This is a substudy anchored within the “Assessing the Progression to Type-2 Diabetes” (APT-2D) study (ClinicalTrials.gov: NCT02838693), which follows up on a large cohort of nondiabetic individuals for 3 years or until they develop T2DM [20].

For the main APT-2D study, healthy participants between the ages 30 and 70 years old, who were not on any long-term medication and had no prior history of diabetes, were recruited from April 2016 to December 2018. Recruitment was done through random sampling via various outreach efforts to the grassroots communities and organizations, and through media releases in order to recruit at least 2300 participants, so as to ensure adequate sample size for the cohort study outcomes.

For this substudy, participants with their penultimate visit (75 g OGTT) being conducted within 3 months from enrolling to this substudy (“Continuous Glucose Monitoring to Assess Glucose Dysregulation in Progression to Type-2 Diabetes” [CGM-APT2D]) and willing to wear a CGM sensor for 14 days were invited to participate. Out of the 449 individuals enrolled in this substudy, 429 participants contributed with CGM data. To narrow the study’s focus to the Asian population in this region, individuals from Europe or the Middle East were excluded from the analysis. Additionally, those diagnosed with T2DM during the study were removed to align with the study’s objective of examining individuals without diabetes. At the end, data from 419 individuals were used.

The main study (APT-2D) was approved by the Domain Specific Review Board of the National Health Group (ref: 2016/00096) in Singapore. This substudy (CGM-APT2D) was approved by the Domain Specific Review Board of the National Healthcare Group (ref: 2020/01085) in Singapore. All participants provided written informed consent prior to joining the studies.

HbA_1c was measured in whole blood by cation-exchange high performance liquid chromatography (Bio-Rad Variant II Turbo; Bio-Rad). Plasma glucose during the OGTT was determined by the AU5822 general chemistry analyzer (Beckman Coulter) at the National University Hospital Referral Laboratory (accredited by the College of American Pathologists).

For each participant, the data encompassed 3 main components: demographic information, including age, sex, and ethnicity; clinical measurements recorded during the clinic visit, consisting of BMI, waist-hip ratio, fasting and 2-hour OGTT glucose (mmol/L), and HbA_1c (%); and glucose profile data for about 14 days obtained from the worn CGM sensor. The CGM data comprised a time series capturing the time and corresponding glucose levels (mmol/L), spaced at approximately 15-minute intervals. These intervals are not strictly equal, and the recording length and timing varied among individuals. Participants scanned the CGM sensor at an average frequency of 9.1 times a day. There were additional heterogeneities in the CGM datasets. For instance, occasional involuntary detachment of the sensor resulted in missing data, while the malfunctioning of sensors in some participants resulted in them wearing more than 1 sensor during the study period. Of the 429 participants, 8% required 1 replacement sensor and 0.5% required 2 replacement sensors.

For ease of comparison, we followed the work by Mao et al [13] and converted the CGM measurements into meaningful features that capture the essential characteristics of an individual’s glucose trajectory. These features encompass measures of centrality (mean) and spread (maximum, minimum, standard deviation, coefficient of variation, and mean amplitude of glycemic excursion), along with proportions of time spent within abnormal ( >7.8 or <3 mmol/L) and normal (3 to 7.8 mmol/L) glucose ranges. We additionally considered measures of glucose excursion, including the average rise and fall and their corresponding rates. Further details about the various CGM features, in terms of how they were defined and derived, are provided in Multimedia Appendix 1.

These summary statistics were computed after excluding the first 24 hours in each consecutive block of CGM recordings, as recommended by the manufacturer because glucose readings on the first day were typically lower compared to subsequent days. In cases where participants wore more than 1 device, we merged CGM recordings from different blocks after eliminating data from the “warm-up” period, given the negligible time gaps between them. During this process, 11 participants were subsequently removed due to insufficient CGM data captured by their devices, preventing the calculation of meaningful CGM summary statistics. Two participants were removed due to invalid HbA_1c values. Consequently, the final dataset comprises information from 406 participants with an average CGM recording length of 12 days.

We categorized these 406 participants based on their diabetic status into those with prediabetes and those with normoglycemia and prediabetes, thereby enabling a comparative analysis between groups. Specifically, for categorical variables such as sex and ethnicity, we used the Pearson chi-square test, while for continuous variables, we explored potential differences in group means through nonparametric Mann-Whitney U tests. P values were computed and a significance level of .05 was applied for interpretation.

Prediction models were then constructed to assess individual prediabetes risk. We explored three distinct sets of predictors in our analysis (Table S1 in Multimedia Appendix 1):

The performances of these models were compared with HbA_1c ≥5.7% alone to classify prediabetes as the benchmark for comparison.

For each of these data configurations, we employed 2 classification algorithms: logistic regression (LR) and support vector machine (SVM). To address potential concerns of overfitting, we implemented repeated random subsampling validation strategies, and evaluated the average prediction performance across 1000 randomly selected test sets, utilizing key metrics such as misclassification rates, specificity, and sensitivity.

In addition to our primary prediction models, we introduced a specialized 2-step prediction strategy. This approach categorized individuals with HbA_1c ≥5.7% as having prediabetes and focused exclusively on predicting the prediabetes risk among individuals with HbA_1c <5.7%. The predictor sets employed for this targeted analysis included the CGM dataset (set 2) and the HbA_1c dataset (set 3). We explored whether this segregation would enhance risk prediction process, allowing for a more nuanced understanding of prediabetes risk among individuals with relatively low HbA_1c levels. A sensitivity analysis was additionally conducted regarding the segregation threshold of 5.7%, but the alternative thresholds failed to yield better prediction accuracy in terms of misclassification rates (Tables S2–S3 in Multimedia Appendix 1). An overview of the methods is summarized in Figure 1.

Demographic information and clinical glucose measurements of the 406 Asian participants without diabetes are summarized in Table 1. Within this cohort, 189 (46.6%) individuals were categorized as prediabetic. The majority of the cohort were women (n=236, 58.1%) and individuals of Chinese ethnicity (n=267, 65.8%), with no significant differences observed in terms of sex and ethnic group distribution between the normoglycemia and prediabetes groups. The age of participants spanned from 33 to 74 years, with half falling between 42 and 57 years old. Those in the prediabetes group were ~6 years older and had a slightly greater BMI and waist-to-hip ratio than those in the normoglycemia group; the prevalence of obesity was 32.3% and 19.8% respectively, under the Singapore classification system [21]. Fasting and 2-hour OGTT glucose levels and HbA_1c were significantly greater in prediabetic than normoglycemic individuals (Table 1 and Figure S1 in Multimedia Appendix 1).

With respect to glucose features derived from the CGM data, the prediabetes group displayed larger variability in glucose levels over time compared to the normoglycemia group, substantiated by significant differences in group means for most measurements related to spread and glucose excursion; only the minimum glucose was not significantly different between the 2 groups (Figure 2 and Table S4 in Multimedia Appendix 1). Notably, our assessment of time-in-range focused on the interval of 3.0‐7.8 mmol/L. This choice stemmed from the consideration that the conventional cut-off for time-in-range (3.9‐10 mmol/L) did not appear pertinent in this nondiabetic population, given the minimal divergence in the proportions of time spent outside this interval between the 2 groups (Table S5 in Multimedia Appendix 1).

We set the benchmark for comparison using the prediabetes cut-off defined by HbA_1c ≥5.7%, resulting in a specificity of 100%, sensitivity of 60%, and a misclassification rate of 18%. The Demo model, which incorporated basic demographic information and obesity-related measures (age, gender, BMI, and waist-to-hip ratio) that are known risk factors for T2DM (Table S1 in Multimedia Appendix 1), demonstrated limited efficacy in distinguishing prediabetes from normoglycemia (Figure 3 and Table 2). The predictive performance greatly improved with the inclusion of CGM features in the model (CGM model), with a prediction sensitivity of 60%‐63% becoming comparable to that using HbA_1c ≥5.7% (Table 2). The diagnostic capability further increased when the HbA_1c value was added as a predictor to the CGM model (HbA_1c model), yielding a higher specificity of 78%‐80% and a sensitivity of 71%‐74%, and hence a substantially lower misclassification rate of 23%‐25% (Table 2). Notably, the sensitivity achieved by the HbA_1c prediction model surpassed that obtained when using HbA_1c ≥5.7% as a single threshold of 60%. In addition, the choice of classification algorithms did not fundamentally impact the predictive performance for all 3 models (Figure 3 and Table 2).

We extended our analysis to assess whether stratifying the population based on their HbA_1c levels could improve the detection of prediabetes using a 2-step approach where those with HbA_1c ≥5.7% were automatically categorized as having prediabetes, thereby focusing the prediction capability on those with HbA_1c <5.7%. Overall, this 2-step approach outperformed the benchmark for comparison using the prediabetes cut-off defined by HbA_1c ≥5.7%, and significantly improved the predictive performance in the HbA_1c model of both algorithms where LR_{ROC AUC} increased from 0.838 to 0.866 and SVM_{ROC AUC} increased from 0.829 to 0.876 (Tables 2 and 3 and Figure 4). In the LR HbA_1c model, there was an increase in sensitivity from 73.6% to 75.7%, resulting in a net reduction of the misclassification rate from 23.2% to 22.3%. A similar trend was also observed when SVM was used (Tables 2 and 3). Interestingly, the inclusion of the HbA_1c value as a variable in the 2-step approach did not improve the predictive performance in both the LR (area under the receiver operating characteristic curve [ROC AUC] of CGM: 0.872 vs ROC AUC of HbA_1c: 0.866) and SVM (ROC AUC of CGM: 0.881 vs ROC AUC of HbA_1c: 0.876) models (Table 3). Assessment of predictive performance using the area under the precision recall curve yielded a similar outcome (Table 3). Nonlinear models such as random forest and Extreme Gradient Boosting were also explored (Table S6 in Multimedia Appendix 1) but were not superior in predictive performance compared to the LR and SVM models in the 2-step approach (Table S7 in Multimedia Appendix 1 vs Table 3). Thus, LR and SVM were chosen for their ease of interpretability and clinical usability.

To further assess the feasibility of using CGM data to accurately identify individuals with prediabetes whose HbA_1c were in the normoglycemic range, we next focused on individuals with HbA_1c levels <5.7% (n=292) and compared the prediction outcomes for this subpopulation. We set the benchmark for comparison using the prediabetes cut-off defined by HbA_1c ≥5.7%, resulting in a specificity of 100%, sensitivity of 0%, and a misclassification rate of 25.7%. The ROC AUC and area under the precision recall curve values of the models were significantly higher than 0.5 and the test results yielded a slightly higher misclassification rate of 30%‐32%, a lower specificity of 73%‐80%, but a much higher sensitivity rate of 40%‐54% compared to the benchmark. SVM was slightly superior to LR in overall predictive performance and the omission of HbA_1c value as a variable resulted in a better model performance (Table 4).