PPG is a low-cost, noninvasive technique that measures the volumetric fluctuation in arterial blood flow [38]. The human wrist is one of the sites for measuring the PPG signal because it has a rich arterial source and an excellent sensor placement with minimal interference to one’s daily activities. The PPG signal comprises superimposed pulsatile alternating current components and direct current voltage components. A PPG signal is obtained by illuminating the light emitting device on the skin surface and measuring the variations in light absorption or reflection that reflect the pulsatile flow patterns, as shown in Figure 1.

The pulsatile alternating current component corresponds to the cardiac cycle, characterizing that the wrist’s blood vessels expand and contract with each heartbeat, whereas the direct current component reflects constant light absorption by venous and arterial blood, as well as other tissues [39]. The PPG signal can detect vascular changes associated with diabetes and contains substantial valuable information from HRV, which is significantly associated with diabetes [40]. Hence, it will be used in this study to extract valuable and meaningful features to identify an individual’s glucose status (elevated or normal).

Before commencing the study, ethical clearance was obtained from the SingHealth Centralised Institutional Review Board of Singapore (2020/2968) on March 21, 2021. All methods were performed in accordance with Singapore’s clinical guidelines and regulations. Informed consent was obtained from all the trial participants or their legal guardians. The clinical trial was registered on ClinicalTrials.gov (NCT05504096) on August 17, 2022.

In total, 500 participants were recruited from KK Women’s and Children’s Hospital in Singapore. Participants’ demographics are summarized in Table 1. For most participants, the blood glucose levels were measured before and after consumption of 75 g of a sugary drink using both the conventional glucometer (Accu-Chek Performa) and the wrist-worn wearable device. Participants who were excluded for the second measurement had high blood glucose measurements ≥11.1 mmol/L on their first measurement and hence were not administered the sugary drink measuring 75 g.

After consuming the sugary drink, 55.1% (266/483) of the participants had high blood glucose (≥7.8 mmol/L). The distribution of blood glucose levels before and after consuming the sugary drink is shown in Figure 2. A statistically significant difference was observed between the 2 distributions (P<.001).

The Actxa Spark+ Series 2, a low-cost and commercially available wrist-worn wearable device, was used in this project. This multifunctional device, built for everyday activities, fitness, and preventive health monitoring, provided an adequate PPG signal quality at 50 Hz. The wearable device is equipped with advanced PPG technology that enables accurate and reliable measurement of heart rate (HR) and other physiological parameters. This is similar to the devices used in Singapore’s nationwide health care campaigns, such as the National Steps Challenge. It is also worth noting that our proposed solution is device agnostic and can be easily integrated into other wearables with PPG capabilities, allowing for a scalable and cost-effective assessment of risk-based populations, including high-risk participants, participants with undiagnosed diabetes, and patients in need of primary prevention interventions.

The raw PPG signal was collected using both wrist-worn wearables in 16-bit binary format. We first performed a digital-to-analog conversion using the following formula:

Liang et al [41] suggested that a fourth-order Chebyshev II filter provides an optimal processing performance for short PPG signals. Hence, we adopted the recommended filter design to remove low-frequency drift and high-frequency noise using a band-pass Chebyshev II filter. The proposed band-pass filter has a lower cut-off frequency of 0.3 Hz and an upper cut-off frequency of 4 Hz.

The filtered PPG signals still contain various forms of outliers, such as peaks with abnormally high amplitudes or distortions in the oscillating waveform, which can be caused by movement from the upper extremity or improper contact between the sensor and skin. Features derived from signals that possess outliers may not be accurate, so a z scores outlier detection with a cut-off value of 3 SDs of the mean was applied. The identified outliers or regions of outliers were replaced with a reasonable estimate via a nearest neighbor interpolation for the HRV feature extraction. Because PPG signals do not change drastically in such a short duration, this method is determined to be an appropriate approach to the problem. Furthermore, the number of outliers was minimal in our data set, and hence should not have affected the features that we generated later. The data preprocessing steps are illustrated in Figure 3.

All experiments and analyses were performed using Python (version 3.9) and relevant libraries (Table 2). The final model was deployed on Amazon Web Services.

Considering AI ethics and the practicality of implementing the algorithm, some demographic data, such as skin color, race, and personal lifestyle habits, were not used as inputs to the models. However, other general personal characteristics associated with the risk of developing T2DM, such as age, gender, BMI, and family health history of diabetes, were added to the feature vector before the feature-selection process.

The redundant or irrelevant features might hinder the performance of the prediction model. To reduce the dimensionality of the input features, we applied an ensemble strategy that uses multiple feature-selection algorithms. This creates an optimal feature subset that minimizes the prediction error rate and is the most relevant for predicting the target variable. The ensemble feature-selection steps are summarized as follows:

In total, 12 features were selected from the entire feature set and ranked based on the results of the feature-selection strategy (Table 3). In our study, these selected features were the most sensitive predictors for capturing the characteristics of a participant’s elevated blood glucose levels.

The selected features could be further divided into 4 main categories. Under the time-domain features, the selected features were the area under the PPG curves. A_FE_mean refers to the average area under the falling edge of each pulse (Figure 4). A_ratio refers to the ratio of the area under the rising edge to the area under the falling edge of each pulse (Figure 4), and both the average and maximum values were deemed relevant to the model’s predictions. A_pulse_iqr refers to the IQR of the total area under each pulse (Figure 4). In the frequency domain, the selected features were the relative powers of the high-frequency bands in both the Welch power spectral density (PSD; Multimedia Appendix 1, F32-F44) and autoregressive PSD (Multimedia Appendix 1, F45-F57).

In the nonlinear domain, the selected features were either related to the energy or the complexity of the signal. LOG_std refers to the SD of log-energy entropy (equation 3), whereas KTE_skew refers to the skewness of the KTE energy measure for each sliding window (equation 2). Furthermore, the complexity feature that was selected was the sum of the MSE over 2 scales, 13 and 14.

Finally, the remaining selected features were demographic features that described the age and BMI of the participants, as well as if they had any family history of diabetes.

Seven widely used machine learning (ML) algorithms, including the naive Bayes classifier, K-nearest neighbors algorithm, logistic regression, random forest, SVM, XGB, and light gradient boosting machine, were trained with the selected features as inputs. We fine-tuned the hyperparameters of each model and validated their performance using the stratified 10-fold cross-validation method. We adopted multiple regularization techniques across various models to prevent overfitting during the model training. Six evaluation metrics, accuracy, sensitivity, specificity, precision, geometric mean (G-mean), and F score, were used to evaluate the model’s performance, as accuracy alone cannot provide a comprehensive examination of model performance due to data imbalance. The G-mean and F score are critical evaluation criteria to assess the models’ performance, as they are robust to significant label imbalance.

The prediction results from each model are reported as the mean and SD of the evaluation metrics, and Table 4 shows the summary of the results. SVM with the radial basis function kernel showed the best prediction performance with an average accuracy of 84.7%, a sensitivity of 81.05%, a specificity of 88.35%, and a precision of 87.51%. In particular, the average G-mean was 84.54% and F score was 84.03%.

The use of deep learning in the medical and health care domain has shown great potential for solving a range of problems, such as detecting specific symptoms or abnormalities [64,65]. However, the interpretability of deep learning models remains a significant challenge, and it is often difficult for clinicians to trust the decisions made using a black-box system. The lack of model interpretability also raises ethical concerns, particularly when the decision fails. Furthermore, our current data set is considerably small (500 participants) compared with typical deep learning models in other domains, which are trained with thousands of data points. Deep learning models are known to perform well with a larger data set and fail to learn meaningful representations when there is a lack of data [66]. Therefore, we did not investigate the use of deep learning in this study.

As the proposed ML model is designed to complement the existing diabetes detection solution and is relatively new to the clinical community, the features selected in the previous section must be interpretable and exhibit a certain level of agreement with existing findings. A family history of diabetes, being male, being aged ≥45 years, and having an increased BMI have been identified as major risk factors in the literature for developing prediabetes or T2DM [63,67,68]. These 4 risk factors were part of the selected predictors, and this paper provides a preliminary attempt to explain how the selected predictors contribute to detecting elevated blood glucose using the Shapley additive explanations (SHAP) framework. SHAP is a game theoretical approach that provides global and local explanations of the association between the ML output and input features [69].

Figure 5A illustrates the SHAP values of each feature across all the predictions from the training set. The features were ranked by their mean SHAP values, with larger values shown in red and smaller values shown in blue. The beeswarm plot revealed that a family history of diabetes, increasing age, and higher BMI are associated with a higher probability of elevated blood glucose levels. These observations are consistent with previous research and demonstrate that the ML algorithm has successfully captured the relationship between these features and elevated blood glucose levels. In addition, other proposed features showed varying levels of impact on the model’s output. However, the gender feature did not have any apparent effect on the model’s predictions.

In Figure 5B, each row in the plot shows how the contributions of different features move the output of the model from the expected value (E[f(x)]) to the actual prediction output f(x) for a single sample with a positive class prediction (blood glucose level ≥7.8 mmol/L) in the test set. The expected value, E[f(x)], is determined using the entire training data set. As expected, most features provide positive SHAP values in this sample, which collectively push the model’s output toward the correct prediction. However, this specific test participant’s BMI was in the healthy range, which pushed the model’s output toward the normal class and might have resulted in a false negative prediction. This indicates that relying on a single feature or demographic data alone may not provide an accurate prediction of blood glucose levels.

Using the SHAP values, we can understand the model’s overall behaviors and how features affect the output positively or negatively, which can help improve the prediction model in the future.

Generally, diagnostic tests are not highly sensitive and highly specific. Therefore, repeated measurements of the wrist-worn wearable device were combined and assessed in an optimum fashion to maximize sensitivity, specificity, and precision.

Consecutive measures of blood glucose were combined in parallel using the “AND” and “OR” rules to assist in the detection of elevated blood glucose measurement levels. The “OR” rule increases the overall sensitivity, and the “AND” rule increases the overall specificity, which is greater than that of either test alone [70].

While the health care landscape is changing, the rapidly aging society and the need for improved population health outcomes call for new models of care to effectively prevent the onset and delay the progression of chronic diseases. Furthermore, short-term health behaviors contribute significantly toward long-term health outcomes, while unattended and frequent glucose spikes might result in prediabetes and eventually diabetes. The availability of noninvasive and device-agnostic blood glucose detection solutions will allow for more frequent and better monitoring of blood glucose levels, thereby reducing the risk of developing T2DM. This study demonstrates that a noninvasive method of assessing diabetes risk using PPG is a viable option to provide a cheaper and accessible modality for the population-wide screening of blood glucose levels. This population-based screening would allow for the earlier detection of DM in the population, especially among those individuals who are unaware of their elevated blood glucose levels. Hence, timely and appropriate lifestyle advice and medical interventions can be provided to prevent diabetes complications. This will subsequently reduce the health care burden for both the individual and the society.

BGEM is a cloud-based solution that can frequently monitor multiple digital biomarkers with minimal disruption to daily life. Developed using the advanced ML operations practice, BGEM can be easily scaled to meet the increasing demand for health care services. The solution includes a user-friendly mobile app that can screen a large population to identify high-risk individuals, people with undiagnosed diabetes, and those who require primary prevention intervention. It also provides timely feedback to users through the app, informing them of their diabetes risk and providing targeted, actionable insights to empower them to take a proactive approach to monitor their glucose levels.

Our pilot study has certain limitations. Since fasting blood glucose measurements were excluded and the criteria to define normal and abnormal levels under fasting conditions differed from our current cut-off, we must refrain from definitively concluding that our model is applicable to fasting conditions. Regarding gender, our feature-selection model did not specifically incorporate it, and our analysis using SHAP demonstrated that gender exerted minimal influence on model predictions. Moreover, all analyses were adjusted for the covariate gender, as required. Therefore, we considered gender to have a limited impact and is not a primary limitation of our findings. To address these limitations, we are actively planning the subsequent phase of data collection. This phase will involve collecting fasting blood glucose measurements in a primary care setting, also allowing for a more balanced gender distribution. More importantly, we could expand our participant pool to encompass participants with prediabetes and diabetes. By addressing these gaps, we aimed to offer a more comprehensive and robust assessment of our model’s applicability and effectiveness.

There was no longitudinal follow-up of the participants. External validation of our model on an independent sample must be undertaken to further assess the detection accuracy and generalizability of the results. Nevertheless, as a preliminary investigation, the potential implications of our findings are significant as they might offer a means to identify previously undiagnosed prediabetes or diabetes cases at the population level. We anticipate that our study will serve as a foundational stepping stone, paving the way for more comprehensive diabetes research using AI and wearable devices. To the best of our knowledge, there is no publicly available data set that systematically examines the relationship between PPG data and blood glucose levels. Acquiring a substantial volume of data is imperative, encompassing a diverse and representative sample spanning the entire spectrum of glucose values and incorporating relevant sociodemographic factors. Such comprehensive data can be obtained through a collaborative effort involving research institutions and industry partners while ensuring strict adherence to local ethical considerations and data privacy regulations.

We demonstrated that the cloud-based ML model can detect elevated blood glucose levels, where consecutive measurements can be combined in an optimal manner to provide high sensitivity, specificity, and precision. However, further research is required to address these limitations.

In this study, we performed sophisticated feature engineering and found that the features derived from the MSE analysis of PPG signals effectively detect blood glucose changes. We will discuss this set of novel features in detail in a separate paper. To reduce bias and evaluate the generalizability of the model, we used a 10-fold cross-validation to assess its performance. The SVM with the radial basis function model performed the best, with an average accuracy of 84.7%, a G-mean of 84.54%, and an F score of 84.03%. Previous models were developed using smaller samples and have lower model performance measures. Our model was developed with a larger sample of 500 participants, and most participants were assessed before and after the consumption of a sugary drink. It also achieved better detection accuracy.