This study evaluated whether PPGRs could be matched and clustered using the CV of the PPGR. Data were from the OhioT1DM dataset [23], comprising 8 weeks of CGM and self-reported lifestyle logs from 12 adults with T1D using insulin pump therapy and Medtronic Enlite CGM systems. CGM readings were taken every 5 minutes, alongside insulin dosages, heart rate, and timestamped dietary and lifestyle events. Data were aligned to a 24-hour clock and stratified by day and week. Despite its modest size [33], the dataset reflects real-world conditions and is widely used in T1D research [34]. Participant demographics are shown in Table 1.

This study is an in silico analysis that relies exclusively on open-access, publicly available data obtained from recognized repositories. No new data were collected, and no interaction or intervention with human participants, animals, or identifiable private information occurred. As such, this research does not meet the criteria for “human subjects research” as defined by institutional and international ethical standards. Ethics approval was not required for this study as it utilized open-access, deidentified data available in the public domain. The analysis involved no direct or indirect contact with human participants or animals, and no identifiable personal data were used [35-39].

A PPGR was defined as a 4-hour window (48 readings) from self-reported mealtime or bolus insulin administration (see Figure 1). When mealtimes were missing, start times were imputed from matching events on the same weekday in other weeks, using the closest average timing. The 4-hour window was chosen to capture the full glycemic excursion, given that glucose in nondiabetics typically normalizes within 90 minutes [40], gastric emptying completes within 4‐5 hours [41], and rapid-acting insulin (eg, Novorapid) peaks at 100‐120 minutes when given 15‐20 minutes before meals [42]. Events with <2 hours of CGM data (24 points) were excluded.

PPGRs were categorized as breakfast (06:00-10:00), lunch (10:00-14:00), snack (14:00-18:00), or dinner (18:00-22:00) in line with circadian and sociocultural eating patterns [41,43,44]. Additional carbohydrate intake within the 4-hour postprandial window was considered part of the original PPGR. A bolus is classified as meal-related if its start time falls within a tight window around a logged meal, specifically from 4 minutes before to 4 hours after the meal; boluses outside every such window are labeled corrections. Correction boluses are then summarized by the PPGR categorization bands (6-10, 10-14, 14-18, 18-22). Where bands did not have an isolated PPGR from a self-reported meal of mealtime bolus, the PPGR was isolated, and where the bands contained meal events, the bolus dose was confirmed to be a correction.

To enable cross-comparison of glucose responses independent of absolute magnitude, each PPGR was normalized by centering it around its peak value. Specifically, the peak glucose reading within each 4-hour window was identified, and all other values were recalculated by subtracting this peak value. This process generated a normalized, peak-centered time series for each event, allowing for uniform trend analysis across individuals and events.

The CV quantified variability by dividing the standard deviation of each isolated PPGR by the mean of all PPGRs for the same individual and meal category, then multiplying by 100 (Equation 1). This approach provided a normalized measure of glycemic variability contextualized within each person’s own physiological response patterns. A CV threshold of less than 36% was adopted as the criterion for acceptable variability, consistent with clinical literature associating lower CV values with reduced risk of hypoglycemia [25,45]. Given that one would expect the PPGR-specific CV to be significantly higher than the 24-hour CV, using the 24-hour threshold of 36% would ensure the internal cohesion of the PPGRs is tighter than the clinically recommended glycemic variability.

Each PPGR was treated as a unique time series and clustered within the specified meal categories. An iterative pairwise algorithm began with an unassigned PPGR and added other unassigned events if the cluster CV remained <36%. Clusters were expanded until no additional events could be included without exceeding the threshold. Unassigned events were reevaluated in later iterations; PPGRs eligible for multiple clusters were assigned to the largest cluster. Events failing all inclusion criteria were labeled outliers (Figure 2).

Each cluster was evaluated using several statistical measures to assess internal consistency and clinical relevance. The standard deviation of glucose values within each cluster was used to estimate dispersion around the cluster mean. The standard error was calculated by dividing the standard deviation by the square root of the number of PPGR events in the cluster, providing a measure of the reliability of the cluster mean. An internal CV for each cluster was also computed, based on the combined glucose values of the cluster and the overall mean for the corresponding meal category. The total number of PPGRs per cluster was recorded to assess distribution across clusters and ensure adequate representation.

Sample size and power calculations used the TTestIndPower function from statsmodels in Python [46], targeting power=0.8 and α=.05 to minimize type I or II errors [47]. Within each meal category, one-way ANOVA tested for differences in mean carbohydrate intake between clusters.

Data preprocessing, normalization, and clustering were implemented in Python 3.10.7 [48] using Visual Studio Code 1.77.3 [49] and standard libraries. The 36% CV threshold, though derived from 24-hour glucose variability, was used to ensure PPGR clusters had tighter internal cohesion than recommended clinical variability limits.

In this study, data from 12 people with T1D were analyzed. Each participant contributed 9900 BG measurements, which were used to generate and cluster over 200 isolated PPGRs per person. CV-based clustering identified stable, participant-specific PPGR patterns, averaging 2.4 for breakfast, 2.7 for lunch, and 3.1 for dinner with most cluster errors below the CGM’s intrinsic error. Carbohydrate intake was largely consistent within clusters, with significant variation in only 2 of 36 meal-participant combinations. The large effect size (Cohen d=1.20) underscores the robustness of this approach. These findings demonstrate the intraindividual reproducibility of the PPGRs within a meal category under free-living conditions in people with T1D, highlighting the possibility to predict the PPGR to a given meal in an individual, a concept already validated in healthy individuals [15] and people living with type 2 diabetes [5]. Using PPGR clusters to highlight how an individual responds to a specific meal type provides insight into the overall meal metabolism without requiring onerous additional information from the people with T1D. This could outline appropriate insulin doses for each PPGR cluster that would optimize TIR. Enhancing the PPGR clustering framework presented here without increasing the burden of disease management, through the utilization of meal tags associated with each PPGR event, could ease the burden of nutrition reporting.

Using the CV as a clustering metric for PPGRs is clinically advantageous because a PPGR-specific CV would be expected to be significantly higher than a 24-hour CV, meaning that applying the 24-hour clinical threshold of 36% ensures tighter internal cohesion than the recommended glycemic variability limit. In a comparative analysis, k-means clustering with dynamic time warping produced moderate cohesion, with 50% (6/12) of participants forming clusters across at least 1 time-of-day period when set to 3 clusters (59% of these meeting the CV threshold) and 58% (7/12) when set to 4 clusters (61% meeting the threshold). By contrast, our proposed method achieves markedly superior performance: in all clusters containing more than 1 PPGR, 100% meet the clinical CV threshold, and cluster formation covers all meal categories for every participant, ensuring both clinical relevance and complete coverage.

In the dataset observed within this work, carbohydrate intake alone does not typically predict which cluster a corresponding PPGR would align to, highlighting the importance of overall meal composition and the physiological factors of the individual in determining the PPGR outcome. In those meals where the average PPGR clusters do not meet the target CV, there is a large discrepancy in the number of PPGRs isolated to each cluster, with many of the clusters resulting from outliers that skew the average CV (as seen in Table S1 in Multimedia Appendix 1). Outliers are to be expected given the real-world setting of the dataset.

The limitations of using carbohydrate alone to predict the PPGR are evident and reflect the known limitations of carbohydrate counting, with studies showing a 20% mean error in carbohydrate estimates in people with T1D [53]. The complexity of carbohydrate-based insulin dosing is compounded by the varying rates of glucose absorption from different types of carbohydrates or meals with multiple components, making them nutritionally complex [53,54]. The absorption of digestible carbohydrates in comparison to simple glucose, known as the glycemic index (GI), is well researched [4,20,54]; however, when used in the real world, with complex mixed meals, the prediction of the PPGR using solely GI has shown low precision [13,55]. The unpredictable interaction of several factors, dietary and otherwise [56-58], contributes to the complexity and interpersonal nature of the management of the PPGR [59].

The number of clusters formed per meal type varies with time of day, highlighting the significant impact of meal timing on the PPGR. The PPGR response to meals with the same nutritional breakdown, consumed at different times of day, is expected to be greater during the night versus during the day [60]. Current treatment guidelines considering insulin-to-carbohydrate ratios, which fluctuate throughout the day, support the argument that carbohydrate content alone should not dictate the PPGR [61,62]. This enhances the argument reflected in this work that carbohydrate content of a meal alone should not be the only determining factor of the PPGR [59,63]. The work here suggests that the PPGR is decided by a combination of the complete nutritional intake and the individual’s physiology, which is supported by the intraindividual PPGR reproducibility to identical meals shown in 800 people [15]. Overall, this highlights the need for the development of a personalized prediction model based on unique PPGR events [58].

Meal timing, along with nutritional composition, has a marked effect on the PPGR [17,24,63], resulting in meals of identical nutritional profiles eliciting different PPGRs when consumed at different times of day (ie, breakfast vs dinner). The treatment relevance of the reproducibility of a PPGR to similar meals at a similar time of day is exacerbated by the fact that habit is a proven driver of food consumption [64,65], both in the type of food and in portion size [66]. To enhance this identified habitual human nature, people with T1D are encouraged to “establish a routine” in the early stages after diagnosis [40]. These factors that enhance the repetitive nature of PPGRs offer the possibility to streamline treatment plans through clustering the PPGRs based on pattern recognition. The traditional treatment strategies of carbohydrate-based insulin dosing have limitations, especially for the case of high fat, high protein foods, which are consistently problematic to manage, along with breakfast cereals possibly due to the high fiber and sugar content [67]. It has been indicated that a person with T1D will use previous experience as a determinant for the bolus dose needed to cover a meal [68]. The optimal indicators of the PPGR in a person with T1D are the carbohydrate and fat content of a meal [20,69]. However, reporting multiple macronutrients compounds the burden on people with T1D, increasing the likelihood of reporting errors. Due to the complexity of PPGR prediction, along with the limitations of carbohydrate counting as a predictor alone, it has been reported that the bolus insulin dose needed to cover the meal consumed was inaccurately estimated in 64% of cases [70]. Therefore, assigning PPGRs to clusters based on an individual’s unique response to a meal would simplify treatment planning and reduce the reliance on determining meal carbohydrate content. The presented PPGR-focused clustering algorithm offers an alternative to traditional evaluation methods, such as area under the curve based on GI, which fail to capture individual physiological responses to meals [59].

The clear establishment of PPGR clusters reflects the recurrent behaviors of people with T1D, potentially offering clinical insights for insulin dose decisions and treatment adjustments to optimize TIR. Utilizing PPGR clusters to predict BG values from the PPGR of a previously consumed meal could offer a reduction in patient anxiety and diabetes burden, due to the foreknowledge of upcoming glycemic fluctuations, ultimately easing the management of T1D. Although this work investigates real-world scenarios, the outcomes are supported by the discernible correlation between PPGRs shown to standardized meals [71]. The assessment of CGM data for patterns shows that interday similarities tend to be higher in older people with T1D [18], which reflects the population in the Ohio dataset used here, whose average age is in the range of 30.9‐56.4 years. The presence of repeatable PPGRs in the real world, as seen in this work, may be attributed to factors such as insulin utilization and the habitual nature of dietary intake and lifestyles. This habitual nature of people’s eating patterns, along with the encouraged ritualistic nature of disease management in T1D, will allow the use of PPGR clusters to enhance a feedback loop to better predict BG levels. The work presented here provides an initial framework for prediction that is independent of the initial BG level and not solely reliant on identifying the amount of carbohydrates consumed, suggesting a more physiological-based glycemic response to food. A bolus targeting solution algorithm based on the PPGR-focused clustering algorithm presented here would allow for an insulin adjustment plan to optimize time in range for each isolated PPGR cluster.

This study presents a PPGR clustering method that incorporates the full 4-hour BG profile, enabling reduced glycemic variability and contextual insights into individual responses to specific meal types. While promising, limitations include reliance on self-reported meals including carbohydrate intake, which is prone to consistent under- or overreporting within individuals [72], and omission of insulin dose, insulin-on-board, and carbohydrates-on-board, all of which influence PPGR. The dataset was derived from multiple clinical studies with nonstandardized insulin regimens, and insulin sensitivity was treated as an independent variable. Future work will address these factors by incorporating dose, timing, and administration data to improve modeling accuracy.

Even with reproducible PPGR clusters, translating findings into insulin dosing strategies remains complex due to the highly individualized nature of bolus timing and dose optimization [3,73]. In both research datasets (eg, OhioT1DM) and real-world contexts, inconsistent meal logging presents challenges. To mitigate this, we propose a meal-tagging mechanism to link PPGRs via pattern recognition rather than precise carbohydrate counts (Lubasinski et al, unpublished data, 2025) [68], enabling clusters to form based on similar nutritional profiles. This will reduce reporting burden, align with natural logging habits, and provide proactive predictions for upcoming meals, including tailored bolus recommendations aimed at optimizing TIR while lowering cognitive load.

Limitations also include a small sample size, reducing statistical power and generalizability. This work is exploratory and should be interpreted with caution due to potential sampling bias. Broader validation is essential, and our methodology has already been applied to a larger public dataset [74] and replicated in follow-up work (Lubasinski et al, unpublished data, 2025). Future multicenter studies with more diverse populations are recommended to strengthen statistical power, enable subgroup analysis, and support the development of scalable, individualized glycemic prediction and adaptive insulin dosing models.

The clinically relevant and safe PPGR clustering technique developed in this study categorizes PPGRs into distinct PPGR clusters based on the participants’ unique response to a meal. This offers a framework for a more personalized and adaptable way to manage PPGRs, independent of initial BG levels or carbohydrate consumption. The reproducibility of PPGRs offers clinical significance through insights for insulin dosing decisions and treatment adjustments to optimize TIR, which may indicate priming insulin doses to flatten the PPGR spike. This will alleviate patient anxiety by providing advance knowledge of glycemic fluctuations and by learning from the past behaviors and outcomes within each cluster. Notably, this work challenges the conventional notion that carbohydrate intake alone determines PPGRs. In summary, we found that PPGRs are reproducible, resulting from a complex interplay of an individual’s unique physiology and their overall nutritional intake, underscoring the need for personalized prediction models.