Page 51 - Read Online

Page 30 of 44 Jung et al. Soft Sci 2024;4:15 https://dx.doi.org/10.20517/ss.2024.02

backpropagation, is critical for the ANN to learn from the data. As the network processes more data, it fine-
tunes its weights, enhancing its predictive accuracy. In DM monitoring, various types of ANNs, such as
recurrent neural networks (RNN), long short-term memory (LSTM) and convolutional neural networks
(CNN), have been employed. Among them, Gu et al. presented the SugarMate, a smartphone-based blood
glucose inference system, which offers a non-invasive alternative to CGM . It integrated smartphone
[272]
sensor data with food, drug, and insulin records to measure physical activity and sleep quality. Facing
challenges of imbalanced and limited data, SugarMate employed Md RNN, a deep RNN model with
3
grouped input layers, for fine-grained blood glucose level inference. This model utilized limited personal
and grouped-user data, achieving 82.14% accuracy in evaluations with 112 participants. Another recurrent
convolutional neural network model was developed for precise blood glucose level forecasting in type 1
[274]
diabetes (T1D) . The model achieves a root-mean-square error (RMSE) of 9.38 ± 0.71 mg/dL over 30 min
and 18.87 ± 2.25 mg/dL over 60 min for simulated cases, and 21.07 ± 2.35 mg/dL for 30 min, 33.27% ± 4.79%
for 60 min in real patient cases. In terms of LSTM-based models, Beauchamp et al. introduced a novel
approach for managing blood glucose levels in T1D using a dual LSTM architecture . This model inverted
[273]
the traditional “what-if” scenario, instead providing recommendations for insulin dosages or carbohydrate
intake to achieve target glucose levels. The technique was further enhanced by embedding the LSTM chain
within a deep residual architecture, optimizing time series forecasting. This model demonstrated significant
improvements for self-managing blood glucose levels in T1D over conventional methods by testing with
real patient data. CNN was also used to forecast accurate glucose levels. Li et al. introduced GluNet, a
[271]
personalized deep neural network framework for accurate short-term blood glucose prediction in T1D .
GluNet employed deep learning techniques and historical data, including glucose measurements, meals, and
insulin doses. In simulations, it achieved a RMSE of 8.88 ± 0.77 mg/dL for 30-minute predictions and
19.90 ± 3.17 mg/dL for 60-minute predictions. Clinical data tests also demonstrated its effectiveness in
glucose forecasting.

Support vector machine
Support vector machines (SVM) are a type of supervised learning algorithm applicable to both classification
and regression tasks. For classification, SVM distinguishes between two distinct classes by determining the
most appropriate hyperplane that divides them, informed by the labeled training data. Depending on the
task complexity and inter-feature relationships, SVM can employ either linear separation or more complex,
non-linear methods. Non-linear classification challenges are managed by kernel functions that elevate the
data into a higher-dimensional space, making the problem linearly separable. When it comes to regression,
SVM aims to establish a function that lies within an acceptable deviation from the given targets, prioritizing
the minimization of errors with a more flexible margin. This approach is similar to non-linear regression,
wherein kernel functions are again leveraged to transform the data, thereby facilitating a linear approach to
an intrinsically non-linear regression problem. For forecasting glycemia, CGM study from 25 DM type 1
patients was carried out in order to explore the viability of on-device computations with SVM . This study
[268]
found that a smartphone could predict glucose levels over a 15-minute horizon with a RMSE of
19.90 mg/dL within 34.89 s. Moreover, for hypoglycemia prediction of DM type 2 patients, Sudharsan et al.
trained a probabilistic model using self-monitored blood glucose (SMBG) values and validated their model
that enables the prediction of a hypoglycemia event in the next 24 h with a sensitivity of 92 % and specificity
of 70% . Bertachi et al. examined using ML to predict nocturnal hypoglycemia (NH) in T1D patients on
[269]
multiple insulin doses . Ten adults were monitored over 12 weeks, with data from CGM and activity
[280]
trackers. ML models, particularly SVM, effectively anticipated NH with over 70% potential avoidance,
demonstrating high sensitivity and specificity. The study validates ML as a viable tool for predicting NH in
T1D management.