In recent years, the proliferation of biosensors and wearable devices has facilitated the ability to perform continuous monitoring of physiological signals. In diabetes management, this has come with the increasing use of continuous glucose monitoring (CGM) devices for helping with glucose control. The current literature on clinical impact of CGM devices shows that continuously monitoring blood glucose concentration levels has benefit in maintaining tight glycaemic control [5,2]. As a next step, glucose prediction offers an opportunity to further improve glucose control by taking actions to avert adverse glycaemic events, such as suspension of insulin delivery in closedloop systems to avert hypoglycaemia.
The general work in this area has typically involved collecting data covering physiological variables such as glucose concentration levels, heart rate, and self-reported data covering exercise,sleep, stress, illness, insulin, and meals. However, public datasets covering ambulatory monitoring of T1DM population are not widely available.
Deep learning [6] facilitates learning the optimal features and has been shown to perform better than other methods involving hand crafted features that have been employed in recent times for predicting glucose concentration levels. However, typically these models require relatively large amounts of data to converge on an appropriate model.
In this work, we employ a multitask learning [1] approach in order to improve the performance of the glucose forecasting in a neural network, where each individual is viewed as a task, using shared layers to enable learning form other individuals.
1 Imperial College London, United Kingdom, email: jsd111@imperial.ac.uk, p.herrero-vinias@imperial.ac.uk, pantelis@imperial.ac.uk
Glucose prediction has been a long-standing area of focus in the diabetes community. As a result, many approaches have existed in order to provide near-time glucose concentration level forecasts. Early work in this area have focused on physiological models and traditional machine learning methods in predicting glucose concentration levels [12,3]. Recent work as seen in the 2018 Blood Glucose Predictive Challenge has seen a move towards deep learning methods with more impressive results [11,9,14,8]. These have used convolutional architectures, recurrent architectures, or a combination of both to model the task of glucose prediction.
In this section, we detail the transformations that are performed on the data prior to training and testing the model for each T1DM participant.
The OhioT1DM dataset 2020 [10] is a dataset comprising 12 unique participants that cover eight weeks of daily living. The participants are given IDs as the data is anonymised. This data comprises physiological data gathered using a continuous glucose monitor (blood glucose concentration levels) and wristband device (heart rate, skin conductance, skin temperature), activity data (acceleration, step count), and self-reported data (meal intake, insulin, exercise, work, sleep, and stressors).
A non-trivial aspect of the datasets used for developing glucose prediction models is the aspect of missingness. This is evident in the Ohio T1DM dataset with missingness present in both physiological variables and self-reported data [4].
Linear Interpolation: The blood glucose values that are missing in this dataset are typically missing at random. This could be attributed to issues around replacing glucose sensors and/or transmitters, or dealing with faulty communication. As a result, we employ linear interpolation in the training set to handle imputation of missing blood glucose concentration levels in the dataset over a period of one hour. In the samples where more than an hour of CGM data is missing the sample is discarded from the training set. This is illustrated with an example sequence in (C) of Fig. 1 On the other hand, features which comprise self-reported data the assumption is made that any missing values represent an absence of said feature. Therefore all missing values in insulin, meal intake and reported exercise are imputed with zero.
The missingness in features from the self-reported data in the testing set is tackled similarly as in the training set. However, this is not the case for blood glucose concentration levels as interpolation when a current value at a given timestep is missing would lead to an inaccurate evaluation of model performance.
Extrapolation: In order to accurately evaluate the performance of the model we cannot always rely on interpolation at test time as this may require, in a real-time setting, an unknown future value to perform interpolation. Consequently, we need to rely on other methods of extrapolation to impute the missing glucose concentration levels. In this scenario (A), for gaps of data less than 30 minutes, we impute missing values with predicted values from the trained model. For missing recent values longer than 30 minutes as in (B), we pad the remaining values with the last computed value. In cases where, a gap larger than 30 minutes is evident in historical data and a current value is present at the given timestep, linear interpolation was then employed instead to provide a more accurate imputation.
To enable training the proposed model effectively, we perform transformation of the relevant input features (blood glucose concentration, insulin bolus, meal(carbohydrate) intake, and reported exercise). The blood glucose concentration levels are scaled down by a factor of 120. Similarly, the insulin bolus is scaled by 100 and meal intake values are scaled by 200 in the same range between features. The exercise values are transformed to a simple binary representation of the presence or absence of exercise, from the recorded exercise intensity on a range from 1-10.
In this section we detail the machine learning technique that is used to provide the means of learning personalised models with the entire dataset. We detail the approach to develop the deep multitask network for personalisation. We provide a summary of the hyperparameters used in training as well and setting up the input for personalised multitask learning.
Multitask learning is an approach in machine learning that can be broadly described as a method of learning multiple tasks simultane-ously with the aim of improving generalisation [1].
Multitask learning for personalisation has been used mainly in affective computing [13] with early work in diabetes management focusing on using multitask learning for developing prediction models for clustered groups of Type 1, Type 2, an non-diabetic participants [7] rather than leveraging similarities within groups such as gender, for personalised glucose predictions.
As seen Figure 2, the output from the shared layers are now fed into the individual(task)-specific fully connected layers of each user.
In a multitask setting of this kind, a multiplicative gating approach is used to ensure that the input corresponding to the particular user trains on just that user in the individual-specific layers. In that sense, at each iteration a batch that consists of data from a particular individual is used to train the shared layers and the layers specific to the individual.
The deep learning model trained in the multitask learning setting is a convolutional recurrent neural network (CRNN) proposed by Li et. al [8] to perform short-term glucose prediction. This forms the basis of the single-task (STL) model. The convolutional recurrent model consists initially of a 3 temporal convolutional layers that perform a 1-D convolution with a Gaussian kernel over the sequence of input to extract features of various rates of appearance, followed by a max pooling layer after each convolution operation. The input is a 4-dimensional sequence that takes a 2-hour window of historical data.
The output from the convolutional layers performs feature extraction and feeds into a recurrent long short-term memory (LSTM) layer that is able to better model the temporal nature of the task.
The output from the shared layers feed into the fully connected layers of each user and to then provide the change in glucose value over the prediction horizon. This is then added to the current glucose value to provide the forecast glucose concentration level.
The loss function used for converging to the appropriate model for the glucose forecasting is the mean absolute error. This is expressed below as:
Figure 2. A detailed look at the formulation of convolutional recurrent networks in a multitask setting. In this setting, each user is represented as a task. In addition, the initial layers (convolutional and recurrent layers) are shared between each user, the next two (dense) layers are shared based on gender, and the last (dense) layer is specific to each user.
where ŷ denotes the predicted results given the historical data and y denotes the reference change in glucose concentration over the relevant glucose prediction, and N batch refers to the batch size.
The following table details provides the details of the hyperparameters used for the model architecture at each layer. The optimiser used for this work is Adam. The learning rate is 0.0053. The model is trained for 200 epochs. This value was obtained through grid search optimisation.
The model is developed on Keras 2.2.2, with a Tensorflow 1.5 backend. The training is performed on an NVIDIA GTX 1050 GPU.
The repository for the code accompanying the paper can be found at: https://github.com/jsmdaniels/ecai-bglp-challenge
The model is tested on data from six participant IDs: 540, 544, 552, 567, 584, 596.
The evaluation of the model is based on two metrics: root mean square error (RMSE) and the mean absolute error (MAE). The extrapolated points are not considered in calculating these metrics. The formulation of these metrics are provided below:
where ŷ denotes the predicted results given the historical data and y denotes the reference glucose measurement, and N refers to the data size.
In order to undertake a comprehensive evaluation of the model performance, the subsequent criteria for assessment are followed:
• Performance evaluation over 30-minute and 60-minute prediction horizon (PH): The RMSE and MAE for each participant is analysed for a the same length of values for both prediction horizons.
The performance of the multitask learning (MTL) approach is evaluated in the context of comparison with the performance of a single task learning (STL) approach which uses only patient specific data.
• Multiple runs for each participant ID: The multitask CRNN (MTCRNN) model uses randomly initialised weights at the start of training. Given the variable nature of this training procedure, the results reported are the average of 5 model runs.
The unit for results reported below is mg/dL. The best performance is in bold. 3 and 4 exhibit the differences in performance as seen in the specific window for participant 596. The increased lag and reduced predictive performance can also be attributed to the higher chance of external activities (insulin, meals, exercise) that influence the blood glucose trajectory occurring over the prediction horizon.
The best predictive performances were achieved by the model with IDs 544, 552, 596 whereas, IDs 540, 567, and 584 exhibited worse performances over both 30 and 60 minute prediction horizons. An investigation of the glycaemic variability, using the coefficient of variation (CV) [2], of the training set of the former set of participants are stable (CV≤36%) whereas the latter group are labile (CV>36%). The multitask learning approach definitively performs better over the single task approach over a 30-minute prediction horizon. However, the performance improvement of the MTL approach over a 60-minute prediction is not consistent across each participant and metric.
One potential issue with multitask learning is the issue of negative transfer. This can be described as a scenario in which one or more of the tasks (individuals) or sampled batches during training are not strongly correlated, degrading the learning in the shared layers, and subsequently the performance at test time.
In this work, we have presented a multitask convolutional recurrent neural network that is capable of performing short-term personalised predictions -19.79±0.06mg/dL (RMSE) and 13.62±0.05mg/dL (MAE) at 30 minutes, as well as 33.73±0.24mg/dL (RMSE) and 24.54±0.15mg/dL (MAE) at 60 minutes. We work towards leveraging population data while still learning a personalised model. In the future, we hope to address further challenges such as negative transfer during learning that could improve the accuracy of individual models. This approach would enable more accurate models to be deployed in the face of limited personal data.
This work is supported by the ARISES project (EP/P00993X/1), funded by the Engineering and Physical Sciences Research Council.
As seen in Table 3, the results shown provide a comprehensive evaluation of the model predictive performance. Evidently, the model performance at PH = 30 minutes is better than the model performance at PH = 60 minutes, given that prediction at 60 minutes is a more complex task than prediction at 30 minutes.