Data Fusion of Activity and CGM for Predicting Blood Glucose Levels* Hoda Nemat1 and Heydar Khadem1 and Jackie Elliott2 and Mohammed Benaissa1 Abstract.2 This work suggests two methods—both relying on minutes. Zhu generated a dilated deep convolutional neural network stacked regression and data fusion of CGM and activity—to predict fed by CGM, insulin, and carbohydrate intake as inputs. Xie applied the blood glucose level of patients with type 1 diabetes. Method 1 an autoregression with exogenous inputs approach to predict BGL uses histories of CGM data appended with the average of activity by exploiting current and past information of CGM data. data in the same histories to train three base regressions: a multilayer Physical activity is a critical factor in diabetes management. perceptron, a long short- term memory, and a partial least squares regression. In Method 2, histories of CGM and activity data are used Therefore, investigation of the activity data in BGL prediction separately to train the same base regressions. In both methods, the models is encouraged [13]. However, developing models with high predictions from the base regressions are used as features to create a accuracy using activity and CGM data is challenging, and limited combined model. This model is then used to make the final studies have been done in this area. Data fusion of activity and CGM predictions. The results obtained show the effectiveness of both data normally result in models with a performance not comparable methods. Method 1 provides slightly better results. with those using CGM alone. This paper proposes two novel CGM and activity data fusion methods to generate BGL prediction models with performance 1 INTRODUCTION comparable with those using CGM data alone. The literature emphasises the importance of the management of type 1 diabetes mellitus (T1DM) in reducing complications associated with the disease [1], [2]. The key role in T1DM management is to 2 DATASET control blood glucose level (BGL) to remain in a normal range [3], To develop BGL prediction algorithms, we used the OhioT1DM [4]. dataset [14]. The dataset contains eight weeks’ worth data of 12 The prediction of BGL from current and past information can be people with T1DM. The data of six patients was released in 2018 for a useful contributor [5]. BGL prediction could provide early the first BGL prediction challenge [15] and data for additional six warnings concerning inadequate glycaemic control to prevent the patients (referred by ID 540, 544, 552, 567, 584, and 596) was occurrence of an adverse glycemic status [6], [7]. released for the second BGL prediction challenge in 2020 [14]. In BGL prediction models could be classified into three main this work, we used the data of the latter six patients. groups: physiological models, data-driven models, and hybrid The dataset includes data of CGM sensor, physical activity band, models. Data-driven models explain the relationship between the physiological sensor, and self-reported life-event. Among the present and past information to BGL prediction. In this regard, different collected data, we explored CGM and activity data which machine learning and time series approaches have been widely used were collected every 5 and 1 minutes, respectively. Detailed [5]. information about the sensors and devices as well as characteristics Many studies have proposed data-driven BGL prediction of the patients has been published [14], [15]. methodologies. Mirshekarian et al. [8], Bertachi et al. [9], In the dataset, there are three types of activity data consisting of Martinsson et al. [10], Zhu et al. [11] and Xie et al. [12] in separate galvanic skin response, skin temperature, and magnitude of studies, developed prediction models to forecast BGL with a acceleration. In this work, we only used the data of the magnitude of prediction horizon of up to 60 minutes. acceleration. Hereafter, for simplicity, ‘magnitude of acceleration’ Mirshekarian’s model was based on a recursive neural network is referred to as ‘activity’. (RNN), which utilised long short- term memory (LSTM) units. CGM, insulin, meal, and activity information were inputs of their model. Bertachi used physiological models of insulin, carbohydrate, 3 METHODOLOGY and activity on board to train an artificial neural network (ANN). This section presents the information about data preprocessing and Martinsson proposed an RNN model trained on historical blood the methodologies developed for the prediction of BGL. glucose information to predict BGL in two horizons of 30 and 60 This paper is submitted to the second Blood Glucose Level Prediction Challenge 2020, the 5th International Workshop on Knowledge Discovery in Healthcare Data. 1 Department of Electronic and Electrical Engineering, University of Sheffield, UK, email addresses: hoda.nemat@sheffield.ac.uk, h.khadem@sheffield.ac.uk, m.benaissa@sheffiels.ac.uk 2 Department of Oncology and Metabolism, University of Sheffield, UK, email address: j.elliott@sheffield.ac.uk Copyright © 2020 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). 3.1 Preprocessing hidden layers used ReLU as the activation function. Mean squared error was the loss function, Adam was the optimiser. The model Missing data in the training set is imputed using linear interpolation. trained with 100 epochs with a learning rate of 0.01. For the testing set, on the other hand, linear extrapolation is used. • Partial least squares regression (PLSR) This is to assure that future data is not seen by the model, and that PLSR carries considerable popularity in different applications, such the model can be used for a real-time application. Thus, we convert as glucose sensing [18]. In this work, PLSR was applied as a CGM and activity data to regular time series without any missing regression tool. Different values were considered for the number of data in 5-minute and 1-minute intervals, respectively. components—ranging from one to the length of the input window. The next step was to unify the resolution of CGM and activity Each time, the predicted residual sum of squares (𝑃𝑅𝐸𝑆𝑆) was data. To do so, we downsampled the activity time series data to 5- calculated as follows. minute intervals by capturing the nearest activity data to each CGM data and discarding the rest. 𝑁 There were a considerable number of unavailable activity data at 𝑃𝑅𝐸𝑆𝑆 = ∑(𝑦 − 𝑦̂𝑖 )2 (1) the beginning and/or end of training and/or test set. This was due to 𝑖=1 the difference in wear time of CGM and activity sensors. For these points average of activity data in the training set is used rather than Where, N is the size of the evaluation set, and 𝑦𝑖 is the reference linear interpolation or extrapolation. Table 1 shows the number of value, and 𝑦̂𝑖 is the predicted value. unavailable activity data for each patient ID. The number of components (𝐴) resulting in the minimum value for 𝑃𝑅𝐸𝑆𝑆/(𝑁 − 𝐴 − 1) is then selected [19]. Table 1. The number of non-existent activity data points in training and testing sets per data contributor. Patient ID testing set training set 3.2.2 Stacked regression 540 547 31 Stacked regression is applied to enhance the performance of BGL 544 0 125 prediction [20]. This technique uses predictions from a number of 552 622 505 models—first-level models—as features to train a new model— 567 0 108 second-level model. In this work, a stacked regression structure was 584 3 123 employed where the three base regressions mentioned in 2.3.1 were 596 80 18 set as its first-level models and a PLSR as the second-level model (Figure 1). Another data preprocessing step was to reframe a time series problem to a supervised learning task. To this end, time series data were transformed into samples with lag observations as input and First-level models future observations as output. We use a rolling window with a 𝑌෠1 Second-level model MLP history length of 6 or 12 data points for the input, which has the information of 30- or 60- minute history, respectively. Also, the 𝑌෠2 Stacking of Final Training PLSR LSTM output of each sample is a vector with 6 or 12 data points 𝑌෠3 𝑌෠1 , 𝑌෠2, and 𝑌෠3 prediction 𝑌෠ set corresponding to prediction horizons of 30- and 60- minute, PLSR respectively. Figure 1. Diagram of the developed stacked regression. 3.2 Regression tools Three base regressions and a stacked regression technique are used as tools to develop the final prediction models. 3.3 Prediction methods We developed two different methods using the stacked regression 3.2.1 Base regressions structure mentioned above to fuse CGM and activity data. Using these methods, models were then created to predict BGL of each • Multilayer perceptron (MLP) patient for both horizons of 30 and 60 minutes. For each prediction MLP [16] is an ANN that can be used for time series forecasting. In horizon, two histories of 30 and 60 minutes were tried for training this work, a single-hidden-layer MLP model was used. The model purposes. comprised a dense layer of 100 nodes with an activation function of rectified linear unit (ReLU) followed by an output layer. Adam and mean absolute error were used as an optimiser and a loss function, 3.3.1 Method 1 respectively. The learning rate was 0.01, and the model was fitted with 100 epochs. This method used the average value of activity data added to the • Long short-term memory (LSTM) window of CGM data to train the first-level models. RNN is also an artificial neural network suitable for working with sequential data. We used a vanilla LSTM recurrent network [17] 3.3.2 Method 2 with vector output which is used for multi-step ahead forecast. The model was composed of a hidden layer with 200 units followed by a In this method, the first-level models were trained twice. Once using fully-connected layer with 100 nodes and an output layer. Both a history of CGM data, and once using a history of activity data, thus producing six first-level models rather than three. Table 3. Evaluation results of the first-level models of Method 1 using a history of 30 minutes. PH: 30 min PH: 60 min Patient ID Model RMSE MAE RMSE MAE PLSR 22.13 16.60 41.09 31.74 540 MLP 21.96 ± 0.29 16.46 ± 0.21 40.53 ± 0.38 30.95 ± 0.33 LSTM 21.22 ± 0.12 15.82 ± 0.08 39.65 ± 0.28 30.38 ± 0.28 PLSR 18.08 13.33 31.80 24.71 544 MLP 17.95 ± 0.07 12.87 ± 0.13 31.61 ± 0.32 24.27 ± 0.71 LSTM 17.62 ± 0.20 12.60 ± 0.32 30.79 ± 0.29 23.02 ± 0.67 PLSR 16.76 12.77 30.23 23.67 552 MLP 16.96 ± 0.19 12.69 ± 0.21 30.38 ± 0.36 23.42 ± 0.61 LSTM 16.44 ± 0.17 12.18 ± 0.22 29.89 ± 0.47 22.53 ± 0.40 PLSR 20.97 15.04 37.41 28.15 567 MLP 21.44 ± 0.63 15.60 ± 0.76 37.96 ± 1.45 29.01 ± 1.35 LSTM 20.61 ± 0.20 14.64 ± 0.32 36.36 ± 0.31 27.08 ± 0.43 PLSR 22.07 16.21 36.85 27.85 584 MLP 21.60 ± 0.12 15.61 ± 0.14 36.54 ± 0.74 27.27 ± 0.89 LSTM 21.55 ± 0.26 15.58 ± 0.27 36.75 ± 1.69 27.62 ± 2.08 PLSR 17.79 12.76 29.63 22.05 596 MLP 18.01 ± 0.16 12.99 ± 0.17 29.75 ± 0.69 21.93 ± 0.38 LSTM 17.23 ± 0.17 12.25 ± 0.29 29.17 ± 0.22 21.29 ± 0.32 PLSR 19.63 14.45 34.50 26.36 Average MLP 19.65 ± 0.24 14.37 ± 0.27 34.46 ± 0.66 26.14 ± 0.71 LSTM 19.11 ± 0.19 13.85 ± 0.25 33.77 ± 0.54 25.32 ± 0.70 Where 𝑦𝑖 , 𝑦̂𝑖 , and N have the same meaning as in (1). 3.4 Evaluation In the Ohio dataset, the last 10 days’ worth of data for each 4 RESULTS AND DISCUSSION contributor was allocated as the testing set and the rest as training [14]. To train and evaluation purposes, we used the training and In this section, the results of RMSE and MAE for prediction testing sets, respectively. Extrapolated data and, the first 60 minutes models are provided for both prediction horizons of 30 and 60 of the test set was excluded when calculating the evaluation metrics. minutes. Models with a performance dependent on random The latter is because the testing set starts immediately after the initialisation ran five times, and the mean and standard deviation of training set, and they are chronologically close to each other. results are reported. We have used the acronym PH for the prediction Summarised statistics of the testing set for each patient is given in horizon in the tables. Table 2. Table 2. The statistics of the patients’ testing set. Original Imputed Evaluation 4.1 Method 1 Patient ID data point data point data point Table 3 displays the evaluation results of the first-level models of 540 2896 3066 2884 Method 1 when a history of 30 minutes is used for training. Based 544 2716 3136 2704 on the RMSE and MAE values, in both prediction horizons, LSTM 552 2364 3950 2352 had the best prediction performance for all patients except 584. For 567 2389 2871 2377 this patient, MLP had the best result. PLSR, as a simple linear 584 2665 2995 2653 regressor, produced results comparable to the non-linear neural 596 2743 3003 2731 network models. Root mean square error (RMSE) and mean absolute error (MAE) Table 4. Evaluation results of the second-level model of Method 1 using a were calculated as follows and considered as evaluation metrics. history of 30 minutes. PH: 30 min PH: 60 min Patient ID RMSE MAE RMSE MAE ∑𝑁 (𝑦𝑖 − 𝑦̂𝑖 )2 𝑅𝑀𝑆𝐸 = √ 𝑖=1 (2) 540 21.19 ± 0.07 15.73 ± 0.09 39.41 ± 0.09 30.04 ± 0.15 𝑁 544 17.40 ± 0.08 12.45 ± 0.08 30.48 ± 0.07 22.90 ± 0.08 552 16.25 ± 0.07 12.02 ± 0.05 29.32 ± 0.09 22.21 ± 0.02 567 20.40 ± 0.07 14.44 ± 0.07 36.12 ± 0.02 27.12 ± 0.07 ∑𝑁 𝑖=1|𝑦𝑖 − 𝑦 ̂𝑖 | 584 21.54 ± 0.06 15.62 ± 0.06 36.27 ± 0.15 27.17 ± 0.16 𝑀𝐴𝐸 = (3) 𝑁 596 17.17 ± 0.10 12.13 ± 0.09 28.77 ± 0.26 20.80 ± 0.17 Average 18.99 ± 0.08 13.73 ± 0.07 33.39 ± 0.12 25.04 ± 0.11 Table 5. Evaluation results of the first-level models of Method 1 using a history of 60 minutes. PH: 30 min PH: 60 min Patient ID Model RMSE MAE RMSE MAE PLSR 22.10 16.58 41.10 31.76 540 MLP 21.58 ± 0.28 16.12 ± 0.22 40.53 ± 1.23 31.12 ± 0.91 LSTM 21.11 ± 0.18 15.56 ± 0.11 39.18 ± 0.37 30.00 ± 0.33 PLSR 18.09 13.33 31.83 24.71 544 MLP 18.09 ± 0.03 13.05 ± 0.08 32.34 ± 1.00 24.80 ± 1.76 LSTM 18.04 ± 0.35 13.06 ± 0.48 30.79 ± 0.39 23.15 ± 0.68 PLSR 16.79 12.78 30.25 23.67 552 MLP 17.58 ± 0.46 13.39 ± 0.70 30.16 ± 0.43 22.89 ± 0.14 LSTM 16.97 ± 0.78 12.59 ± 0.55 30.69 ± 0.70 23.19 ± 0.55 PLSR 20.99 15.03 37.51 28.21 567 MLP 21.71 ± 0.92 15.80 ± 1.06 37.34 ± 0.78 28.02 ± 0.76 LSTM 20.74 ± 0.50 14.75 ± 0.59 36.67 ± 0.98 27.52 ± 1.06 PLSR 22.04 16.19 37.04 27.97 584 MLP 22.10 ± 0.25 15.98 ± 0.23 37.13 ± 0.74 27.68 ± 0.89 LSTM 21.66 ± 0.10 15.63 ± 0.12 36.76 ± 0.46 27.18 ± 0.44 PLSR 17.62 12.66 29.48 21.97 596 MLP 18.05 ± 0.29 12.71 ± 0.27 29.71 ± 0.35 21.83 ± 0.21 LSTM 17.58 ± 0.19 12.55 ± 0.34 29.55 ± 0.52 21.63 ± 0.34 PLSR 19.60 14.43 34.53 26.38 Average MLP 19.85 ± 0.37 14.51 ± 0.43 34.54 ± 0.75 26.06 ± 0.78 LSTM 19.35 ± 0.35 14.02 ± 0.36 33.94 ± 0.57 25.44 ± 0.57 Table 4 shows the evaluation results of the second-level model of Method 1 when a history of 30 minutes was used for training. Comparing these results with those in Table 3, the second-level 4.2 Method 2 model resulted in better prediction performance than all the first- In this section, the evaluation result of Method 2 is presented. To level models for all patients and both prediction horizons. This be concise, the results of the second-level model only are reported, means that the stacked regression technique helped improve which are the final predictions of the method. prediction performance. Table 7 shows the evaluation results of Method 2 using a 30- Table 5 displays the evaluation results of the first-level models of minute history. Comparing these results with those in Table 4, the Method 1, when a history of 60 minutes was used for training. As prediction performance of Method 2 was comparable with that of results show, for both prediction horizons, LSTM had the best Method 1 for all patients, except patient 552. This may be due to the performance for a majority of the patients. In overall, PLSR existence of a large number of missing activity data points in this provided the second-best results. patient’s data (as can be seen in Table 1). The evaluation results of the second-level model of Method 1 using 60-minute history are shown in Table 6. In comparison with Table 7. Evaluation results of Method 2 using a history of 30 minutes. Table 5, it can be observed that the stacked regression technique PH: 30 min PH: 60 min advanced the prediction performance for all patients for this history, Patient ID RMSE MAE RMSE MAE too. Also, in comparison with Table 4, Method 1 had a better overall 540 21.26 ± 0.09 15.89 ± 0.07 39.48 ± 0.16 30.26 ± 0.19 performance when it used a history of 30 minutes than a history of 544 17.59 ± 0.11 12.62 ± 0.12 30.68 ± 0.15 23.14 ± 0.20 60 minutes. 552 19.85 ± 4.51 12.65 ± 0.46 35.70 ± 3.32 23.76 ± 0.40 567 20.52 ± 0.12 14.49 ± 0.12 36.39 ± 0.20 27.14 ± 0.19 Table 6. Evaluation results of the second-level model of Method 1 using a 584 21.72 ± 0.17 15.78 ± 0.10 36.53 ± 0.13 27.45 ± 0.08 history of 60 minutes. 596 17.24 ± 0.11 12.19 ± 0.07 28.83 ± 0.11 21.03 ± 0.13 PH: 30 min PH: 60 min Patient ID Average 19.70 ± 0.85 13.94 ± 0.16 34.60 ± 0.68 25.46 ± 0.20 RMSE MAE RMSE MAE 540 20.98 ± 0.13 15.50 ± 0.14 39.05 ± 0.17 29.68 ± 0.18 Table 8 lists the evaluation result of Method 2 using a history of 544 17.66 ± 0.09 12.66 ± 0.08 30.42 ± 0.36 22.82 ± 0.42 60 minutes. Comparing these results with those in Table 6, the 552 16.30 ± 0.09 12.04 ± 0.06 29.38 ± 0.24 22.26 ± 0.21 evaluation results for both methods were close to each other. Also, 567 20.52 ± 0.17 14.54 ± 0.10 36.52 ± 0.10 27.31 ± 0.14 comparing these results with those in Table 7, Method 2 made better 584 21.62 ± 0.17 15.63 ± 0.08 37.01 ± 0.28 27.64 ± 0.20 predictions using a history of 60 minutes than a history of 30 596 17.45 ± 0.08 12.27 ± 0.09 28.92 ± 0.27 20.92 ± 0.19 minutes. Average 19.09 ± 0.12 13.77 ± 0.09 33.55 ± 0.24 25.11 ± 0.23 Table 8. Evaluation results of Method 2 using a history of 60 minutes. 5, p. e11030, 2019. PH: 30 min PH: 60 min [6] J. Vehí, I. Contreras, S. Oviedo, L. Biagi, and A. Bertachi, Patient ID “Prediction and prevention of hypoglycaemic events in type-1 RMSE MAE RMSE MAE diabetic patients using machine learning,” Health Informatics J., p. 540 20.89 ± 0.05 15.49 ± 0.11 39.30 ± 0.35 29.80 ± 0.21 1460458219850682, 2019. 544 17.70 ± 0.14 12.68 ± 0.13 30.71 ± 0.22 23.25 ± 0.29 [7] C. Berra et al., “Hypoglycemia and hyperglycemia are risk factors 552 16.73 ± 0.51 12.33 ± 0.18 34.67 ± 3.51 23.47 ± 0.58 for falls in the hospital population,” Acta Diabetol., vol. 56, no. 8, 567 20.57 ± 0.14 14.63 ± 0.11 36.70 ± 0.30 27.48 ± 0.18 pp. 931–938, 2019. 584 21.72 ± 0.06 15.71 ± 0.05 36.85 ± 0.09 27.69 ± 0.13 [8] S. Mirshekarian, R. Bunescu, C. Marling, and F. Schwartz, “Using LSTMs to learn physiological models of blood glucose behavior,” 596 17.53 ± 0.21 12.26 ± 0.18 28.88 ± 0.21 21.02 ± 0.17 Proc. Annu. Int. Conf. IEEE Eng. Med. Biol. Soc. EMBS, pp. 2887– Average 19.19 ± 0.18 13.85 ± 0.13 34.52 ± 0.78 25.45 ± 0.26 2891, 2017. [9] A. Bertachi, L. Biagi, I. Contreras, N. Luo, and J. Vehí, “Prediction of Blood Glucose Levels And Nocturnal Hypoglycemia Using 5 SUMMARY AND CONCLUSION Physiological Models and Artificial Neural Networks.,” in 3rd International Workshop on Knowledge Discovery in Healthcare This work contributes to the prediction of BGL by proposing two Data, 2018, pp. 85–90. methodologies for data fusion of CGM and activity using stacked [10] J. Martinsson, A. Schliep, B. Eliasson, C. Meijner, S. Persson, and regression. O. Mogren, “Automatic blood glucose prediction with confidence In the first method, the average value of activity data added to a using recurrent neural networks,” 3rd Int. Work. Knowl. Discov. Healthc. Data, vol. 2148, pp. 64–68, 2018. window of CGM data was used as input to train prediction models. [11] T. Zhu, K. Li, P. Herrero, J. Chen, and P. Georgiou, “A Deep Initially, three base regression models consist of MLP, LSTM, and Learning Algorithm for Personalized Blood Glucose Prediction.,” PLSR were trained. Subsequently, predictions from these base in 3rd International Workshop on Knowledge Discovery in models were used as features to train a new PLSR model which then Healthcare Data, 2018, pp. 64–78. made final predictions. [12] J. Xie and Q. Wang, “Benchmark Machine Learning Approaches with Classical Time Series Approaches on the Blood Glucose In the second method, the same base regressions were trained Level Prediction Challenge.,” in 3rd International Workshop on once using windows of activity data and once using CGM data. The Knowledge Discovery in Healthcare Data, 2018, pp. 97–102. predictions of all trained base models were then fed as features to a [13] M. H. Jensen, C. Dethlefsen, P. Vestergaard, and O. Hejlesen, new PLSR model for its training process. The new PLSR was used “Prediction of Nocturnal Hypoglycemia From Continuous Glucose to make refined predictions. Monitoring Data in People With Type 1 Diabetes: A Proof-of- Concept Study,” J. Diabetes Sci. Technol., vol. 14, no. 2, pp. 250– The results obtained show that Method 1 (average value of 256, 2020. activity data added to the window of CGM data) had a slightly better [14] C. Marling and R. Bunescu, “The OhioT1DM Dataset for Blood performance than Method 2 (first-level models trained twice, once Glucose Level Prediction: Update 2020,” in 5th International with a history of CGM data, once using a history of activity data). In Workshop on Knowledge Discovery in Healthcare Data, 2020. overall, Method 1 using a history of 30 minutes had the best results [15] C. Marling and R. C. Bunescu, “The OhioT1DM Dataset For Blood Glucose Level Prediction.,” in 3rd International Workshop by providing a RMSE of 18.99 and 33.39 for the prediction horizon on Knowledge Discovery in Healthcare Data, 2018, pp. 60–63. of 30 minutes and 60 minutes, respectively. [16] F. Murtagh, “Multilayer perceptrons for classification and regression,” Neurocomputing, vol. 2, no. 5–6, pp. 183–197, 1991. [17] S. Hochreiter and J. Schmidhuber, “Long short-term memory,” 6 SOFTWARE AND CODE Neural Comput., vol. 9, no. 8, pp. 1735–1780, 1997. [18] H. Khadem, M. R. Eissa, H. Nemat, O. Alrezj, and M. Benaissa, To implement the models, we used Python 3.6, TensorFlow 1.15.0 “Classification before regression for improving the accuracy of and Keras 2.2.5. Also, Pandas, NumPy and Sklearn packages of glucose quantification using absorption spectroscopy,” Talanta, python were used. The codes were run on a commodity laptop. The vol. 211, 2020. codes of our implementation are available at: [19] S. Wold, M. Sjöström, and L. Eriksson, “PLS-regression: a basic tool of chemometrics,” Chemom. Intell. Lab. Syst., vol. 58, no. 2, https://gitlab.com/Hoda-Nemat/data-fusion- pp. 109–130, 2001. stacking.git [20] L. Breiman, “Stacked regressions,” Mach. Learn., vol. 24, no. 1, pp. 49–64, 1996. REFERENCES [1] G. S. Jeha et al., “Continuous glucose monitoring and the reality of metabolic control in preschool children with type 1 diabetes,” Diabetes Care, vol. 27, no. 12, pp. 2881–2886, 2004. [2] L. S. Schilling et al., “A new self-report measure of self- management of type 1 diabetes for adolescents,” Nurs. Res., vol. 58, no. 4, p. 228, 2009. [3] E. R. Seaquist et al., “Hypoglycemia and diabetes: A report of a workgroup of the american diabetes association and the endocrine society,” J. Clin. Endocrinol. Metab., vol. 98, no. 5, pp. 1845– 1859, 2013. [4] K. Makris and L. Spanou, “Is there a relationship between mean blood glucose and glycated hemoglobin?,” J. Diabetes Sci. Technol., vol. 5, no. 6, pp. 1572–1583, 2011. [5] A. Z. Woldaregay, E. Årsand, T. Botsis, D. Albers, L. Mamykina, and G. Hartvigsen, “Data-driven blood glucose pattern classification and anomalies detection: machine-learning applications in type 1 diabetes,” J. Med. Internet Res., vol. 21, no.