Page 52 - Read Online

Jung et al. Soft Sci 2024;4:15 https://dx.doi.org/10.20517/ss.2024.02 Page 31 of 44

Decision tree
A decision tree (DT) is a method used to simplify data classification through a series of binary decisions,
leading to a clear partition of data. It operates by splitting data at nodes based on attribute values until the
data in each partition is sufficiently similar. DTs are utilized extensively in data mining and ML for
regression and classification and are favored for their interpretability, offering valuable insights into key
data attributes. For discrete target values, classification trees are utilized, while regression trees handle
continuous ones. DTs can be adapted to various forms, including tree ensembles, to suit specific analytical
needs and outcomes. One study introduced a non-invasive sweat sensor that measures glucose levels in
eccrine sweat using electrochemical impedance spectroscopy . With just 1-5 µL of sweat, the device
[279]
provided real-time glucose measurements every few minutes. A ML algorithm, specifically DT regression,
was used to process sensor data from three human subjects, achieving a high accuracy (R² of 0.94 and RMSE
of 0.1 mg/dL).

Autoregressive integrated moving average
The autoregressive integrated moving average (ARIMA) model is a sophisticated linear method for
forecasting future values in time series data, extending from past values. It integrates differencing operators
with autoregressive and moving average components, making it a more generalized form of the
autoregressive moving average (ARMA) model. The strength of ARIMA lies in its ability to accommodate a
wide array of non-stationary series, thereby enhancing its applicability and accuracy in time series
forecasting. A study monitored five DM patients for four days and forecasted glycemia by adapting ARIMA
models with prediction horizons of 30 and 60 min, obtaining RMSEs of 22.9 and 42.2 mg/dL, respectively
[281] . In addition, the ARIMA model was employed to forecast future trends in blood glucose levels,
[282]
specifically targeting hypoglycemia and hyperglycemia . This predictive framework was developed using
CGM data from 100 patients, encompassing both T1D and type 2 diabetes (T2D). The model effectively
generated early warnings, achieving a high sensitivity of 100% while maintaining a low false rate of 9.4%.

Transfer learning
Transfer learning is a ML technique where a model developed for one task is reused as the starting point for
a model on a second task. This approach leverages pre-trained models to solve similar problems,
significantly addressing the issue related to the need for a large amount of training data. It is particularly
effective in deep learning, where substantial datasets and extensive training are typically required. Transfer
learning is instrumental in various applications, including image and speech recognition, where it utilizes
knowledge gained from one domain to improve learning in another, enhancing efficiency and performance
in the learning process. Regarding DM monitoring, a study developed deep-learning methods for predicting
blood glucose levels in T2D patients using continuous monitoring data, aiming to improve glycemic control
and reduce complications . The main challenges were small, imbalanced patient datasets and rare hypo/
[277]
hyperglycemic events. By employing transfer learning and data augmentation, including techniques such as
mix-up and generative models, the study achieved over 95% prediction accuracy and 90% sensitivity within
a clinically relevant one-hour horizon. The approach was also effective for T1D, demonstrating its broader
applicability. De Bois et al. also addressed the challenge of insufficient data in healthcare deep learning by
proposing a multisource adversarial transfer learning framework . This method enhanced data transfer
[278]
across multiple health sources, learning a generalized feature representation. Applied to diabetic glucose
forecasting, the approach showed improved statistical and clinical accuracies, particularly effective with
diverse datasets or limited data scenarios. The adversarial transfer learning framework outperformed
standard methods, promoting a more universal feature representation and improving deep model training
with shared healthcare data.