Predicting Progression Patterns of Type 2 Diabetes using Multi-sensor Measurements

Abstract

Type 2 diabetes – a prevalent chronic disease worldwide – increases risk for serious health consequences including heart and kidney disease. Forecasting diabetes progression can inform disease management strategies, thereby potentially reducing the likelihood or severity of its consequences. We use continuous glucose monitoring and actigraphy data from 54 individuals with Type 2 diabetes to predict their future hemoglobin A1c, HDL cholesterol, LDL cholesterol, and triglyceride levels one year later. We use a combination of convolutional and recurrent neural networks to develop a deep neural network architecture that can learn the dynamic patterns in different sensors’ data and combine those patterns with additional demographic and lab data. To further demonstrate the generalizability of our models, we also evaluate their performance using an independent public dataset of individuals with Type 1 diabetes. In addition to diabetes, our approach could be useful for other serious and chronic physical illness, where dynamic (e.g., from multiple sensors) and static (e.g., demographic) data are used for creating predictive models.

1. Introduction

Type 2 diabetes (T2D) is one of the leading causes of death and disability worldwide (Braunwald, 2019). In T2D, cells cannot respond normally to insulin, the hormone that regulates blood glucose (Kahn, Cooper, & Del Prato, 2013). As a result, individuals with T2D may experience higher or lower than safe blood glucose levels, increasing their risk for serious health consequences including heart disease and stroke. The prevalence of T2D is steadfastly increasing – with the number of T2D patients globally expected to surpass 450 million in 2030 (Cho et al., 2018). Behavioral interventions are the foundation of T2D management (e.g., diet and lifestyle modifications). In this context, predicting diabetic progression can inform T2D management. The goal of diabetic progression prediction is to predict one or more diabetes-related parameters (e.g., hemoglobin A1c). A reliable approach to modeling T2D progression prediction can help foresee potentially dangerous health consequences such as cardiovascular and kidney disease, stroke, neuropathy, and eye complications. These predictions can inform diabetes management, thereby reducing the likelihood of such health consequences.

T2D progression trends reflect a sometimes obscure and complicated set of determinants that should be inspected before designing a predictive model. These determinants consist of a broad range of parameters and different types of physical conditions that are generally known as T2D risk factors. These risk factors include, but are not limited to, age, ethnicity, body mass index, diet, physical activity, and family history. While most of these risk factors appear to be independent of one other, there exist both clear and complex correlations and interactions among them (Wu, Ding, Tanaka, & Zhang, 2014). Due to this complex nature of T2D, advanced techniques incorporating multiple risk factors and their interactions are necessary for predicting progression of this disease. Advanced machine learning (ML) methods such as deep learning algorithms offer powerful tools for learning complex and hidden associations among T2D risk factors and for predicting disease progression among individuals with T2D (Kavakiotis et al., 2017). The use of such methods is generally warranted when a multi-variable and complex model is desired, and for creating such models, it is crucial to have a dataset capable of reflecting many potential causal factors into the learning model. One limitation of existing T2D prediction models is that they use a limited number of variables, and many of these models rely on static data like weight, age, and lab records from patients, but neglect more advanced and dynamic data such as patients’ physical activity patterns (Kavakiotis et al., 2017). Even when physiological data (e.g., blood glucose or physical activity) are included, it is mostly from only one of the mentioned variables (Georga, Protopappas, Polyzos, & Fotiadis, 2015; J Simon, Schrom, Castro, Li, & J Caraballo, 2013; Namayanja & Janeja, 2012). These limited models do not consider all the valuable information for T2D progression that can enhance long-term predictions. In designing a comprehensive model for T2D progression, considering various potentially effective parameters of disease progression is a necessity. Our proposed model includes an automated platform that can reveal the underlying hidden correlations among various physiological signals and offers a generalized framework for the predictions. Specifically, the proposed model in this paper relies on physiological measurements collected using two types of wearable devices: 1) continuous glucose monitoring (CGM), which continuously measures and records blood glucose levels, and 2) actigraphy, which records raw acceleration data for tracking physical activity and sleep. Both devices are relatively common sensors that can easily be worn. We investigate how the wearable signal interactions can inform our prediction tasks. In a multi-sensor dataset, three types of signal correlations can be considered: 1) intra-sensor interactions, 2) inter-sensor interactions, and 3) temporal patterns. Intra-sensor interactions refer to the interactions among the measurements of a single sensor in different dimensions. For example, measured values of acceleration in three dimensions of physical space have dependencies with each other. These dependencies can determine the intensity and type of physical activity (e.g., walking, running, sitting) in an interval. Since low physical activity is a risk factor for T2D (Hamasaki, 2016), extracting these local correlations can help build the learning model. Inter-sensor interactions refer to those interactions among measurements of different sensors. For example, the dependencies between blood glucose levels and physical activity intensity may carry some information about T2D risk (Riddell & Perkins, 2009). Finally, extracting the temporal patterns of the recorded measurements can be informative for the learning model. In this study, in addition to the dynamic wearable data, there are other types of information (lab test results and physical indices) that are considered static. Our model uses these data as a supplement to the wearable signal measurements to enhance the prediction performance.

In the present study, we build a model that predicts four important indices of T2D progression. The first index is hemoglobin A1c (HbA1c) or glycated hemoglobin, which is a type of hemoglobin that bonds to the glucose molecules. HbA1c is considered the primary indicator for the T2D status and severity (Khan, Sherwani, Ekhzaimy, Masood, & Sakharkar, 2016). The second index is high-density lipoprotein (HDL) or ‘good’ cholesterol, which tends to decrease in individuals with T2D. Low HDL cholesterol is a risk factor for heart disease (Haase, Tybjærg-Hansen, Nordestgaard, & Frikke-Schmidt, 2015). The third index is low-density lipoprotein (LDL) cholesterol or ‘bad’ cholesterol. The last index is triglycerides, a type of fat in the blood. These indices are important for individuals with T2D as they can serve as useful indicators of T2D complications (Bitzur, Cohen, Kamari, Shaish, & Harats, 2009). Specifically, the main contributions of our work include: 1) presenting a comprehensive predictive model of T2D progression using static and dynamic data, and 2) achieving a high accuracy for predicting HbA1c as among the most important sequalae of T2D severity with potential clinical implications.

2. Related Work

ML-based approaches have been widely used to design T2D progression models. Most of the existing models predict certain T2D outcomes depending on the stage of the disease. However, these models differ in their methodology and the data used. One common approach is to extract the most important features in T2D progression using feature selection techniques. For instance, a wrapper method has been used to identify an optimal set of T2D predictors (Bagherzadeh Khiabani et al., 2015). With a similar modeling approach, a variant of the random forest method was employed to evaluate the accuracy of different features in predicting blood sugar (Georga et al., 2015). In another related study, unsupervised association rule mining was used to derive the connections among different types of diabetes risk factors (J Simon et al., 2013). ML-based techniques have been also used to model the likelihood that an individual without the disease will be diagnosed with the disease in the future. This problem has been mostly considered as a classification problem in the literature. For example, a survey study reviewed the presented prediction models in six separate studies (Kavakiotis et al., 2017). The methods for these six studies included logistic regression, naive Bayes, linear discriminant analysis, random forest, artificial neural networks, and support vector machines (SVM). The results of the review suggest that using the same set of features, SVM obtains the best prediction performance as compared to the other techniques. In another work, a method based on the random forest was presented (Casanova et al., 2016). The authors showed a higher robustness of their method against overfitting. Using a Bayesian scoring algorithm, the space of events for a diabetes patient was modeled in another study (Anderson et al., 2016). Also, in a comprehensive study, a rotation forest was employed as an ensemble method to combine 30 ML algorithms for diagnosing diabetes (Ozcift & Gulten, 2011). Recently, few studies have also studied the T2D progression in a regression (versus classification) setting. These studies (Rahimloo & Jafarian, 2016) (Kopitar, Kocbek, Cilar, Sheikh, & Stiglic, 2020) mostly address the question of early detection rather than the progression prediction (Kopitar et al., 2020).

One of the main concerns regarding the progression of T2D is its associated health consequences. Accordingly, detecting the short- and long-term complications of T2D has been the focus of many diabetes modeling studies. This body of work has studied the complex relations between diabetes and other health conditions such as high blood pressure, obesity (Felber & Golay, 2002), Alzheimer’s disease (Nicolls, 2004), cancer, obstructive sleep apnea, and eye damage (Orgel & Mittelman, 2013). For example, Sudharsan, Peeples, and Shomali (2014) used SVM and random forest to predict the occurrence of hypoglycemia (low blood glucose) among T2D patients, and Ibrahim et al. (2015), in a prevention-focused study, presented a combination of fuzzy logic classifiers and a clustering method to classify diabetic eye damage scenarios.

Recently, deep learning has received a lot of attention in T2D modeling. Among the earlier studies in this area was the work by Mohebbi et al. (2017) that experimented several deep learning methods to identify the occurrence of T2D based on a collection of CGM datasets. They achieved their best performance using convolutional neural networks (CNNs). CNN-RNN (RNN: recurrent neural network) architectures have been also popular in the field and were utilized for multivariate temporal data learning. For instance, a variation of these models in the form a CNN-LSTM (LSTM: long short-term memory) deep learning network was proposed for learning from heart rate variability (HRV) data to diagnose diabetes by Swapna, Vinayakumar, and Soman (2018).

One critical limitation of prior research on diabetic prediction on modeling is that only a few models have used multi-modal data for predicting the progression of this disease. The closest study to ours is a recent study on type 1 diabetes (Akbari & Chunara, 2019) that used blood glucose data and contextual information to predict future disease status. However, the model presented in this past work does not aim to predict the progression pattern of T2D and does not consider other physiological measures.

Extracting short- and long-term patterns from wearable sensors data generally involves using methods for studying time-series datasets and working with different data modalities. For instance, the “A-Wristocracy” method used hand-engineered features from the acceleration sensor data in a deep neural network for identifying human complex activities (Vepakomma, De, Das, & Bhansali, 2015). Various versions of RNNs have been also widely used in pattern extraction from time-series signals. For instance, RNN and LSTM networks were used in activity and gait recognition studies (Hammerla, Halloran, & Plötz, 2016) (Guan & Plötz, 2017). Recently, applying the wearable sensors data in modeling diabetes using deep learning has gained researchers’ attention. These studies mostly use measured CGM values as the time-series input of the model (Cappon, Acciaroli, Vettoretti, Facchinetti, & Sparacino, 2017) (Ashiquzzaman et al., 2018). DiabDeep (Yin, Mukadam, Dai, & Jha, 2019) bypassed the feature extraction advantages and used the wearable measurements directly. More recently, (Faruqui et al., 2019) applied the self-reported CGM measurements of patients in an LSTM network, capable of predicting future blood glucose values in a multi-step process.

In this work, we aim to utilize more than just CGM values as the used wearable sensor measurements in the deep learning prediction model for T2D prediction. Specifically, we also will use tri-axial actigraphy data as well as one-time assessments of patient background variables and clinical lab results (see Figure 1).

Figure 1.

Different types of data used in this study. From left to right: Activity monitoring device (ActiGraphy), CGM measurement, and static data containing physical attributes and lab test results.

3. Data Collection

The data used in this study are part of a larger longitudinal observational project in collaboration with Christiana Care Health System that received IRB approval. Adults with T2D and their spouses/partners were recruited from an active endocrinology clinic. The T2D patients were between 33–78 years of age. The purpose of the larger longitudinal study was to examine psychosocial factors that influence blood glucose levels and T2D management among T2D patients and their spouses (Soriano & Laurenceau, 2020). In the current study, only the data from patients with T2D (not their spouses) were used. For each patient, various types of data were collected. First, during the baseline assessment, patients completed self-report questionnaires, and additionally, their blood and urine samples were collected. Also, they were provided with two types of wearable devices (CGM and ActiGraphy), which they wore for the next seven days.

Blood and urine samples were collected a second time at a one-year follow-up. This way, our dataset contained three types of data: 1) CGM data, 2) Activity (ActiGraphy) data, and 3) data related to the demographics, physical attributes, and lab test results (referred to as the static data in this paper). We show patient i’s entry record with D_i, where D_i = (CGM_i, Acc_i, DL_i), and CGM_i, Acc_i, DL_i refer to the three types of data described. Figure 1 shows these three types of data. Of the 63 patients in our dataset, 9 were missing HbA1c lab results at one-year follow-up and were therefore excluded from analyses. As a result, data from 54 patients were included in our analyses.

CGM measurements –

At the time of the baseline assessment, each patient was equipped with a Dexcom G4 Platinum Professional CGM device. Patients were trained on how to operate the device (e.g., how to wear the sensor, how to calibrate the device to obtain reliable readings). The CGM devices monitored the blood glucose levels for each patient in five-minute intervals for 24 hours per day across seven consecutive days. Due to slight differences in scheduling, the number of blood glucose readings (D) varied among the patients–ranging from 1445 to a maximum of 2016 measurements. For the user i in the dataset, the series of CGM measurements on D time points (t₁, t₂ , …, t_n) can be shown as CGM_i = C_t1,i,C_t2,i,…C_tD,i, where C_tk,i is the CGM value, measured in t_k for the patient i.

ActiGraphy data –

Over the course of the same seven days that patients wore the CGM, they also wore an actigraphy device on their non-dominant wrists. Specifically, patients wore the ActiGraph wGT3X-BT device. Actigraphy records the physical movements of each patient in three-dimensional space in 0.1 seconds intervals. These data can be used to determine the type and level of physical activity as well as sleep patterns. We used the raw data measured by the ActiGraph devices, which was extracted from the ActiLife 6.13.3 software (the commercial software provided by the manufacturer ActiGraph, LLC). As part of the initial processing phase, a minimal level of noise removal and missing data interpolation were applied to the acceleration signals. To preserve the original raw patterns of the change in the three-dimensional measurements, we did not use the available software for additional processing such as using the movement patterns to identify the type of activity that an individual was doing (e.g., walking or sitting). Each measurement contained three values (X_i, Y_i, Z_i), showing the acceleration in three dimensional coordinates of the device at a corresponding time point. This way, each patient had approximately 6 million ActiGraphy measurements over the seven-day recording period.

Static data –

In addition to the wearable sensor data, patients’ sociodemographic data (age, gender), physical attributes (body weight, height, and waist circumference), as well as some lab tests were collected in our study. The lab bloodwork measured the values of HbA1c, HDL cholesterol, LDL cholesterol, non-LDL cholesterol, total cholesterol, and triglycerides. These values were remeasured during the one-year follow-up assessment. Some of the characteristics of the static data used in this study are shown in Table 1.

Table 1.

Statistical information about the static data (mean ± standard deviation).

Data	Data Category	Mean
Age	Demographic	(60.4±10.1) years
Gender	Demographic	(F=19/M=31)
Cholesterol	Lab test results	(163.4±40.3) mg/dL
HDL cholesterol	Lab test results	(49.9±14.3) mg/dL
LDL cholesterol	Lab test results	(81.2±31.2) mg/dL
Triglycerides	Physical attributes	(164.0±112.0) mg/dL
Weight	Physical attributes	(216.0±50.6) lbs
Height	Physical attributes	(67.7±3.9) inches
Waist circumference	Physical attributes	(106.4±16.4) cms

4. Predictive Model

In developing our model, we specifically aimed to create a model that can predict the changes in four important T2D sequelae (HbA1c, HDL cholesterol, LDL cholesterol, and Triglycerides) after one year for patients diagnosed with T2D. Our deep neural architecture is inspired by a deep architecture originally proposed for extracting the hidden patterns from multi-sensor measurements using a CNN-RNN architecture (Yao, Hu, Zhao, Zhang, & Abdelzaher, 2017). We extended this architecture by including static data and making additional adjustments to achieve higher performance. In our proposed framework, a hybrid structure is used for learning efficiently from the available types of data. Our expression of signal measurements and their correlations facilities supplementing these data with the static records. Specifically, the proposed pipeline for processing the datasets consists of four main steps. First, the raw wearable signal data representation is modified. Second, the model extracts the correlation of signals. Third, the temporal properties of the resulting sequence are extracted. Finally, the static data are added to the model. The proposed model was implemented using the TensorFlow and Keras software libraries. The code is publicly available on GitHub¹.

4.1. Conversion of Frequency-Domain

Raw CGM and ActiGraphy signals contain information about the magnitude of the measurements at various times. As shown by others, the raw signal format might not the best format for finding local frequency patterns and for extracting the dependencies between the signal measurements (Guo, Wu, Ding, & Feng, 2008). Instead of this raw format, transforming the raw signals to other variations can help in identifying local frequency patterns in the sequential data. We therefore transformed the time-domain signals to the frequency-domain signals and applied discrete Fourier transform (DFT) on the raw signals for each subject i. DFT converts a signal from the time-domain (signal strength as a function of time) to the frequency-domain (signal strength as a function of frequency) and shows the signal’s spectral content, divided into discrete bins (frequency bands). To maintain the temporal information on the signals (which can be extracted and learned by the model), the signals were divided into T windows with a length of τ data points for each window. Each sequence S_i, had a fixed number of windows. The DFT function was applied to each window separately. In this transformation, τ amplitude and phase pairs were generated from each window. This way, the dimensions of the input vector (to the model) for each signal were as following:

$$ActiGrpah:\left[3\times T\times 2\tau \right]$$

$$\text{CGM}:\left[1\times \text{T}\times 2\tau \right]$$

For the actigraphy data, number 3 refers to the number of measured dimensions (x, y, z), and for both CGM and actigraphy, T is the number of formed windows, and 2τ is the number of entries (amplitude and phase) in each window. Table 2 also shows these different elements of our model. We had D ≈ 6 × 10⁶ datapoints for the patients that were divided into T intervals. Since the number of data points in our application was very large, another level of hierarchy was added in the implementation above to the intervals. We grouped the T intervals into F frames of size T/F. The CGM data was expanded by concatenating it twice to itself in the first dimension (3×DFT). Then, the data was reshaped in a tensor with the dimensions of [3×F×T/F×2τ] for both CGM and actigraphy measurements and was then fed into the network.

Table 2.

The main elements of the model

Element Type	Shown as
Patients’ timepoints	D
Frequency domain timepoints	2D
Number of intervals	T
Datapoints per interval	2D/T
Number of frames	F
Number of intervals in each frame	T/F

4.2. Signal Correlation Extraction

After the initial preprocessing steps, the model was ready to analyze intra- and inter-sensor measurement dependencies as show in Figure 2. We use a CNN architecture to extract the dependencies in the frequency domain of signals. The CNN layers extract features from the multi-dimensional data. In the current model, we use two separate CNN layers (Figure 3). First, the model processes the intra-sensor measurement dependencies, i.e., the interactions between the actigraphy signals for the three dimensions (X_i, Y_i, Z_i). This was implemented by using three convolutional layers, consisted of 64 convolutional filters on the input data. For the CGM signal, there is only one dimension of measurement in the data and there are no intra-sensor interactions to be extracted. However, we used a unified architecture for both actigraphy signals and CGM measurements to simplify the architecture. The outputs of these two separate convolutional layers are then concatenated and fed to a second stack of convolutional layers for processing inter-sensor measurement dependencies. This second stack of layers will specifically look for the interactions between the CGM and actigraphy signals. The output is the flattened vector shown in Figure 3.

Figure 2.

Two types of wearable sensor signal dependencies for signal samples. Inter-sensor dependency (green arrows). This type of dependency relates to the interactions between the measured CGM value and the three measured acceleration values in the corresponding time point. Intra-sensor dependency (blue arrows). This refers to the interactions among the three measured acceleration values using ActiGraphy signal in one timepoint.

Figure 3.

The CNN architecture. Two series of inputs (CGM and ActiGraphy) represented as transformed discrete signals using discrete Fourier transformation (DFT). First, the extended CGM signals and ActiGraphy signals are fed into the first series of convolutional layer.

4.3. Temporal Data Training

In the third step of our pipeline, temporal patterns of the whole sequence of frames (intervals) were extracted. As mentioned earlier, the whole sequence of data was shaped into F number of frames with a size of T/F. Here, each frame represents one data point in the time series sequence. An RNN architecture with gated recurrent unit (GRU) cells was used to learn meaningful features from these time sequences. Specifically, two stacked GRU layers were used to process F frames sequentially.

4.4. Static Data Training

One of the main steps toward having a comprehensive prediction model for the T2D was the addition of static data into the learning model. This data included demographic, physical attributes, and lab test results. Because of their informative nature as diabetic risk factors, we expected to achieve better performance by adding this information to the model. Converting the time-series data into the frequency domain sequences using DFT, and further extracting the pattern dependencies among the signals, lifted the domain inconsistency barrier for adding this static data into the sensor measurement information. Here, we used a dense (fully connected) layer prior to the concatenation of the static data with the output of GRU layers (applied to the dynamic data). After the concatenation, we used an additional dense layer serving as the output layer of our architecture. Figure 4 shows the architecture of the GRU network and the supplementary learning partition for including the static data.

Figure 4.

The merging parts. This part of the architecture receives the flattened output of the CNN network and merges it with the physical attributes data using the dense and concatenation layers.

5. Model Evaluation

5.1. Evaluation using a separate dataset

While the goal of the proposed model was to predict the progression patterns of T2D and the model was presented as it was applied to our own T2D case study, we evaluated this model on a separate but closely related dataset, too. We examined the proposed model on the publicly available OhioT1DM dataset (Marling & Bunescu, 2018). This dataset was originally developed for predicting the blood glucose levels of 12 subjects with Type 1 diabetes (T1D), and has been used in a couple of public prediction competitions (Martinsson et al., 2018; Yin et al., 2019). It contains eight weeks of data related to CGM readings, insulin levels, physiological sensor recordings, and self-reported life events. CGM measurements in a two-hour window and three static variables of age, weight, and gender were used in the training process. Following the similar prediction procedure presented for this dataset, we used the model to predict the blood glucose levels, as determined by the CGM measurements. For the standard task of predicting the CGM value in 60 minutes in the future, we have achieved an RMSE (root mean squared error) of 34.2. The performance of the models that previously used this T1DM dataset for predicting the disease progression is higher than our achieved accuracy. This is not surprising as we have not considered many complicated correlations between the T1DM risk factors and its progression behavior.

5.2. Evaluation on our own T2D dataset

We evaluated the performance of the model in a series of experiments. In these experiments, we used a GRU network with 32 cells, and therefore, the number of frames (F) was set to 32 to match the number of GRU cells. We set the number of intervals (T =200) to break the 6 million data points (D) into 300,000 data point per interval, which in turn was determined by the batch size fitting in the memory. We tested our models in both regression and classification modes. As testing the model with regression does not confine the prediction model into the predefined classes, by having access to the actual estimated values, the end-users (such as the providers) may have a clearer picture of the predicted progression, enabling them to make more informed decisions. In the regression mode, our goal was to predict the final difference (indicating the amount of change within one year) of a certain T2D sequelae s (e.g., HbA1c) as shown below:

$$\Delta s=s_{\text{next}-year}-s_{current}$$

As discussed earlier, the four T2D sequelae that were used as the model outputs were HbA1c, HDL cholesterol, LDL cholesterol, and triglycerides. In the classification mode, the model was used to predict whether the value of a sequela s will be increased or decreased within the one-year period. Here, two binary classes (increase and decrease) were considered, and the class C_s was defined as:

$$If\hspace{0.17em}\Delta s\hspace{0.17em}\ge \hspace{0.17em}0,\hspace{0.17em}{C}_s\hspace{0.17em}=\hspace{0.17em}\text{Increase}$$

$$If\hspace{0.17em}\Delta s\hspace{0.17em}<\hspace{0.17em}0,\hspace{0.17em}{C}_s\hspace{0.17em}=\hspace{0.17em}\text{Decrease}$$

For the regression mode, we used mean squared error and for the classification mode, we used binary cross-entropy as the loss functions. In all experiments, we used 10-fold cross-validation to generate the training and test sets.

To demonstrate the difference in the performance of our model when it receives various modalities of data, we have performed a set of experiments related to three scenarios of data availability: 1) using only the wearable sensor data (i.e., the measurements that were recorded by CGM and ActiGraphy devices), 2) using the static data only, and 3) using both wearable sensors and static data. Table 3 shows the accuracy (for the classification task) and RMSE (for the regression task) of our model based on these three scenarios of data availability. For calculating RMSE for a sequela s, considering ∆s as the actual measured change in s for each patient, $\widehat{\Delta }s$ as the predicted change, and N as the number of predicted records, we used the following formula:

$$RMSE=\text{sqrt}(\frac{{\sum }_{i=1}^N{\left(\widehat{\Delta }s-\Delta s\right)}^2}{N})$$

We have used three baseline methods for predicting the progression of the disease. The first baseline method was the random forest classifier and regressor using the bagging techniques of the ensemble algorithm series with 10 estimators (Breiman, 2001). The second method was XGBoost (Chen & Guestrin, 2016) classifier and regressor based on gradient boosting. The third method was a model based on the popular Wide and Deep architecture (Cheng et al., 2016). We used the wide part of the network for processing the static data, and the deep part was used for the dynamic part. In this model, no frequency domain conversion was used, and instead of using the raw time-series data, preprocessed data using the corresponding software for each wearable device was used. A two-layer LSTM network (1445 number of cells in each layer) was used for implementing the deep part of the model for processing signal data. We have used the TensorFlow library for implementing the random forest and XGBoost baselines and Keras library for implementing the Wide and Deep network. Table 4 shows the performance of our complete model (utilizing both wearable and static data) in comparison with the baseline methods. Besides accuracy and RMSE, we also report the performance of the model in the classification mode using the area under the ROC curve (AUROC) and in the regression mode (predicting the sequelae change), using the Normalized RMSE values (NRMSE = RMSE/Mean). This normalized error rate is useful for comparing the quality of the same model for various labels with different ranges. Additionally, to study the impact of the different number of blood glucose readings (for each individual) on the performance of our model, we report the results of a series of regression analysis experiments studying the relationship between these two in the Appendix. We also report the model’s performance for each individual sample in Table A1. Finally, to improve the explainability of the model based on the input features, we ranked the features in our dataset. We calculated the SHAP values using the SHAP toolbox (Lundberg & Lee, 2017). SHAP estimates the importance of features by removing the features from the input one by one and evaluating the model performance in their absence. A SHAP value demonstrates the relative contribution of each feature to the output of a certain model. A higher SHAP values shows a higher contribution (importance) of a feature in determining the outcome of the predictive model. Figure 5 shows the normalized SHAP values for four predicted sequelae. The log of the original SHAP values was multiplied by 10,000 to obtain the normalized numbers shown in this figure. For the time-series wearable signals, the mean SHAP valued across the time frames are reported.

Table 3.

Accuracy and RMSE (mean ± SD) of the proposed model compared with three scenarios of data availability.

	Accuracy				RMSE (mg/dL)
Sequalae Type	HbA1c	HDL	LDL	Triglycerides	HbA1c	HDL	LDL	Triglycerides
Wearable Data	0.87±0.02	0.81±0.02	0.82±0.02	0.90±0.03	1.56±0.07	6.17±0.46	10.04±0.73	17.69±1.00
Static Data	0.69±0.02	0.67±0.01	0.71±0.02	0.70±0.03	2.64±0.12	10.60±0.63	10.41±0.78	18.10±1.22
Wearable+Static	0.90±0.02	0.87±0.02	0.88±0.02	0.92±0.02	1.37±0.07	5.82±0.43	9.93±0.71	17.38±1.89

Table 4.

Type 2 diabetes sequelae prediction performance. For four algorithms of Random Forest (RF), XGBoost, Wide and Deep, and the proposed method, classification accuracy (Acc.), classification AUROC, regression RMSE, and regression NRMSE are reported.

	RF	XGBoost	W&D	Proposed
Accuracy	0.69±0.01	0.72±0.01	0.86±0.01	0.90±0.02
AUROC	0.65±0.02	0.69±0.02	0.83±0.02	0.88±0.02
RMSE	2.77±0.1	2.63±0.1	1.68±0.1	1.37±0.8
NRMSE	0.35	0.34	0.21	0.17
a) HbA1c
	RF	XGBoost	W&D	Proposed
Accuracy	0.73±0.02	0.75±0.02	0.80±0.02	0.87±0.02
AUROC	0.74±0.03	0.75±0.03	0.79±0.02	0.82±0.03
RMSE	10.67±0.7	10.58±0.6	6.20±0.5	5.82±0.4
NRMSE	0.24	0.24	0.14	0.13
b) HDL cholesterol
	RF	XGBoost	W&D	Proposed
Accuracy	0.77±0.02	0.77±0.02	0.81±0.02	0.88±0.02
AUROC	0.77±0.02	0.79±0.02	0.79±0.02	0.82±0.02
RMSE	12.03±0.9	11.14±0.6	10.46±0.7	9.93±0.7
NRMSE	0.09	0.09	0.08	0.08
c) LDL cholesterol
	RF	XGBoost	W&D	Proposed
Accuracy	0.84±0.02	0.86±0.02	0.89±0.02	0.92±0.02
AUROC	0.79±0.03	0.79±0.04	0.80±0.03	0.82±0.03
RMSE	19.99±1.2	14.49±1.0	18.37±1.2	17.38±0.9
NRMSE	0.21	0.15	0.20	0.19
d) Triglycerides

Figure 5.

The normalized SHAP value of the used features for predicting the T2D sequalae.

6. Discussion

In this study, we presented a new architecture for predicting the progression of T2D using static (e.g., patient demographics) and dynamic (e.g., wearable sensor) data. The results in Table 3 show that the performance for our model in terms of accuracy and AUROC was improved by including both (static and dynamic) datasets. Specifically, for the classification test on the HbA1c, which is the most informative sequalae in monitoring the T2D progression, the accuracy was improved by 3.41% in our model, compared to the Wide and Deep approach. While evaluating the performance of our model using only the error values may show its relevance to the practical cases less clearly, we can use some relevant measures in the field as a comparison. As an established threshold for the medical laboratories, the error rate of ±6% is considered as the maximum tolerable error in measuring HbA1c (Little, 2014). This is a threshold set by the US Food and Drug Administration (FDA) agency for measuring the HbA1c in lab tests. In our model, we achieved an RMSE of 1.37%, meaning any predicted value of HbA1c in our model will have an average error rate of 1.37%. This is lower than the medically acceptable error rate by the FDA and may demonstrate the applicability of our proposed method to actual T2D management programs. Besides HbA1c, our proposed model also shows a good performance in estimating the future HDL and LDL cholesterol values, achieving 82% AUROC for both values in the classification task, and NRMSE of 0.13 and 0.08 respectively, in the regression task. We have also compared the performance of our model with three baseline methods based on random forest, XGBoost, and Wide and Deep architecture. As Table 4 shows, the accuracy of the proposed model surpasses the baseline scenarios in predicting all four T2D sequelae. The underlying structure of the Wide and Deep method and our model are very similar, while our model could consistently achieve more competitive results. This shows that extracting the frequency-domain localizations at multiple levels could enhance the overall performance of our model. T1D and T2D are two different diseases with unique properties of their own. While our study is primarily focused on T2D, in lieu of a comparable public T2D dataset along with published predictive models on it, we have picked a T1D dataset to evaluate our predictive model from a technical standpoint. We evaluated the performance of our model using the OhioT1DM dataset for predicting blood glucose values in the next 60 minutes in a separate case study related to type 1 diabetes. This was the closest publicly available dataset to our problem that we could find. In this experiment, with the absence of four primary targeted sequelae for prediction, the value of CGM itself was used as the predicted value. Among the existing studies, two of the best results were reported by (Yin et al., 2019) (RMSE of 33.3) and (Martinsson et al., 2018) (RMSE of 33.2). We have achieved an RMSE of 34.2 on the same tasks, using the model presented in this work. While our model design was not specifically tuned for the OhioT1DM data (unlike the two existing studies mentioned earlier), obtaining similar performance to the state of the art in the field further demonstrates the potential of our method for studying similar problems. We expect that the true potential of our method can be particularly revealed in scenarios where more than one type of signal is available. Our method can effectively extract intra-sensor and inter-sensor dependencies.

According to our SHAP experiments shown in Figure 5, the importance degree (feature ranking) of the temporal measurements was found to be higher than physical attribute records consistently across the prediction tasks for the four sequelae. Among the temporal measurements, CGM measurements were found to be the highest-ranking features, and the acceleration measurements had the next highest ranks. As discussed above, we have observed that the performance of the model without considering the static features is lower than a model utilizing static and dynamic (temporal) data together. For triglycerides, this difference is smaller than the other three. This smaller difference might be due to the lower dependency of triglycerides to the available physiological features. Another possible explanation for this could be the higher variance of the triglyceride values compared to other three sequalae, making the prediction task essentially more challenging for all models. Based on the importance ranking of each time-series signal compared to the other features, it can be observed that excluding each of the four wearable data sources would result in worse performance. The higher importance of acceleration in the Z dimension, compared to X and Y dimensions may be related to the greater contribution of this ActiGraphy measurement in indicating the physical activity intensity (Bai et al., 2016). The Z dimension corresponds to the movement of the ActiGraphy device toward a perpendicular axis to its screen. Weight and waist circumference were found to be the two most important factors among the physical attributes for predicting future values in our model.

One unique property of our study was demonstrating that our model can predict future values of four major T2D sequelae, using data collected over a relatively short period. While we did not explicitly extract specific activity patterns (such as sleeping and running), we expect that our model was able to implicitly use those patterns for its predictions, as actigraphy patterns generally are considered as a reliable source of activity recognition. Our multi-level architecture enabled our model to identify the underlying patterns from various sources of data.

Potential limitations of our study include the age range of the participants as well as the sample size of patients. Because of the focus of the original study was on adults, we did not have access to the data from younger individuals. We note that as the prevalence of T2D increases with the age, our model can still be useful on a large variety of T2D patients. It should be also noted that HbA1c levels alone are generally not enough for the diagnosis of T2D (American Diabetes Association, 2021). Accordingly, a limitation of our study is the lack of blood glucose and serum insulin measurements. Including such additional physiological and biochemical data would further improve our work. While our study has mainly focused on diabetes, the method we described may be useful for studying a variety of similar health problems and predicting future health outcomes. The scope of such health problems is very large, as the presence of both static and dynamic datasets is ubiquitous in many health problems.

7. Conclusion

In this study, we presented a new model for predicting the progression of type 2 diabetes sequelae using a combination of dynamic (wearable sensor) and static (demographic and lab test) data. Our model followed a deep neural architecture using a hybridized CNN and RNN network, and specifically predicted future HbA1c, HDL cholesterol, LDL cholesterol, and triglycerides values. It also used various fusion techniques to extract the dependencies among the different signals. We showed that our model was able to achieve better predictive performance in comparison with several baselines including the methods that use portions of versus both static and dynamic data.

Appendix. Error Rates for each patient

This section presents the observed error rates of every patient in the study while predicting the four sequalae, using the proposed model.

Table A1.

The achieved RMSE for four sequalae of HbA1c, HDL Cholesterol, LDL Cholesterol, LDL Cholesterol, and Triglycerides (TG), for each of the 54 patients. Size shows the number of available CGM measurements.

ID	Size	HbA1c	HDL	LDL	TG
1	1990	1.71	6.54	7.55	18.16
2	1975	1.49	5.87	11.02	10.23
3	2006	0.87	5.78	9.23	18.61
4	1973	1.51	5.50	12.71	14.63
5	1986	1.15	5.68	8.02	10.26
6	1972	1.90	5.22	8.59	23.15
7	1924	1.74	4.86	9.92	21.08
8	1895	1.31	5.14	7.54	12.17
9	1935	1.10	6.54	11.42	11.83
10	1956	1.50	5.56	7.95	20.28
11	1995	0.82	5.08	9.89	13.93
12	1991	0.81	7.24	12.48	23.93
13	1915	1.40	4.95	9.91	26.52
14	1550	0.85	5.57	11.85	27.46
15	1925	1.99	4.39	8.16	19.58
16	1680	1.22	5.19	8.47	12.69
17	1840	1.12	4.86	12.15	16.55
18	1925	1.05	5.09	8.13	20.90
19	1970	0.77	6.04	11.05	18.94
20	1840	2.31	5.86	8.87	24.96
21	1915	1.84	5.83	10.30	9.53
22	1698	1.73	5.47	7.42	20.51
23	1445	1.65	5.75	9.69	20.22
24	1590	1.18	7.56	10.04	23.69
25	1870	0.96	5.95	9.60	17.12
26	1987	1.80	5.97	11.48	16.76
27	1930	2.06	6.64	12.40	25.36
28	2008	1.64	5.27	11.27	16.72
29	1840	1.83	5.44	8.55	11.92
30	1959	0.97	6.70	7.27	10.84
31	1690	0.83	5.97	8.73	10.43
32	1882	0.95	7.58	12.08	17.01
33	2016	2.01	6.26	11.23	9.02
34	1550	1.66	4.65	9.85	20.73
35	1845	0.66	5.64	9.41	7.85
36	1886	1.83	6.92	8.05	14.43
37	1995	1.84	6.01	12.10	17.35
38	1529	1.53	6.62	12.48	8.51
39	2000	0.83	7.96	8.92	16.49
40	1992	0.86	6.54	7.55	18.16
41	1980	0.89	5.87	11.02	10.23
42	1655	1.19	5.78	9.23	8.61
43	1907	1.70	5.50	12.71	14.63
44	1893	0.92	5.68	11.02	10.26
45	1915	1.57	5.22	8.59	23.15
46	2008	1.14	4.86	9.92	21.08
47	1880	2.00	5.14	7.54	32.17
48	1790	1.28	6.54	11.42	21.83
49	1945	1.06	5.86	9.95	20.28
50	1938	1.52	5.08	9.29	13.93
51	1905	0.85	7.24	12.48	23.93
52	1920	1.23	5.95	9.91	22.52
53	1904	1.71	5.57	11.85	17.46
54	1547	1.49	4.89	8.16	19.58
All	1871	1.37	5.82	9.93	17.38

We have also run a series of regression analysis experiments to study the correlation between the number of available blood glucose readings and the difference from the reported population mean RMSEs, for the four sequalae. Here are the results of our regression analysis. For HbA1c, R = 0.03, and P-value = 0.85, for HDL Cholesterol, R = 0.07, and P-value = 0.61, for LDL Cholesterol, R = 0.04, and P-value = 0.78, and finally, for Triglycerides, R = 0.07 and P-value = 0.62 were obtained. Collectively, we have observed in the current variances in the length of blood glucose readings, have a minimal effect on the performance of the method.