SVR, classic statistical methods such as ARIMA, deep learning neural networks, or even a naive persistence algorithm, to name a few.
Gu et al. [
The aim of this work is to predict glucose levels in advance in order to avoid situations of hyperglycemia or hypoglycemia, as well as others negative effects on the health. For instance, chronic hyperglycemia may induce fatigue and vision problems among others. For that purpose, we created a model capable of predicting the glycemia of type 1 diabetes, where values must be as accurate as possible. In this context, the metrics are the RMSE and the MAE. The smaller these values, the more reliable the model. The real gain for a patient is to be able to make decisions at any given time considering the prediction of future values, and possibly avoid glycemia-related discomforts while minimizing intrusive methods. 4 4.1
Our approach can be summarized by the following steps: 1. Data importation 2. Data preprocessing 3. Implementation of the CRNN prediction model with multi-step forecasting 4. Training, testing and tuning on selected features 5. Delivery of the forecasted blood glucose levels
Data importation involves loading, merging, and aligning values from multiple sources under the same time scale. The selected features are basal/bolus insulin, carbohydrates, and blood glucose levels.
In the preprocessing step, all variables must have measurements carried over at the same time. This requires the use of subsampling or oversampling methods and, corollary, defining imputing methods. Linear interpolation is used for the glucose level on the training set to resample the time series at a frequency of 5 minutes. The others features may be imputed with null values when required, as their nature is sparse. We can also use domain-specific functions, such as an equation describing the absorption rate over time for carbohydrates. The preprocessing steps are summarized by : 1. Save all the blood glucose timestamps 2. Resample the features to a time delta of 1 second 3. Forward fill the missing values by using the last available values 4. Fill the left missing values with 0 5. Resample the features to a time delta of 5 minutes 6. Smooth each feature with a 1D Gaussian filter over a window containing the past 2 hours of data
We use the saved timestamps at preprocessing step 1 to generate the results at the same timestamps as the measured values.
Linear interpolation is used during the training process to impute the missing values. This allows to have more data points for the model to be trained on. However, we should note that linear interpolation is not ideal for big gaps of missing values. The interpolation is not used at test time as it could lead to a data peek. Meaning the predictions would be contaminated by future values.
The CRNN model target values are based on the following equation: yt+L = bgt+L bgt; for L = 1; 2; : : : ; 12 (1) where bgt is the blood glucose value at time t, L is the lag value in timesteps for the horizon, and y the label to predict, that is the differentiated value of the blood glucose level.
For instance, if the blood glucose level is 80 mg=dL at the current time and 60 mg=dL 30 minutes later, the label for a prediction horizon of 30 minutes at the current time would be 20 mg=dL.
Respectively, as the model does not predict directly the blood glucose level but only the difference from the last known value. The predicted blood glucose level is obtained with the following equation: bbgt+L = bgt + y^t+L; for L = 1; 2; : : : ; 12 (2) where bgt is the blood glucose value at time t, bbgt+L is the predicted blood glucose level at time t + L, y^t+L is the predicted blood glucose level difference at time t with lag L, L is the lag value in timesteps for the horizon, and y the label to predict. It is important to note that the CRNN only outputs the values y^t+L.
The model is capable of giving a prediction for each 5 minutes prediction horizon up to 60 minutes that is 5, 10, ..., 60 minutes. This feature may give valuable information to a user and thereby improve their blood glucose level control. An example of such a prediction is given in Fig. 5.
The overall architecture of the CRNN is based on [
The training of a model is done on a contributor-basis, that is one
model is trained per contributor. The reason is that each glucose
response is individual, and a one-population model does not seem
reasonable for CGM.
Our results reported in this work are based on the OhioT1DM dataset
[
The reported results are based on the following signals: glucose level, basal insulin, bolus insulin, and meal for 6 data contributors (540, 544, 552, 567, 584, and 596). While more signals were available, we decided to use only these signals. Indeed, we performed several tests with the complete dataset and the preliminary results indicated better results with a limited set of features.
The use of a sliding window consisting of the last 2 hour data,
resulting in 24 data points is based on [
The detailed architecture of the CRNN is presented in Table1.
Layer description Convolution 1D Max pooling 1D Convolution 1D Max pooling 1D Convolution 1D Max pooling 1D LSTM Dense Dense Dense
Output dimension (Batch size, 24, 8) (Batch size, 12, 8) (Batch size, 12, 16) (Batch size, 6, 16) (Batch size, 6, 32) (Batch size, 3, 32) (Batch size, 64) (Batch size, 256) (Batch size, 32) (Batch size, 12)
The model is pretrained on batches of size 1024 over 1000 epochs, with an RMSProp optimizer. The learning rate is initially set to 0.001 and is reduced with a factor of 0.1 when the model does not progress after 3 epochs. Early stopping is used similarly with a patience of 50 epochs in order to regularize the model. The last model’s weights with the lowest validation loss are then restored.
For each data contributor of 2020 the pretrained model is loaded and trained similarly as the pretraining stage. With the only difference that the learning rate is reduced with a patience of 15 epochs and that one model is saved for each contributor. 4.2.3
The RMSE and the MAE are calculated for the six contributors using the following equations.
v
n RM SE = tuu n1 X(y^i i=1 yi)2 (3) where y^i is the predicted value, yi the truth value and n the number of observation.
n M AE = 1 X n i=1 jy^i yij (4)
The first observation is the errors systematically increase for each contributor over time. It is not surprising that the larger the prediction window, the larger the error in general grown up.
A comparison between contributors was also performed (Fig. 2 and Fig. 3). As we can see, for example with the contributor 596, the prediction curve for 30 minutes and 60 minutes follows the real curve with an RMSE = 13.34 and MAE = 9.08 [mg/dL], and RMSE = 27.74 and MAE = 19.13 [mg/dL]. These specific curves are also detailed in Fig. 4.
The evaluation of the prediction from the contributor-side can be performed by a multi-step prediction (as illustrated in Fig. 5). Prediction is made through one forward pass of a horizon of the last 2 hours data and outputs the horizon for the next hour represented by the orange curve. In the given example, the model seems to have predicted well the tendencies.
A comparison of our approach with other algorithms is
summarized in Table 2. We notice that the RMSE of our work at horizon 30
is smallest and at 60 minutes similar to [
y^t+L = bgt 1; for L = 1; 2; : : : ; 12 (5) Prediction horizon 30 60
Metrics (mg/dL) MAE RMSE MAE RMSE (1) Li’s CRNN
NA 21:0:7 2:35
NA 33:27 4:79 (2) BASELINE Overall 18:13 0:00 25:76 0:00 30:70 0:00 42:00 0:00 (3) CRNN
With the hypothesis of a hypo-glycemia starting below 70 mg=dL and a hyper-glycemia starting above 150 mg=dL, we believe that our model delivers relevant and actionable results for real patients. In our opinion, a RMSE of 17:45 mg=dL for a 30-minute horizon indicates that data-driven decisions could be made in regard to avoiding hypo or hyper-glycemic related events. 5
In this paper we described a model for predicting future blood sugar
levels of people with type 1 diabetes. A CRNN approach was
proposed with the advantages of using only 4 different signals and very
little signal processing. The evaluation was performed using RMSE
and MAE metrics, with different horizons and on multiple
contributors. The results were compared with different algorithms.The results
report low error rates given the problematic of glucose prediction,
and in our opinion could be considered for real-world
implementation. Yet, several research tracks remain to be explored. For instance,
testing additional features that can influence blood sugar levels such
as stress or illness. We can also think of extracting manual features
from the given signals with signal processing methods, and defining
domain-specific imputation methods, such as for the absorption of
carbohydrates over time. It would be also interesting to further
personalize predictions as suggested in [
During the making of the camera ready version, we found an error in the preprocessing stage thanks to the great reviews. The missing values were treated using a linear interpolation during the testing of the model. Thereby the predictions were contaminated by future values. This error was corrected by removing the interpolation and dealing with missing values as explained in this paper version.
It is available at https://github.com/JonasFreibur/BLGP-HES-SO
The authors would like to thank the HES-SO for funding this research.