Identification of Physical Activity Type in People with Diabetes: A Spectrogram-Based Approach

Abstract

Background:

Individuals with type 1 diabetes (T1D) require close glucose monitoring to prevent both short- and long-term complications. Physical activity (PA) is a significant source of variability in metabolic dynamics, leading to glycemic fluctuations that depend on the type, intensity, and duration of the exercise. Accurately monitoring and classifying the type of PA is crucial for optimizing glycemic control and minimizing the risk of hypoglycemia.

Method:

This study utilizes the largest clinical trial of PA in people with T1D to date, the Type 1 Diabetes and Exercise Initiative (T1DEXI), which included both structured and unstructured PA sessions, to develop an online classification approach for identifying the type of PA (aerobic, interval, resistance). A computationally efficient convolutional neural network (CNN) was trained on time–frequency representations (spectrograms) of step count and heart rate signals, readily available from wearable devices, from the structured PA sessions of the T1DEXI dataset. The proposed methodology presents an ad hoc process for designing the spectrograms based on the CNN architecture to optimize the classifier’s performance.

Results:

The CNN-based classification approach was implemented using spectrograms of 5- and 30-min signals, resulting in two classifiers that achieve high classification accuracy when evaluated on the structured PA sessions. The 5-min classifier was then applied to unstructured PA sessions, where the predicted distribution of glucose changes for the activity types was consistent with clinical evidence.

Conclusion:

These results demonstrate the potential of the proposed approach for its integration into decision support systems or automated insulin delivery systems, enabling improved glucose management during exercise in T1D.

Keywords

physical activity classification convolutional neural network spectrogram type 1 diabetes

Introduction

Type 1 diabetes (T1D) is a chronic condition characterized by the autoimmune destruction of pancreatic beta cells, leading to insulin deficiency. This results in elevated blood glucose levels that can cause severe long-term complications if not properly managed. The treatment of T1D involves the administration of exogenous insulin, which requires close monitoring of glucose levels to prevent hypoglycemic and hyperglycemic episodes.¹

The glycemic dynamics in T1D are strongly affected by various sources, mainly food intake and other factors such as physical activity (PA) and stress.^2,3 Significant research has been dedicated to modeling and controlling glucose levels in response to food intake and insulin administration.⁴ Nonetheless, the presence of PA also introduces an important source of variability in metabolic dynamics, leading to significant decreases or increases in glucose levels.^5,6 This behavior depends on the type, intensity, duration, and specific conditions of the patients during exercise, between others.^7,8 Therefore, it is essential to improve insulin administration and glucose prediction strategies by taking into account the specific type of physical exercise being performed.^9,10 Currently, there is no consensus within the scientific community on how to model the effect of PA. As such, the ability to accurately monitor and classify the type of PA in patients with T1D in real time is crucial for optimizing glycemic control and compensating for dangerous fluctuations in glucose. This information could be integrated into clinical decision support systems (DSS) or automated insulin delivery (AID) systems, allowing for the adjustment of insulin administration based on the type of activity and its expected glycemic impact.

Numerous studies and clinical trials have been conducted to address the challenges associated with PA monitoring and classification in the context of T1D.^11–14 Some works have explored the use of heart rate (HR) and/or accelerometer and gyroscope-based measures for classifying the type and intensity of PA.^15–18 For example, Cho et al.¹⁵ developed a random forest-based framework for real-time detection and classification of activity types and intensity in controlled environments, utilizing triaxial accelerometer, HR, and continuous glucose monitor (CGM) data. Others have incorporated additional physiological variables, such as skin conductance or skin temperature, to enhance the classification capabilities.^19,20 In contrast, Abdel-Latif et al.³ explored the classification of PA and acute psychological stress in T1D using portable sensor data. In that work, a recurrent neural network (RNN) with long short-term memory (LSTM) architecture and a decision tree-based strategy is employed, demonstrating the feasibility of a multitask classification approach. While some of these studies have provided valuable insights, they have primarily focused on evaluating their methods in structured and predefined exercise sessions within controlled environments. This approach, even though useful for initial validation, may not fully capture the complexities of real-world PA patterns encountered by individuals with T1D. In their daily lives, people often engage in a variety of unstructured and varied physical activities, which can introduce additional challenges for accurate monitoring and classification.

To address the previously mentioned limitations, the Type 1 Diabetes and Exercise Initiative (T1DEXI) clinical trial was designed to collect detailed data on glucose levels, PA, and other relevant factors in patients with T1D during their daily routines and structured exercise sessions.²¹ By including a diverse range of activities and contextual information, the T1DEXI dataset enables the analysis of glucose dynamics in a more realistic, everyday setting. Researchers have used this rich dataset to conduct studies that estimate the probability of hypoglycemia risk during physical exercise and in the postexercise period using advanced statistical modeling techniques.^22,23 Digital twin models have also been developed for the simulation of continuous glucose response.^24,25 Additionally, associations between daily step count classifications and continuous glucose monitoring metrics have been analyzed.²⁶ As conclusions of the T1DEXI trial, the authors found that prolonged moderate-intensity aerobic activities have been found to present the greatest drop in glucose levels, followed by interval activities, while resistance activities show the smallest drop or even an occasional rise.²¹ Therefore, differentiating among exercise types may be beneficial to determine the most appropriate insulin management or nutritional intake strategy for exercise in individuals with T1D. Building on this, Fushimi et al.²⁷ developed the first online classifier to identify the type of exercise using HR and triaxial accelerometer data from the T1DEXI dataset. Although their LSTM-based classifier achieved good accuracy in classifying both structured and unstructured exercise sessions, the use of RNN approaches can be computationally demanding and challenging to implement on low-power devices, such as wearable fitness trackers or insulin pumps.

The studies mentioned above and several others in the literature highlight the use of machine learning^28,29 and deep learning (DL) techniques³⁰ in the field of healthcare. Within DL methods, convolutional neural networks (CNN) have emerged as a powerful tool, especially in image processing, due to their high effectiveness in extracting relevant features from visual data. This approach has been applied for the classification of biomedical signal spectrograms, which serve as an alternative representation of time-series data in the time–frequency domain. For instance, studies have exploited CNNs for the classification of electrocardiogram signals,³¹ electroencephalogram data for epilepsy diagnosis,³² and the early detection of diabetes based on skin impedance and HR variability³³ or from HR signals.³⁴ Finally, the work by Hsu et al.¹⁶ proposes a DL algorithm that uses accelerometer and gyroscope data for the classification of sports (tennis, golf, volleyball, etc.) in only 10 healthy subjects without diabetes.

This article proposes an online classifier capable of accurately determining the type of PA (aerobic, interval, resistance) performed by individuals with T1D. Beyond classification, this work also seeks to validate the classifier’s performance by contrasting the results against the expected glycemic changes for each activity type, analyzing the distribution of these changes in real sessions versus the classifier’s outcomes. Differing from previous works, the methodological approach utilizes spectrogram representations of step count and HR signals from the T1DEXI dataset to train a CNN classifier, which is computationally more efficient than other neural network-based methods typically used for time-series analysis. Moreover, step count data were used instead of triaxial accelerometer data, a signal that is usually unavailable in practice. Conversely, step count is accessible from most commercial devices and also eliminates the need for additional preprocessing steps. Furthermore, the lower sampling rate of the step count signal (0.1 samples per second) compared with the accelerometer data (32 samples per second across three axes) results in fewer samples, further reducing the computational requirements. As a result, overall, the proposed method offers practical advantages for real-time and real-world scenarios where low-latency inference is crucial. Initially, a 30-min signals classifier was used to design and evaluate the spectrogram generation methodology. Then, to enable real-time applications for variable-duration and unstructured PA events, the methodology was extended to a computationally efficient 5-min CNN-based classifier. This adaptation makes it suitable for the dynamic nature of PA while facilitating its integration into clinical DSS or AID systems. This aims to improve glucose management during exercise in patients with T1D by offering a more precise understanding of the type of activity being performed.

T1DEXI Study

The dataset used in this work originates from the T1DEXI study,²¹ which was conducted under real-life conditions and collected a wide variety of data related to exercise in individuals with T1D. This study involved a total of 497 adults diagnosed with T1D for at least 2 years prior to the start, with an age of 37 ± 14 years and a glycated hemoglobin (HbA1c) value of 6.6% ± 0.8%. The participants were under one of the three most common insulin administration treatments at the time of the study: commercially approved hybrid closed-loop, standard insulin pump, or multiple daily injections.

Each participant was randomly assigned with six structured exercise sessions based on videos that the participants followed at home. The exercise types include aerobic ( $n = 162$ ), interval ( $n = 165$ ), or resistance ( $n = 170$ ) over a 4-week period. In addition to the exercise sessions designed for the study, the participants continued with their usual physical activities. Throughout the study period, each volunteer recorded information about their diet, insulin administration, and glucose levels measured using a CGM device, specifically the Dexcom G6.

Additionally, the patients used the Verily Study Watch for the continuous collection of HR and movement data (from which step count is inferred). During the structured exercise sessions, the participants were asked to wear a Polar H10 chest strap HR monitoring system for a more accurate monitoring of the exercise relative intensity.

This valuable resource has been used in this work as the basis for generating a dataset, which was employed to train and evaluate exercise classification models. The use and analysis of this clinical dataset is highly relevant, as the data were collected under everyday life conditions over an extended period, unlike other studies conducted in controlled environments. This allows for the design and evaluation of machine learning methods in more challenging and realistic scenarios, where daily activities (diet, PA, sleep quality, etc.) and life events (stress, etc.) present an impact on the glucose level.

Methodology

The proposed classification approach was supported by a methodology comprising data analysis and preparation, signal processing and spectrogram generation, and the application of CNN for image classification, as detailed in the following subsections.

Data analysis and preparation

A data filtering process was carried out to ensure data consistency and the validity of the analysis and results. Only those days in which the participants recorded some PA were selected, considering participants with the three types of insulin administration treatment. It is important to note that the missing samples in CGM and wearable devices are a significant constraint when extracting information from the sampled signals.^35,36 To address this limitation, only sessions with more than 90% of the available samples were selected from the Verily Study Watch, which recorded HR and step count at a sampling rate of 10 s. This requirement helps to maintain the signal integrity used for the analysis, minimizing the impact of missing data.³⁶ Interpolation was performed for any remaining missing values within the selected sessions.

After the data filtering process described earlier, a total of 462 participants were considered for the analysis. The participants had a mean age of 37 ± 14 years and had been diagnosed with T1D for an average of 19 ± 12 years. The mean HbA1c value of the study population was 6.6% ± 0.7%, indicating generally good glycemic control. The dataset included 1798 structured exercise sessions and 5121 unstructured PA sessions recorded by the participants.

The filtered dataset was divided into two parts for the training and validation process of the classification approach:

Dataset Structured Sessions (SS): This dataset is used for training and validation and consists of the structured exercise sessions labeled as aerobic, interval, and resistance.

Dataset Unstructured Sessions (US): This dataset is only used for validation and concentrates on the daily life activities recorded by the participants, which are self-categorized at the time of recording.

The structured exercise sessions in Dataset SS were used for the training and initial evaluation of the developed classifier, as they provide a more controlled and well-defined set of physical activities, with a fixed duration of 30 min. In contrast, Dataset US, containing the unstructured daily life activities, was utilized for the validation of the classifier in a more realistic scenario. The unstructured sessions in Dataset US had a mean duration of 58.6 ± 47.2 min, with a median of 45 min and an interquartile range of 32–60 min, reflecting the variable nature of physical activities performed in daily life. This division of the dataset allows for a comprehensive evaluation of the classifier's performance with previously unused data, ensuring its robustness and applicability in both structured exercise settings and real-world daily activities.

To further characterize the exercise sessions, the change in CGM during each session was calculated as follows²⁷:

Δ CGM = CGM (T_{f}) - CGM (T_{i}),

(1)

where

T_{i}

and

T_{f}

represent the starting and ending times of the exercise, respectively. The change in glucose during the activity can serve as an additional representative indicator of the activity type, as previous studies have reported significant differences in the ΔCGM values across three exercise categories: aerobic, interval, and resistance.²¹ For the structured exercise sessions in Dataset SS, the statistical differences in the

Δ

CGM values between the different exercise types were evaluated using the Mann–Whitney U test. Results indicated that there are significant differences between the aerobic and resistance exercises (

p ≪ 0.05

), as well as between resistance and interval exercises (p ≪ 0.05). However, no significant difference was observed between aerobic and interval exercises (

p = 0.26

), within this specific dataset.

The $Δ$ CGM values across different exercise types indicate that aerobic activities resulted in a larger mean glucose drop compared with resistance activities, while interval exercises have an intermediate effect, more similar to aerobic, consistent with previously reported findings.²¹ This analysis of the glucose dynamics provides valuable insights that will be exploited to further validate the developed classifier.

Signal processing and spectrogram generation

The spectrogram is a powerful tool for analyzing nonstationary signals whose frequency components vary over time. Unlike the traditional Fourier transform, which assumes stationarity, the spectrogram provides a time–frequency representation that captures the temporal evolution of a signal's spectral content. This property makes it particularly suited for applications involving signals with time-varying characteristics, such as physiological time series.

The short-time Fourier transform (STFT) is used to compute the spectrogram representation. This signal processing technique serves to analyze the frequency and amplitude of localized waves. It is particularly useful for signals with time-varying characteristics, allowing the examination of changes in frequency and amplitude over time. For a discretized digital signal $x [n]$ , the STFT is calculated as follows³¹:

STFT {x [n]} = X (i, f) = \sum_{n = - \infty}^{\infty} x [n] \cdot w [n - i] \cdot e^{- j 2 π f n},

(2)

where

x [n]

is the input signal,

i

represents the discrete time index,

f

is the frequency variable in Hertz, and

w [n]

is the window function used to segment the signal into overlapping time frames. In this study, a Hanning window was selected, defined as follows:

w [n] = \{\begin{matrix} 0.5 [1 - \cos (\frac{2 π n}{n_{fft} - 1})], & 0 \leq n \leq n_{fft} - 1 \\ 0, & o therwise, \end{matrix}

(3)

where

n_{fft}

represents the window length. In practice, to calculate the STFT, the fast Fourier transform (FFT) algorithm is applied to each window of the input signal

x [n]

The spectrogram is a two-dimensional representation where the columns represent time, the rows represent frequencies, and the element magnitude reflects the energy of the signal, represented with a color intensity scale. Each element of this matrix is obtained by calculating the STFT and converting it to the decibel scale as follows:

s_{m, f} = 20 \cdot \log_{10} | X (m \cdot h {op}_{lenght}, f) | \cdot

(4)

Here, $s_{m, f}$ is the element at the $m$ th window index (column) and $f$ th frequency index (row) of the spectrogram matrix, and the $h o p_{length}$ parameter represents the displacement, in samples, between successive time frames or window positions.

The window size ( $n_{fft}$ ), and the number of samples that the window shift ( $h o p_{length}$ ) are critical parameters in the spectrogram generation process. The $n_{fft}$ determines the frequency resolution, while the $h o p_{length}$ controls the temporal resolution and overlapping between consecutive windows. These parameters will be further explored in the design of the classification model in the “Classifier Design” section.

CNN for image classification

CNNs are a specialized class of artificial neural networks designed to process grid-structured data, such as images. Figure 1 depicts a standard CNN architecture, consisting of multiple stages that analyze and transform input data. The process begins with convolutional and pooling layers, where learnable filters extract meaningful features from the input. These features are gradually refined through repeated convolutional and pooling stages, allowing the network to capture increasingly complex patterns.

FIG. 1.

Typical convolutional neural network architecture for processing grid-structured data.

The key to their success lies in the convolutional layers, which act as powerful feature extractors by applying sliding filters across the input. These filters identify localized patterns and hierarchical features that are critical for classification tasks, making CNNs particularly effective for spatial data such as spectrograms.³⁷ In the final stage, the output of these layers is passed to a fully connected network, where weights are adjusted, and activation functions are applied to produce the final prediction or classification.

Convolutional layers perform the core operation by convolving the input with filter kernels to generate feature maps, enabling the network to identify and leverage patterns within the data. Mathematically, this operation can be expressed as follows¹⁶:

c_{i, j}^{l} = H (b_{k}^{l} + \sum_{m = 1}^{M} \sum_{n = 1}^{N} w_{m, n}^{l, k} x_{i + m - 1, j + n - 1}^{l - 1, k}),

(5)

where

l

represents the layer index,

M

and

N

are the dimensions of the filter, or

kerne l_{size}

b_{k}^{l}

is the bias of the

k

-th feature map, and

w_{m, n}^{l, k}

are the weight parameters of the filter for the layer input

x_{i + m - 1, j + n - 1}^{l - 1, k}

. The activation function

H

, typically a Rectified Linear Unit (ReLU), introduces nonlinearity into the network, ensuring that only positive values are passed through while effectively setting negative values to zero.

After the convolutional layers, the feature maps are flattened into a one-dimensional vector and passed through fully connected layers. These layers establish dense connections between all neurons in successive layers, enabling high-level feature integration for classification. The fully connected layer can be represented as follows:

y = H (W^{⊤} x + b) = H (\sum_{i = 1}^{N} w_{i} x_{i} + b),

(6)

where

x

is the input vector,

y

is the output vector,

W

is the weight matrix,

b

is the bias, and

H

is the nonlinear activation function; again, the ReLU function is usually used.

The final layer of the CNN applies the softmax activation function to the last fully connected layer. This softmax function converts the output of the fully connected layer into a probability distribution over the target classes:

P (c | y) = \arg \max_{c \in C} \frac{\exp (y_{c})}{\sum_{i = 1}^{N_{C}} \exp (y_{i})},

(7)

where

c

is the class, and

N_{C}

is the total number of classes.

In this work, the design of the CNN architecture will focus on the optimal configuration of the convolutional layers and fully connected layers, without the use of pooling layers. Indeed, pooling layers are commonly used in CNN design to reduce dimensionality. However, in the current approach, the spectrogram dimensions are already small from their generation. Therefore, pooling layers are not employed to preserve spatial resolution and maintain a high resolution for the effective extraction of relevant features.

The specific selection of filter sizes, the number of filters, and other hyperparameters will be explored in the next section to adapt the CNN model to the dimensions of the input spectrogram images.

Classifier Design

This section details the design of the CNN-based classifier, including the spectrogram generation process, modifications for online operation, and the training and validation strategy.

Spectrogram generation process

Typically, the architecture of the CNN for the classification task is designed based on the input size. This is because the application of convolutional filters decreases the size of the image throughout the network when no padding (adding extra pixels around the leading edges to preserve their dimensions) is applied to the input arrays.

As mentioned in the “Signal Processing and Spectrogram Generation” section, the pair of parameters ( $n_{fft}$ , $h o p_{length}$ ) defines the size of the sliding window used for computing the FFT on each column of the spectrogram, as well as the step size by which this window is shifted. These parameters directly influence the dimensions of the generated spectrograms in the following way:

n_{rows} = n_{fft} / 2 + 1,

(8)

n_{cols} = \frac{signa l_{length} + h o p_{length}}{h o p_{length}},

(9)

where

n_{rows}

and

n_{cols}

correspond to the number of rows and columns, respectively. The variable

signa l_{length}

refers to the total number of samples in the input signal.

An ad hoc process is proposed for designing an effective parameter pair. This approach allows a CNN to adjust its first layer according to the input size while maintaining the structure and configuration of the subsequent layers unchanged, including their defined input and output dimensions. The proposed process will be detailed in the context of classifying 30-min signal spectrograms. Some network hyperparameters that define the architecture are fixed, while others remain variable depending on the model input. Figure 2 shows the CNN architecture along with the most important aspects of this design.

FIG. 2.

Convolutional neural network design for spectrogram generation process using 30-min signals.

In each convolutional layer $l$ , the size of the kernel ( $kerne l_{size}$ ) applied to the input determines the size of the stage output in each dimension $d$ . The stage output size in each dimension is calculated as follows:

Outpu t_{size}^{l, d} = \frac{inpu t_{size}^{l, d} - kerne l_{size}^{l, d} + 2 \cdot paddin g^{l}}{strid e^{l, d}} + 1,

(10)

where

l

represents the convolutional layer index and

d

represents the dimension, the stride (the step size of the convolutional filter) is set to 1, and the padding to 0. For more information on these hyperparameters, refer to LeCun et al.³⁷ For the fixed part of the network, an architecture consisting of two convolutional layers for feature extraction and two fully connected (dense) layers for classification is selected. Each convolutional layer uses 16 filters, while the dense layers include 16 neurons in the first layer and 3 neurons in the second. All layers utilize the ReLU activation function, except for the final dense layer, which employs a softmax activation function to output the class probabilities for the three categories.

The kernel, input, and output sizes for each convolutional layer are defined as follows. For the second layer, a fixed output size of $5 \times 5$ pixels and a kernel size of $3 \times 3$ are specified. Since the input size of the second layer is equal to the output size of the first layer, applying Equation (10) yields an output size of $7 \times 7$ pixels for the first layer. The first layer takes the HR and step count spectrograms as inputs, whose dimensions are variable and need to be determined to optimize the network's performance. Once the input size is established, the kernel size for the first layer is calculated using Equation (10), resulting in dimensions of $(n_{rows} - 6) \times (n_{cols} - 6)$ . Consequently, the minimum required spectrogram size for the HR and step count inputs (to the first layer) is determined to be $7 \times 7$ pixels. The maximum spectrogram size is determined by the signal length used for its calculation.

With the CNN architecture defined, the ranges for $h o p_{length}$ and $n_{fft}$ can be determined. Using Equation (9), the maximum and minimum values for $h o p_{length}$ are calculated as $h o p_{length} = signa l_{length} / (n_{cols} - 1)$ . Recalling that the sampling time of the signals is 10 s, for the 30-min case (180 samples) and a minimum spectrogram size of 7, the maximum $h o p_{length}$ is 5 min (30 samples), as larger shifts would violate the minimum spectrogram size. The minimum $h o p_{length}$ is set to 1 min (6 samples), defining the range as [6,30] samples. Using Equation (8), the range for $n_{fft}$ can also be obtained analytically in a similar manner. The maximum $n_{fft}$ is constrained by the signal length, resulting in 30 min (180 samples). The minimum $n_{fft}$ value is determined by the minimum spectrogram size of 7, resulting in 1 min (12 samples). Thus, the range for $n_{fft}$ is defined as [12,180] samples.

Finally, multiple network trainings and evaluations were performed to find the best combination of $n_{fft}$ and $h o p_{length}$ . This process was repeated five times for each of the 16 possible combinations, obtained by dividing in width and height the possible values into 4. In the “Training Strategy” section, the training scheme and evaluation methodology are described in detail. From this process, the following values result: an $n_{fft}$ of 126 samples, or 21 min, and a $h o p_{length}$ of 22 samples, or 3.7 min, resulting in spectrograms of $64 \times 9$ pixels.

Modifications for online classification

While the approach of using 30-min signal spectrograms proved to be straightforward for spectrogram generation, this classifier is not suitable for variable-duration real-time applications, as it can only provide a classification after 30 min of the PA session. This characteristic further restricts the proposal evaluation with unstructured PA, as seen in Dataset US.

To extend the methodology to a variable-duration approach, more appropriate for potential integration into real-life applications, a voting-based CNN classifier using 5-min spectrograms with a 2-min step between consecutive spectrogram calculations was developed. Therefore, after the initial 5-min classification, a probable class value will be obtained every 2 min. At any given time during the activity, the “current” classification will be based on the most frequently selected class by this classifier since the beginning of the session. The voting decision is determined as follows: if either the aerobic or resistance class receives more than 45% of the votes, that class is selected. Otherwise, the interval class is assigned. This process addresses the challenges inherent to the interval class, which often combines aerobic and resistance components depending on the exercise protocol.⁷ Consequently, given its mixed nature, the session is classified as intervals if either the intervals class exceeds 45% of the votes or if neither of the remaining two options (aerobic and resistance) exceeds the 45% threshold for a definitive classification. Additionally, the initial period of 5 min is considered a warm-up and therefore discarded.

The design of the CNN architecture and the selection of the ( $n_{fft}$ , $h o p_{length}$ ) pair followed the same procedure described in the previous subsection, using the same architecture but with different fixed hyperparameter values: the second convolutional layer has a $3 \times 3$ output with a $2 \times 2 kerne l_{size}$ , while the first layer uses a $kerne l_{size}$ dependent on the input $(n_{rows} - 3) \times (n_{cols} - 3)$ . This structure resulted in a minimum spectrogram size of $4 \times 4$ pixels, determined by evaluating Equation (10). Subsequently, by applying Equations (8) and (9), and considering a signal length of 5 min or 30 samples, the ranges for the parameters were determined to be $6 \leq n_{fft} \leq 30$ samples and $1 \leq h o p_{length} \leq 10$ samples. Finally, the pair was experimentally set to $n_{fft} = 30$ and $h o p_{length} = 4$ .

Training strategy

Both the 30- and 5-min spectrogram-based classifiers were trained following a structured scheme, with a 60/20/20% split for training, validation, and testing, respectively. Overall, 60% of the samples were used for adjusting the classifier’s weights or parameters, while the 20% validation set was used to monitor the model’s performance during the training process. Finally, the remaining 20% that the CNN had not seen during training was used for the final testing of the trained model. This training scheme was intended to avoid overfitting.

Due to class imbalance in the sessions, which included 591 aerobic, 590 interval, and 635 resistance exercise sessions, a class-weighted loss function was applied to address this issue. The class weights were defined as the inverse of the normalized training set counts.

The training was conducted using the TensorFlow/Keras framework in Python, employing the Adam optimizer and utilizing a loss function suitable for multiclass classification with sparse labels, specifically the Categorical Crossentropy. The models were trained for 100 epochs with a batch size of 16. Additionally, a “callback” function was implemented to stop the training when the validation performance started to decrease in order to avoid overfitting.

Results and Discussion

This section presents the performance evaluation of the developed CNN-based activity type classifiers. The results include their performance on structured exercise sessions using both the 30-min and the 5-min classifiers. Additionally, the variable-duration real-time 5-min classifier is validated using unstructured PA data.

Classifiers performance on structured sessions

The proposed classifiers performance is evaluated using the Dataset SS test subset. The overall accuracy, which is the percentage of correct predictions over the total number of samples, is reported. The “recall” for each class is also provided, representing the percentage of correct predictions for a given class according to the true label. Additionally, the confusion matrix is presented, displaying the counts of exercise sessions classified by the model versus the true labels of those sessions.

The confusion matrices for the 30- and 5-min spectrogram-based classifiers are presented in Tables 1 and 2, respectively. The 30-min classifier achieved an overall accuracy of 82.5%, with class-specific recalls of 80.5% for aerobic, 74.1% for interval, and 92.1% for resistance exercise. This demonstrates the model's ability to accurately identify the different exercise types from the 30-min spectrograms. When transitioning to the computationally efficient 5-min variable-duration online classifier, the recall for interval sessions remained consistent at 72.4%. Resistance sessions showed an improvement, achieving a recall of 94.4%. However, a slight decrease was observed for aerobic sessions, with a recall of 65.3%, resulting in an overall score of 77%.

Table 1.

Confusion Matrix Showing the Predicted Class by the 30-Min Spectrogram Exercise Type Classifier Versus the True Class on the Test Set

No. of sessions (recall %)	Predicted class
True class	Aerobic	Interval	Resistance
Aerobic	95 (80.5)	14 (11.9)	9 (7.6)
Interval	21 (18.1)	86 (74.1)	9 (7.8)
Resistance	5 (4.0)	5 (4.0)	116 (92.1)

Table 2.

Confusion Matrix Showing the Predicted Class by the 5-Min Spectrogram Exercise Type Classifier Versus the True Class on the Test Set

No. of sessions (recall %)	Predicted class
True class	Aerobic	Interval	Resistance
Aerobic	77 (65.3)	33 (28.0)	8 (6.8)
Interval	13 (11.2)	84 (72.4)	19 (16.4)
Resistance	4 (3.2)	3 (2.4)	119 (94.4)

Although the recall for aerobic sessions decreased compared with the 30-min case, the value remains comparable to those reported in other studies, while achieving higher recall for interval and resistance classes.^15,27 Furthermore, aerobic sessions misclassified as nonaerobic are predominantly categorized as interval sessions, which have a similar impact on glucose levels, particularly in the T1DEXI case study, as shown in the previous subsections. Only 6.8% of aerobic sessions are misclassified as resistance, whose effect on glucose levels is expected to be either neutral or slightly elevating.

The analysis of the $Δ$ CGM is presented in Figure 3, which illustrates the distribution of $Δ$ CGM for the sessions labeled by the 30- and 5-min classifiers compared with the actual distribution of those sessions from the test set. The summary of $Δ$ CGM values in mean and standard deviation (mean ± SD) for structured exercise sessions labeled by both classifiers is shown in Table 3. This table demonstrates the close alignment between the predicted $Δ$ CGM by the 30- and 5-min classifiers and the actual observed glucose changes during the real structured exercise sessions. This consistency across the three exercise modalities suggests that the classifier indirectly captures the underlying characteristics of the glucose changes associated with the PA types. Despite the slight decrease in 5-min classification accuracy for the aerobic class compared with the 30-min case, the impact on glucose levels remains highly consistent with the real dataset across the three classes.

FIG. 3.

Distribution of $Δ$ CGM across different types of structured exercise: real and classified. CGM, continuous glucose monitoring.

Table 3.

Summary of $Δ$ CGM (mean ± SD) for Labeled Structured Exercise Sessions by Both Classifiers

	Aerobic	Interval	Resistance
Real	−17.5 ± 31.8	−12.1 ± 26.5	−4.4 ± 25.8
30-min	−17.0 ± 32.2	−11.6 ± 25.9	−5.6 ± 26.1
5-min	−16.7 ± 32.7	−11.4 ± 27.5	−7.4 ± 26.1

CGM, continuous glucose monitoring; SD, standard deviation.

It is important to highlight that even though the 5-min classifier provides useful information that has the potential to improve glucose control, considering the immediate effect of exercise, the 30-min classifier is also relevant due to the long-term effect of PA on glucose metabolism.

Validation for unstructured sessions

The 5-min online classifier is also validated on unstructured PA sessions using the Dataset US, which contains sessions of various durations. For these unstructured sessions, which were not used for training, the classification is based on the most frequently selected class by the CNN throughout the session. The unstructured sessions typically lack a predefined protocol and fixed duration, and they were labeled by the participants themselves based on the activity they performed. Examples of the self-categorized activities in Dataset US include Walking, Weightlifting, Biking, and various other daily activities or sports. These sessions inherently carry a higher degree of uncertainty due to variations in activity intensity (e.g., two individuals may walk or cycle at different speeds), regularity (the activity could be performed at a consistent intensity or include periods of higher and lower intensity), and, naturally, duration.

The mean glucose change and the statistical differences between the labeled activity classes are evaluated, in addition to the label provided by the patient. The distribution of $Δ$ CGM for the unstructured sessions, grouped by predicted class, is shown in Figure 4. The analysis reveals a mean glucose change of −26.7 ± 45.5 mg/dL for the aerobic class, −15.4 ± 41.1 mg/dL for the interval class, and –14.2 ± 45.7 mg/dL for the resistance class. Applying the Mann–Whitney U test reveals statistically significant differences between the aerobic and resistance classes ( $p ≪ 0.05$ ). These findings are consistent with previous literature²¹ and with the analysis of Dataset SS in the “Data Analysis and Preparation” section, which also demonstrated significant differences between these two activity types. However, in Dataset US, statistically significant differences are observed between the interval and aerobic classes ( $p ≪ 0.05$ ), whereas no significant differences are found between the interval and resistance classes ( $p = 0.81$ ). This contrasts with the results from Dataset SS, where significant differences were observed between the interval and resistance classes, but not between the interval and aerobic classes. This discrepancy can be attributed to two factors. First, the mixed nature of interval exercise, which combines aerobic and resistance components,⁷ may lead to intermediate effects on glucose dynamics, making the differences with other classes less consistent. Second, the distribution of predicted classes in Dataset US is unbalanced, with aerobic activities representing 51.0% of the total data, resistance activities accounting for 42.2%, and interval activities comprising only 6.8%. This imbalance could limit the statistical power to detect differences involving the interval class. Nevertheless, it is more relevant to differentiate between aerobic and resistance activities, as they have opposing effects on glucose levels. As shown in Figure 4, the overall trend remains consistent: Aerobic exercise leads to the greatest reduction in glucose levels, followed by interval exercise, and finally resistance exercise.

FIG. 4.

Distribution of $Δ$ CGM associated with unstructured exercise sessions predicted by the 5-min online spectrogram classifier.

Table 4 presents data for various sessions based on the patients’ reported labels. For this analysis, only activities with at least 100 reported cases were selected to ensure a representative and consistent dataset. The table includes, from left to right, the total number of sessions, the number (and percentage) of sessions classified as aerobic, interval, and resistance, followed by the mean and SD of $Δ$ CGM for each class in the last three columns. Examining the classification columns for aerobic, interval, and resistance (3rd to 5th columns) in relation to the patients’ self-reported labels, it is evident that the classifier demonstrates the expected behavior. Activities such as Walking, Jogging/Running, Hiking, and Cardio are predominantly classified as aerobic, while Biking, Weightlifting, Outdoor chores, Cleaning/Housework, and Pilates/Yoga fall under the resistance category. On the other hand, activities like CrossFit/Fitness Class and Interval training do not exhibit a clear classification. These latter activities require special consideration due to their inherent variability, which depends heavily on the specific format and intensity applied by the patient.

Table 4.

Analysis of Self-Reported Activity Categories from Unstructured Sessions: Percentage of Sessions Assigned to Each Exercise Class by the 5-Min Online Spectrogram Classifier, and the Corresponding Mean ± SD of the Glucose Change ( $Δ$ CGM)

Category	# Sessions	#_Aerobic (%)	#_Interval (%)	#_Resistance (%)	$Δ$ CGM_Aerobic	$Δ$ CGM_Interval	$Δ$ CGM_Resistance
Walking	2015	1591 (79.0)	64 (3.2)	360 (17.9)	−27 ± 41	−16 ± 30	−3 ± 39
Biking	485	76 (15.7)	17 (3.5)	392 (80.8)	−36 ± 46	−23 ± 42	−34 ± 48
Weightlifting	404	25 (6.2)	61 (15.1)	318 (78.7)	−15 ± 49	−15 ± 40	−9 ± 38
Jogging/Running	381	338 (88.7)	7 (1.8)	36 (9.4)	−30 ± 50	−27 ± 41	4 ± 49
Outdoor Chores	293	28 (9.6)	31 (10.6)	234 (79.9)	−50 ± 57	−23 ± 38	−23 ± 44
Cleaning/Housework	277	19 (6.9)	15 (5.4)	243 (87.7)	−15 ± 39	1 ± 33	−9 ± 42
Hiking	200	161 (80.5)	8 (4.0)	31 (15.5)	−36 ± 47	3 ± 55	18 ± 74
CrossFit/Fitness Class	196	57 (29.1)	53 (27.0)	86 (43.9)	−18 ± 43	−15 ± 37	−9 ± 35
Pilates/Yoga	166	3 (1.8)	2 (1.2)	161 (97.0)	1 ± 12	14 ± 21	−12 ± 35
Interval Training	154	51 (33.1)	35 (22.7)	68 (44.2)	0 ± 39	−5 ± 36	−5 ± 33
Cardio	118	73 (61.9)	6 (5.1)	39 (33.1)	−31 ± 47	−57 ± 33	−34 ± 52

The observed patterns in $Δ$ CGM associated with the predicted exercise reveal that the expected relationship (6th–8th columns of Table 4), $Δ$ CGM_Aerobic < $Δ$ CGM_Interval < $Δ$ CGM_Resistance, is consistently observed for several activity categories, such as Walking, Jogging/Running, Hiking and CrossFit/Fitness Class. In other activities such as Weightlifting, Outdoor Chores, and Cleaning/Housework, the $Δ$ CGM values generally follow the expected trend. However, for the interval class, the values are closer to one of the other two categories, and in the case of Cleaning/Housework, the trend is not observed specifically for the interval class. These discrepancies may be attributed to the unique characteristics of the interval class, as previously discussed. For the Biking category, which does not follow the expected pattern for the resistance class, this could be due to certain rides that should have been classified as aerobic but were instead categorized as resistance, likely due to the lack of step count information. Finally, the last three categories in the table, having fewer representatives, exhibit more variability, likely due to greater sensitivity to outliers. Nevertheless, when analyzed by category and considering the percentage distribution across classes, it is reasonable that Pilates/Yoga is primarily classified as resistance, Interval Training spans all three classes depending on the dominant activity, and Cardio is predominantly classified as aerobic. Overall, the classifier’s performance demonstrates consistency with the self-reported labels and the expected trends in $Δ$ CGM across the different classes, further validating its reliability.

It is reasonable to suppose that an announcement of the exercise event by the user with an activity category could aid the proposed method in providing a preliminary classification. That is, if the user reported Walking as the activity, a preliminary class could be aerobic. Nonetheless, Table 4 demonstrates that, for instance, 79.0% of the sessions labeled as Walking were actually classified as aerobic. Meanwhile, 17.9% of the Walking sessions were classified as resistance. The mean $Δ$ CGM for the walking sessions classified as resistance was −3 mg/dL, which is consistent with the expected glucose behavior in resistance sessions. This highlights the importance of utilizing the continuous signals available from the wearable devices (such as HR) to predict the actual effect of an exercise event on the subject’s metabolism.

It is worth noting that the choice between the 5- and 30-min classifiers depends on the application’s objectives and the activity duration being analyzed. The 5-min spectrogram-based classifier proves useful in the context of an AID, where online adjustments of insulin administration are needed, as well as for activities with indefinite durations. Conversely, the 30-min classifier, which demonstrated superior performance in the classification of fixed-duration sessions, could be more suitable for analyzing the long-term exercise effects. This is particularly relevant given that the residual exercise impact on glucose levels can persist for up to 24 h postactivity.⁷ Moreover, considering the clinical recommendation of 150 min of PA per week, typically distributed as 30-min sessions on weekdays,⁷ the 30-min classifier is better suited for these fixed-duration activities and postactivity evaluations.

It is important to note that the proposed classification approach offers practical advantages in terms of computational efficiency and real-time capabilities. Recent studies have demonstrated the feasibility of deploying CNN architectures on low-power computing devices for various applications,³⁸ including spectrogram-based audio classification tasks.³⁹ Compared with RNNs, like LSTMs, the feedforward structure of CNNs results in lower computational complexity, making them more suitable for deployment in practical, real-world scenarios where rapid, low-latency inference is crucial. Furthermore, the use of spectrograms as the input representation has been shown to be feasible on low-power devices. This suggests that the CNN-based activity recognition classifier, utilizing spectrogram-based feature extraction, can be effectively implemented in resource-constrained settings, such as sports wearable devices or insulin pumps, while maintaining real-time performance.

Conclusions

A CNN-based classifier capable of accurately identifying the type of PA (aerobic, interval, resistance) performed by individuals with T1D was proposed and validated with the largest clinical trial of PA in T1D to date. The 30-min spectrogram-based classifier achieved an overall accuracy of 82.5% on structured sessions, demonstrating the model’s ability to effectively distinguish between the different exercise modalities. To enable variable-duration real-time applications, the methodology was extended to a computationally efficient 5-min online classifier, which maintained a good global accuracy of 77% on the structured session test set. Additionally, when applied to unstructured PA sessions, the $Δ$ CGM distribution of the labeled sessions by the 5-min classifier was consistent with clinical evidence, showing significantly greater decreases for the sessions classified as aerobic, intermediate decreases for interval, and smaller decreases for resistance exercises.

The methodology for generating spectrograms and designing the CNN architecture constitutes a valuable contribution, serving as a foundation for optimizing performance and advancing the construction of spectrogram-based classifiers. Furthermore, the low computational requirements of the real-time classifier make it suitable for deployment in resource-constrained settings, such as wearable devices.

The observed consistency between the predicted activity types and their associated glycemic responses highlights the classifier’s potential impact on PA management, offering a step forward to enable real-time exercise guidance and control for individuals with T1D.

Nevertheless, the proposed methodology assumes prior announcement or detection of activity initiation, which could limit its applicability in unstructured settings. As a future direction, incorporating a “no PA” class could enhance the classifier's robustness and broaden its use in diverse contexts, or using an exercise detector.^15,40

Next steps should prioritize the integration of this technology into real-time clinical systems, validating its performance across diverse populations and settings. These efforts will harness the potential of CNN-based activity classifiers to improve glucose management and advance care in T1D.

Authors’ Contributions

M.D.S.: Conceptualization, writing—original draft, methodology, validation, formal analysis, investigation, data curation, and visualization. F.A.I., E.F., and F.G.: Conceptualization, writing—review and editing, supervision, and funding acquisition. All authors read and approved the final article.

Footnotes

Abbreviations

References

DiMeglio

, Evans-Molina

, Oram

. Type 1 diabetes. Lancet, 2018; 391(10138):2449–2462; doi: 10.1016/S0140-6736(18)31320-5

Stults-Kolehmainen

, Sinha

. The effects of stress on physical activity and exercise. Sports Med, 2014; 44(1):81–121; doi: 10.1007/s40279-013-0090-5

Abdel-Latif

, Askari

, Rashid

, et al. Multi-task classification of physical activity and acute psychological stress for advanced diabetes treatment. Signals, 2023; 4(1):167–192; doi: 10.3390/signals4010009

Ming

, Guo

, Zhang

, et al. Recent advances in the precision control strategy of artificial pancreas. Med Biol Eng Comput, 2024; 62(6):1615–1638; doi: 10.1007/s11517-024-03042-x

Chushig-Muzo

, Calero-Díaz

, Fabelo

, et al. Characterizing the impact of physical activity on patients with Type 1 Diabetes using statistical and machine learning models. Appl Sci, 2024; 14(21); doi: 10.3390/app14219870

Moser

, Zaharieva

, Adolfsson

, et al. The use of automated insulin delivery around physical activity and exercise in Type 1 Diabetes: A position statement of the European Association for the Study of Diabetes (EASD) and the International Society for Pediatric and Adolescent Diabetes (ISPAD). Horm Res Paediatr, 2024; 1–28; doi: 10.1159/000542287

Riddell

, Gallen

, Smart

, et al. Exercise management in type 1 diabetes: a consensus statement. Lancet Diabetes Endocrinol, 2017; 5(5):377–390; doi: 10.1016/S2213-8587(17)30014-1

Behrens

, Brinkmann

. Exercise and type 1 diabetes mellitus. Ger J Sports Med, 2023; 74(7):242–247; doi: 10.5960/dzsm.2023.579

Mosquera-Lopez

, Wilson

, El Youssef

, et al. Enabling fully automated insulin delivery through meal detection and size estimation using Artificial Intelligence. Npj Digit Med, 2023; 6(1):39; doi: 10.1038/s41746-023-00783-1

10.

, Daniels

, Liu

, et al. Convolutional recurrent neural networks for glucose prediction. IEEE J Biomed Health Inform, 2020; 24(2):603–613; doi: 10.1109/JBHI.2019.2908488

11.

Chetty

, Shetty

, Fournier

, et al. Exercise management for young people with type 1 diabetes: a structured approach to the exercise consultation. Front Endocrinol, 2019; 10:326; doi: 10.3389/fendo.2019.00326

12.

Yardley

, Hay

, Abou-Setta

, et al. A systematic review and meta-analysis of exercise interventions in adults with type 1 diabetes. Diabetes Res Clin Pract, 2014; 106(3):393–400; doi: 10.1016/j.diabres.2014.09.038

13.

Reddy

, Wittenberg

, Castle

, et al. Effect of aerobic and resistance exercise on glycemic control in adults with Type 1 Diabetes. Can J Diabetes, 2019; 43(6):406–414.e1; doi: 10.1016/j.jcjd.2018.08.193

14.

El Fatouhi

, Héritier

, Allémann

, et al. Associations between device-measured physical activity and glycemic control and variability indices under free-living conditions. Diabetes Technol Ther, 2022; 24(3):167–177; doi: 10.1089/dia.2021.0294

15.

Cho

, Aiello

, Ozaslan

, et al. Design of a real-time physical activity detection and classification framework for individuals with Type 1 Diabetes. J Diabetes Sci Technol, 2024; 18(5):1146–1156; doi: 10.1177/19322968231153896

16.

Hsu

Y-L

, Chang

H-C

, Chiu

Y-J

. Wearable sport activity classification based on deep convolutional neural network. IEEE Access, 2019; 7:170199–170212; doi: 10.1109/ACCESS.2019.2955545

17.

Jacobs

, Resalat

, El Youssef

, et al. Incorporating an exercise detection, grading, and hormone dosing algorithm into the artificial pancreas using accelerometry and heart rate. J Diabetes Sci Technol, 2015; 9(6):1175–1184; doi: 10.1177/1932296815609371

18.

Cescon

, Choudhary

, Pinsker

, et al. Activity detection and classification from wristband accelerometer data collected on people with type 1 diabetes in free-living conditions. Comput Biol Med, 2021; 135:104633; doi: 10.1016/j.compbiomed.2021.104633

19.

Turksoy

, Hajizadeh

, Hobbs

, et al. Multivariable artificial pancreas for various exercise types and intensities. Diabetes Technol Ther, 2018; 20(10):662–671; doi: 10.1089/dia.2018.0072

20.

Sevil

, Rashid

, Maloney

, et al. Determining physical activity characteristics from wristband data for use in automated insulin delivery systems. IEEE Sens J, 2020; 20(21):12859–12870; doi: 10.1109/jsen.2020.3000772

21.

Riddell

, Li

, Gal

, et al. Examining the acute glycemic effects of different types of structured exercise sessions in Type 1 diabetes in a real-world setting: The Type 1 Diabetes and Exercise Initiative (T1DEXI). Diabetes Care, 2023; 46(4):704–713; doi: 10.2337/dc22-1721

22.

Mosquera-Lopez

, Ramsey

, Roquemen-Echeverri

, et al. Modeling risk of hypoglycemia during and following physical activity in people with type 1 diabetes using explainable mixed-effects machine learning. Comput Biol Med, 2023; 155:106670; doi: 10.1016/j.compbiomed.2023.106670

23.

Bergford

, Riddell

, Jacobs

, et al. The Type 1 Diabetes and Exercise initiative: Predicting hypoglycemia risk during exercise for participants with Type 1 Diabetes using repeated measures random forest. Diabetes Technol Ther, 2023; 25(9):602–611; doi: 10.1089/dia.2023.0140

24.

Young

, Dodier

, Youssef

, et al. Design and in silico evaluation of an exercise decision support system using digital twin models. J Diabetes Sci Technol, 2024; 18(2):324–334; doi: 10.1177/19322968231223217

25.

Crespo-Santiago

, Cescon

. Physiology-informed deep learning modeling of Type 1 Diabetes dynamics: mapping data to virtual subjects. IFAC-Pap, 2024; 58(15):235–240; doi: 10.1016/j.ifacol.2024.08.534

26.

Turner

, Marak

, Gal

, et al. Associations between daily step count classifications and continuous glucose monitoring metrics in adults with type 1 diabetes: analysis of the Type 1 Diabetes Exercise Initiative (T1DEXI) cohort. Diabetologia, 2024; 67(6):1009–1022; doi: 10.1007/s00125-024-06127-2

27.

Fushimi

, Aiello

, Cho

, et al. Online classification of unstructured free-living exercise sessions in people with Type 1 Diabetes. Diabetes Technol Ther, 2024; 26(10):709–719; doi: 10.1089/dia.2023.0528

28.

Kumar Shukla

, Rakhra

, Singh

, et al. The role of machine learning in health care diagnosis. In: 2022 4th International Conference on Artificial Intelligence and Speech Technology (AIST). 2022; pp. 1–6; doi: 10.1109/AIST55798.2022.10064906

29.

Woldaregay

, Årsand

, Walderhaug

, et al. Data-driven modeling and prediction of blood glucose dynamics: Machine learning applications in type 1 diabetes. Artif Intell Med, 2019; 98:109–134; doi: 10.1016/j.artmed.2019.07.007

30.

Ravi

, Wong

, Deligianni

, et al. Deep learning for health informatics. IEEE J Biomed Health Inform, 2017; 21(1):4–21; doi: 10.1109/JBHI.2016.2636665

31.

Huang

, Chen

, Yao

, et al. ECG Arrhythmia classification using STFT-Based spectrogram and convolutional neural network. IEEE Access, 2019; 7:92871–92880; doi: 10.1109/ACCESS.2019.2928017

32.

Mandhouj

, Cherni

, Sayadi

. An automated classification of EEG signals based on spectrogram and CNN for epilepsy diagnosis. Analog Integr Circuits Signal Process, 2021; 108(1):101–110; doi: 10.1007/s10470-021-01805-2

33.

Das

, Ghosh

, Guha

, et al. Early detection of diabetes based on skin impedance spectrogram and heart rate variability noninvasively. In: 2017 1st International Conference on Electronics, Materials Engineering and Nano-Technology (IEMENTech). 2017; pp. 1–5; doi: 10.1109/IEMENTECH.2017.8076984

34.

Yildirim

, Talo

, Ay

, et al. Automated detection of diabetic subject using pre-trained 2D-CNN models with frequency spectrum images extracted from heart rate signals. Comput Biol Med, 2019; 113:103387; doi: 10.1016/j.compbiomed.2019.103387

35.

Jacobs

, Herrero

, Facchinetti

, et al. Artificial intelligence and machine learning for improving glycemic control in diabetes: best practices, pitfalls, and opportunities. IEEE Rev Biomed Eng, 2024; 17:19–41; doi: 10.1109/RBME.2023.3331297

36.

Rehman

, Contreras

, Beneyto

, et al. The impact of missing continuous blood glucose samples on machine learning models for predicting postprandial hypoglycemia: an experimental analysis. Math Basel, 2024; 12(10); doi: 10.3390/math12101567

37.

LeCun

, Kavukcuoglu

, Farabet

. Convolutional Networks and Applications in Vision. In: Proceedings of 2010 IEEE International Symposium on Circuits and Systems 2010; pp. 253–256; doi: 10.1109/ISCAS.2010.5537907

38.

Estrebou

, Fleming

, Saavedra

, et al. Lightweight convolutional neural networks framework for really small TinyML Devices. In: Smart Technologies, Systems and Applications. ( Narváez

, Proaño

, Morillo

, et al. eds) Springer International Publishing: Cham; 2022; pp. 3–16; doi: 10.1007/978-3-030-99170-8_1

39.

Pau

, Naramo

, Dimbiniaina

. Coding Mel Spectrogram Using Keras and Tensorflow for Home Appliances Tiny Classification. In: 2023 IEEE International Conference on Consumer Electronics (ICCE). 2023; pp. 1–5; doi: 10.1109/ICCE56470.2023.10043378

40.

Yang

C-C

, Hsu

Y-L

. A review of accelerometry-based wearable motion detectors for physical activity monitoring. Sens Basel, 2010; 10(8):7772–7788; doi: 10.3390/s100807772