Continuous glucose monitoring for automatic real-time assessment of eating events and nutrition: a scoping review

Background Accurate dietary assessment remains a challenge, particularly in free-living settings. Continuous glucose monitoring (CGM) shows promise in optimizing the assessment and monitoring of ingestive activity (IA, i.e., consumption of calorie-containing foods/beverages), and it might enable administering dietary Just-In-Time Adaptive Interventions (JITAIs). Objective In a scoping review, we aimed to answer the following questions: (1) Which CGM approaches to automatically detect IA in (near-)real-time have been investigated? (2) How accurate are these approaches? (3) Can they be used in the context of JITAIs? Methods We systematically searched four databases until October 2023 and included publications in English or German that used CGM-based approaches for human (all ages) IA detection. Eligible publications included a ground-truth method as a comparator. We synthesized the evidence qualitatively and critically appraised publication quality. Results Of 1,561 potentially relevant publications identified, 19 publications (17 studies, total N = 311; for 2 studies, 2 publications each were relevant) were included. Most publications included individuals with diabetes, often using meal announcements and/or insulin boluses accompanying meals. Inpatient and free-living settings were used. CGM-only approaches and CGM combined with additional inputs were deployed. A broad range of algorithms was tested. Performance varied among the reviewed methods, ranging from unsatisfactory to excellent (e.g., 21% vs. 100% sensitivity). Detection times ranged from 9.0 to 45.0 min. Conclusion Several CGM-based approaches are promising for automatically detecting IA. However, response times need to be faster to enable JITAIs aimed at impacting acute IA. Methodological issues and overall heterogeneity among articles prevent recommending one single approach; specific cases will dictate the most suitable approach.


Introduction
Nutrition has a major impact on people's health and well-being (1)(2)(3)(4)(5)(6)(7)(8).However, accurately assessing nutrition and dietary intake remains challenging, with the most precise tools often involving high costs, participant and staff burden, or privacy issues (9)(10)(11)(12)(13)(14).Yet, valid and reliable measurement of dietary behavior is essential to accurately detect changes in research settings and guide patient counseling in clinical practice (e.g., weight loss programs).Technological advances in recent years have led to new approaches for accurately assessing dietary intake that try to overcome some of the shortcomings of traditional dietary assessment methods (9,(15)(16)(17)(18).
An attractive technology-based option for assessing the consumption of calorie-containing foods and beverages (ingestive activity, IA) is continuous glucose monitoring (CGM).CGM involves using a sensor that measures glucose concentrations in the interstitial fluid (19,20) as a proxy for blood glucose levels (20,21).CGM has become an important tool in diabetes care (19,(22)(23)(24)(25).For instance, it is an integral component of artificial pancreas (AP) systems designed to automate and improve blood glucose regulation in individuals with type 1 diabetes mellitus (T1DM) via the utilization of CGM, an insulin infusion pump, and a control algorithm (26).Beyond diabetes management, CGM is gaining popularity for use in healthy individuals and athletes (20,27).Several CGM devices show satisfactory accuracy data (20,28,29).
The automatic and (near-)real-time detection of IA via CGM could offer benefits in (clinical) practice, including a reduced participant and staff burden.In addition, interventionists could monitor meal plan adherence more closely and detect deviations from intervention goals as they occur.Consequently, targeted and personalized countermeasures could be deployed proactively.One particularly useful approach would be CGM-based detection of IA in the context of Just-In-Time Adaptive Interventions (JITAIs).JITAIs aim to exploit the full potential of remote monitoring combined with delivering intervention content in the moment/context when it is most needed and the patient is likely to be (most) receptive (30).Preliminary results show promising effects of JITAIs on predicting and preventing dietary lapses (31).If detection times of the CGM-based approaches in question were extremely short (e.g., less than a few minutes), JITAIs could aim at acutely impacting IA (e.g., sending a prompt asking a person to terminate a meal).If detection times were relatively short (e.g., less than an hour), JITAIs could aim at altering subsequent IA (e.g., a dinner meal).In both cases, information on IA would be much more readily available than with traditional dietary assessment methods (e.g., 24-h recalls).
However, there may also be challenges associated with the use of CGM for the automatic monitoring of IA.On the one hand, there are system-inherent challenges.For example, postprandial rises in blood glucose vary in timing and extent depending on meal composition, meal quantity, inter-individual variability, and many other factors (32-40).Further, there is a delay between interstitial fluid and blood glucose concentrations (20,41,42).On the other hand, blood glucose levels are not only influenced by IA but also by other factors such as physical activity, stress, and diurnal fluctuations (20,41,(43)(44)(45)(46)(47)(48)(49)(50)(51)(52)(53)(54)(55).Thus, false positive detections (e.g., erroneously flagging a meal due to glucose increases caused by stress) and false negative detections (e.g., erroneously not flagging a meal because other factors render the glucose response too flat) might occur.
In the past years, research examining the use of CGM for the automatic detection of IA has accumulated.Recent publications reviewed options for the automatic detection of IA using wearable−/ sensor-based methods (15-18), but they did not specifically address CGM.The present article aims to close this research gap and answer the following guiding questions: 1 Which approaches using CGM for the automatic detection of IA in (near-)real-time have been investigated, and have these approaches relied solely on CGM or also used other data (e.g., sensors/wearables)? 2 How accurate are these approaches in detecting IA? 3 Can these approaches be used in the context of JITAIs?

Methods
The reporting of this review is based on the updated Preferred Reporting Items for Systematic Reviews and Meta-Analyses Extension for Scoping Reviews (PRISMA-ScR) guideline (56).

Search strategy
The primary systematic literature search was conducted on 09 September 2022, using the IEEE Xplore, PubMed, Scopus, and Web of Science databases.An identical supplementary search was conducted on 02 October 2023.The search term was developed and refined by two authors (JB, CH) to capture all relevant publications, and the search term contained: (intake OR uptake OR eating OR ingest* OR meal OR drink* OR beverage OR consum* OR oral) AND (monitor* OR assess* OR detect* OR estimat* OR measur* OR sens*) AND ("continuous glucose monitoring" OR "real time continuous glucose monitoring" OR "real-time continuous glucose monitoring" OR "flash glucose monitoring" OR "intermittently scanned continuous glucose monitoring" OR CGM OR rtCGM OR isCGM OR "artificial pancreas" OR "artificial beta cell*" OR "artificial beta-cell*" OR "artificial β-cell*" OR "artificial β cell*") AND (algorithm OR "deep learning" OR "machine learning" OR "neural network*" OR AI OR "artificial intelligence").Because a fully-closed-loop AP system must first detect meals to adequately manage the following increases in glucose by delivering insulin to the patient (57), the search also included AP systems.
JB conducted the database searches and removed duplicates for the primary and supplementary searches.Two authors (JB, CH) screened the titles and abstracts against the predefined eligibility.In discrepancies, a consensus was reached via discussions and ineligible publications were discarded.For the primary search, JB screened the full texts of the remaining publications for eligibility and consulted with CH, who then independently screened these full texts in cases of uncertainty.In addition, one other author (SHF) conducted independent cross-checks for a randomly selected 20% of the full texts.For the supplementary search, two authors (JB, CH) independently screened the full texts of the remaining publications.Again, in case of discrepancies, a consensus was reached via discussions, and subsequently, ineligible publications were discarded.Another author (CG) was consulted for her technical/mathematical expertise during the screening process.JB hand-searched the reference lists of eligible publications for any additional relevant literature.In several cases, the corresponding authors of articles were contacted, e.g., to receive full texts or raw data or to clarify results.

Eligibility criteria
We included publications if the following inclusion criteria were met: (1) publications were written in English or German and published until 2 October 2023; (2) publications are original articles published in peer-reviewed scientific journals or conference papers; (3) at least one performance measure of the automatic detection of IA is reported explicitly.For example, the accuracy was calculated by comparing the CGM-based (did not have to exclusively rely on CGM as input) approach against a ground-truth method (e.g., selfreported or observed IA); (4) a CGM-based approach was used to detect IA in vivo in free-living, semi free-living, or laboratory settings.This also included trials of AP systems if criterion 3 was met; (5) only the most recent publication on a specific approach by a particular research group was included if it supersedes preceding publications.
We excluded publications for the following exclusion criteria: (1) the approach was not tested in human participants (e.g., in silico studies); (2) no outcome results were reported (e.g., study protocol publications); (3) outcomes did not include an explicit performance measure describing the results of the automatic detection of IA (e.g., only figures showcasing the CGM trends over time); (4) the methodology was described without sufficient detail.We did not apply restrictions regarding publication date or participant age.

Data extraction
The following information was extracted: (1) first author and publication year; (2) a summary of the study; (3) sample size and, if available, sex and age of participants; (4) participants' diabetes status [no diabetes, prediabetes, T1DM or type 2 diabetes mellitus (T2DM)]; (5) scope of the study (duration/number of IA events) and, if available, information on the IA events (e.g., meal composition); (6) groundtruth/criterion method(s); (7) performance measure(s); (8) details on the CGM device and if applicable other relevant devices used in the study.
JB extracted relevant information from the original publications, and in cases of uncertainty, the respective publications were doublechecked by CH.Two other authors (CG, SHF) also double-checked the extracted information.One other author (CG) further extracted technical details of the tested approaches.

Data synthesis
We synthesized the evidence qualitatively, focusing on answering the three research questions outlined above.Although using an explicit cutoff (e.g., ≥80% F1-score or accuracy) is desirable for performance evaluation and has been used in a related review (16), this approach was not feasible, as only a few publications reported accuracy and/or F1-score values.Furthermore, we appraised the included publications critically.We considered the following aspects of being of concern: (1) errorprone methods for identifying the ground truth of IA [e.g., selfreported IA (58) or retrospective identification from CGM data, whereas inpatient settings with observed IA were generally assumed to be less error-prone]; (2) a sample consisting exclusively of individuals with diabetes as this might limit generalizability to non-diabetic populations; (3) meal announcement/mealaccompanying insulin boluses, as there might be an interference with the (early) postprandial blood glucose levels that are relevant for the automatic detection of IA; (4) algorithm inputs other than CGM since ultimately a CGM-only approach would be desirable to minimize costs and effort.
Many of the screened publications were excluded from the present review because they did not include details on the detection of IA but instead focused on measures of glycemic control (e.g., time in specific glucose ranges).Further, some publications were excluded because they included graphic CGM data with IA marked as such but did not provide quantitative data on the detection of IA.Another common reason for exclusion was investigation in silico, often using virtual patients with T1DM.
We were unable to retrieve the full text of one publication despite several efforts to contact the authors directly.This publication was excluded; however, it was considered likely ineligible based on its abstract.
In several aspects, there was substantial heterogeneity among the included publications.First, the number and type of performance metrics reported for the tested approaches differed substantially.Commonly reported performance metrics included the number of true and/or false positives and/or negatives (including frequencies per day and rates) (64,66,68,70,73,75,84,85,87), sensitivity (60, 61, 70, frontiersin.org(Continued)       (62,65,84,85), and area under the curve (82).However, even when the same metrics were reported, their definition was sometimes inconsistent across publications.For instance, the detection window for true positive detections ranged from 60 to 180 min, depending on the publication (66,73,75).In addition, the general study setup varied between publications, including differences in the sample composition, the use of meal announcement/mealaccompanying insulin boluses, the ground-truth method used for identifying IA, the devices used, and the scope of the data collection (Table 1).

Performance of the approaches
We identified several CGM-based approaches for the automatic detection of IA that achieved high values in the respective performance metrics (Table 1).However, the substantial heterogeneity in the applied methodology and reporting of results needs to be considered.
For instance, Godoy and colleagues achieved 98.9% sensitivity, 96.7% specificity, and 97.6% accuracy with their feedback schemebased algorithm (70).Notably, the algorithm uses certain patientspecific parameters, such as the insulin sensitivity factor derived from participants' usual diabetes treatment (70).Similarly, the algorithm by El Fathi et al. successfully detected 12/12 meals without any false positives and a detection time of 35 min (64).In two publications using the same dataset, Bertrand et al. investigated several IA detection approaches in individuals without diabetes (60,61).A range of performance metrics is reported in both publications.In the first publication, the highest achieved mean sensitivity was 66.8%, and the highest achieved mean specificity was 77.3%, for example (60).In the second publication, the highest achieved mean sensitivity was 66.8%, and the highest achieved mean specificity was 97.7% (61).Importantly, in both publications, the models did not exclusively rely on CGM as input (60,61).Similarly, Palacios et al. had a sample of individuals without diabetes (82).However, their models, too, did not exclusively rely on CGM as input, but also utilized other physiological variables such as heart rate and skin temperature (82).Palacios et al. reported the area under the receiver operating characteristic curve (AUC-ROC) and the area under the precision-recall curve (AUC-PR) (82).For cold-start cases with a window size of k = 110 min, they reported an AUC-ROC of 89.1% and an AUC-PR value of 80.3% (82).For non-cold-start cases and k = 20 min, the AUC-ROC was 99.6%, and the AUC-PR was 96.4% (82).
The performance of CGM-only approaches, which hold particularly great value for practical applications, varied substantially.Sensitivities varied between 20.8% (73) and 96.8% (85).Where reported, average false positive detections ranged from 0.4 (75) to 2.8 (73) IA events per day.Selected publications further reported a false positive rate of 20.8% (84), a false discovery rate of 16.6% (74), and a median precision value of 73.0% (66).Moreover, publications reported detection times between 11.8 min (mean) (73) and 45.0 min (median) (66).Importantly, all CGM-only approaches were tested on samples consisting exclusively of individuals with (pre)diabetes, some of which also used meal announcements/insulin boluses.A detailed description of the performance metrics for each of the included publications is provided in Table 1.

Detection times
The detection time is the relevant metric to evaluate whether the identified CGM-based IA detection approaches could be used in the context of JITAIs.

Discussion
The primary objective of this review was to examine whether CGM can be used to automatically detect IA in (near-)real time.In sum, there are various promising approaches that show satisfactory to excellent performance on measures such as sensitivity and specificity.However, the performance of CGM-based methods for automatically detecting IA varies.Similarly, detection times vary, but currently, they appear too long to administer JITAIs for acutely altering IA.Methodological issues and overall heterogeneity among publications make it difficult to recommend the bestperforming approach.
4.1 Which approaches using CGM for the automatic detection of IA in (near-) real-time have been investigated, and have these approaches relied solely on CGM or also used other data (e.g., sensors/ wearables)?
Our results indicate that both CGM-only approaches and those supplemented with other input data (e.g., accelerometry, photoplethysmography, temperature sensors) have been tested.Moreover, various algorithms have been used to detect IA.Since approaches using different sensor modalities and/or programming methods were successful at automatically detecting IA, it is evident that various solutions can be used for automated, CGM-based IA detection.

How accurate are these approaches in detecting IA?
Our review showed that the performance evaluation of any single approach depends on the respective case and priorities.For example, if the goal is to combine a CGM-based approach with smartphone prompts to enable comprehensive diet logs, the method should have high sensitivity to avoid missing a potential IA (false negative).In this case, specificity would only play a minor role as nothing is lost by sending a prompt in response to a false positive detection -the prompt can remain unanswered by the patient/participant.In contrast, when the goal is to use the CGM-based approach as a standalone IA assessment tool, high specificity would be critical to avoid artificial inflation of the number of daily meals, for example.Thus, a single best approach for all scenarios could not be identified.The substantial heterogeneity of the applied methods and reporting of results, including the broad range of the number and type of reported performance metrics and their varying definitions, made it difficult to compare the performance of the different approaches.
However, collectively, our results demonstrate that there are indeed several relatively well-performing CGM-based approaches for the automatic detection of IA.One example is the feedback schemebased algorithm by Godoy et al., which achieved near-perfect sensitivity, specificity, and accuracy (70).However, this algorithm relies on several patient-specific parameters as input that are derived from participants' usual diabetes treatment (70).Thus, it remains to be determined whether this approach could be adapted to work equally well in individuals without diabetes, for whom these data are not routinely assessed.Similarly, methodological issues further limiting studies' internal and/or external validity pertain to using meal announcements or insulin boluses and focusing on samples with diabetes in the reviewed studies.All included publications suffered at least one such methodological limitation (Table 2).Several reviewed articles reported solutions that relied solely on CGM as input for their IA detection algorithms.Performance among these approaches varied, but sensitivities ≥90% were achieved by several groups (73,84,85), and false positive occurrences < 1 per day were reported (75).This suggests that inputs other than CGM are not necessary to achieve excellent performance in automatically detecting IA.
Of note, some algorithms that incorporated inputs other than CGM might also work with CGM as their only input for the specific goal of IA detection.For example, in Bertrand et al. 's machine-learning algorithms, data from two wrist-worn activity trackers were incorporated in addition to the CGM data (61).However, the 20 most important features were derived from the CGM data (61).Hence, it is likely that an adaptation of their algorithm that relies exclusively on the CGM data as input might also achieve good -albeit likely somewhat worse -IA detection performance.Similar cases can be made for other publications in which insulin data were used as input in addition to the CGM data (59, 63,64).These results suggest that it is possible to automatically detect IA using CGM-based and even CGM-only algorithms.

Can these approaches be used in the context of JITAIs?
Generally, to successfully administer a dietary JITAI, IA must be detected in (near-)real-time.However, precisely how short the detection would have to be depends on the specific goal, as outlined before.Detection times as fast as 9.0 min were reported (70), but most approaches needed 20 to 40 min to detect IA (Table 1).This can generate feedback on IA much faster than traditional dietary assessment methods, such as 24-h recalls, thus creating opportunities for earlier intervention.For instance, detecting deviations from a standardized study procedure (e.g., when IA is detected in a fasting window) is likely possible.Further, when deviations from a specific meal plan (e.g., low carbohydrate) are detected, the plan could be adjusted for the subsequent meals on the same day.Automated meal detection could further trigger behavioral intervention prompts regarding portion size and eating rate (i.e., reminders to eat more slowly) for future meals.However, in most cases, detection times are too long to modify/influence IA truly in the moment it occurs (e.g., a participant on a ketogenic diet has likely already finished a carbohydrate-rich meal by the time it is detected).Regardless, it is debatable if that is really the goal and what intervening during an eating event would look like.x, no; ✓, yes;?, no information available; CGM, continuous glucose monitoring; IA, ingestive activity."x" in all columns would signal minimal methodological concerns and thus an overall rating of four stars in the rightmost column; the more "✓, " the greater the methodological concerns, and consequently, the fewer stars are awarded in the rightmost column.Note that for the overall rating, no available information ("?") was treated as eliciting methodological concerns ("✓").a While insulin pump data were also used in this paper, the meal detection algorithm relies only on glucose data as input.b The authors state that self-reported IA is not seen as the ground-truth method in their work; however, in the absence of direct observation of IA in this study, herein, we assume self-reported IA as the ground-truth.

Implications for clinical and research practice
Our review shows that several CGM-based options for the automatic detection of IA exist.Ultimately, the specific use case will dictate the most suitable approach.Different approaches might be appropriate depending on factors such as the budget, population, targeted level of wearing comfort, and goal of the automatic IA detection.
Notably, other innovative methods for the automatic detection of IA, such as those using wearable-, sensor-, and image-based methods (9,(15)(16)(17)(18), are also promising.These methods may even be superior to CGM-based approaches regarding detection times.Wang and colleagues identified several devices that can quickly detect IA (16), such as a headband device that can detect eating events via chewing sounds within only 3 min (16,90).Similarly, a pilot study by Kumar and colleagues investigating the use of abdominal sounds to detect IA found an average detection time of only 4.3 min (91).It has even been demonstrated that eating events can be predicted ahead of time (16,92).Yang et al. used a camera, a GPS device, and an accelerometer to predict eating and food-purchasing events up to 4 min in advance (92).The authors found that a trained gradient-boosting model achieved a mean accuracy of 72.9% in predicting eating events 0-4 min in advance (92).This highlights that different methodologies might have inherent strengths and limitations.The suboptimal detection times might be considered an inherent limitation of CGM-based approaches.Recent advances have tried to solve the CGM-inherent lag time issue (93), but more research is needed.It remains to be seen whether these limitations inherent to using CGM for automatically detecting IA can be overcome.On the other hand, one key benefit of using CGM might be its unobtrusiveness, which could facilitate its acceptance in practice.This unobtrusiveness contrasts many other, more obtrusive approaches such as glasses and camera-based methods (9,16,18).
A promising prospect might be to use a sensitive CGM-based approach that sends a prompt to the patient/participant asking them to log IA in case of a true positive detection.Thus, the CGM-based approach would serve as an automated reminder.That way, a false positive detection does not automatically lead to erroneous IA information but needs to be verified by the person.In this context, the suboptimal detection times also likely would be acceptable.

Limitations and directions for future research 4.5.1 Sample characteristics
Unsurprisingly, most publications included samples with diabetes, as the primary use case for automated IA detection is AP systems.However, to examine the potential of CGM-based approaches for detecting IA in various populations, more research in more diverse populations, including healthy individuals, should be conducted.This is particularly important as the generalizability of previous findings to non-diabetic individuals is likely limited, for instance, due to the usually far lower variations in blood glucose levels in persons without diabetes (77) as compared to persons with diabetes (78).Thus, there may likely be systematic differences in the performance of such approaches in individuals with diabetes compared to those without diabetes.Moreover, in many studies, meals were announced to the system, and/or manual insulin boluses accompanied the registered meals.For example, Ornetzeder and colleagues evaluated the detection performance of three previously published algorithms (79-81) using meals accompanied by insulin boluses (75).While the resulting performance metrics of this publication and similar others are promising, they need to be interpreted considering the applied insulin boluses.Ornetzeder et al. argue that this potential distortion was deemed acceptable due to a lack of alternative, insulin bolus-free datasets and the time it takes for the administered insulin to achieve its peak action (75).However, it is still possible that the results of CGM-based IA detection approaches might differ in scenarios without exogenous insulin infusions.Specifically, the administered insulin might flatten the blood glucose excursions from the meal's start, making its automatic detection less likely.In line with this, Faccioli et al. state that some of their false negatives might have been related to the attenuated postprandial CGM curves following the administration of meal-accompanying insulin therapy (66).At the same time, it should be considered that the postprandial glucose excursions of individuals with insulin-dependent diabetes would be much more pronounced without insulin treatment than in non-diabetic individuals (82).As such, it could be argued that by administering meal boluses, the postprandial glucose excursions of individuals with diabetes more closely approximate those of individuals without diabetes.Direct evidence is, of course, still necessary to increase confidence in any conclusions.Thus, future studies should ultimately enroll more individuals without diabetes.

Research focus
Moreover, it also needs to be considered that for the initialization of closed-loop systems, background information (e.g., treatment management, physical characteristics of the patient) is typically provided to the system (59).This information may only sometimes be readily available in other contexts.In addition, the goals of algorithms geared toward use in closed-loop/AP systems might differ from approaches aimed at the use for automatic detection of IA in general.For instance, in their AP-oriented work, Kölle et al. focused on glucose excursions caused by larger meals because smaller meals or snacks, which do not cause a substantial increase in blood glucose levels, do not necessarily need to be detected and trigger an insulin bolus to ensure adequate glucose control (73).Yet, in a scenario where the automatic detection of IA via CGM is meant to provide information on any IA -irrespective of its size -this argument does not hold up.This example highlights the potential differences in the setup of algorithms depending on the goal.
Taken together, fundamentally different circumstances and goals may be pursued, and thus, algorithms may be constructed differently, depending on the research question.Consequently, it might be possible to further optimize algorithms to automatically detect IA in research or clinical settings other than closed-loop/AP systems.

Comparability of approaches
There was substantial heterogeneity in how the performance of the investigated approaches was evaluated across the reviewed publications.Thus, as noted by others (66), a direct comparison between the approaches is difficult due to differences in the utilized datasets, preprocessing, and evaluation methods.Differences like these ultimately hamper the search for the best-performing approaches.Performance metrics reported in publications should include at least the following measures: the number of true positives, false positives, and false negatives, which can be used to calculate important metrics such as sensitivity and precision; the detection time, defined as the time from the start of the IA to the time the algorithm detects the IA, whose reporting allows researchers and practitioners to judge whether a specific approach could be used to administer JITAIs, for example.A short detection time followed by a prompt could also allow for more immediate self-reported IA.More accurate self-reports could be the consequence due to diminished recall bias.

Future avenues
In general, more research should be dedicated to using CGM for the specific goal of automatically detecting IA in a broad range of populations, particularly in individuals without diabetes.Such approaches have several potential benefits, but prior research has mainly focused on using CGM for diabetes care and AP systems.However, as explained, algorithms will likely be constructed differently for the specific goal of automatically detecting IA.Moreover, previous findings will have to be replicated and extended in non-diabetic samples to overcome the currently limited generalizability.
Depending on the use case, several advancements would be necessary to rely exclusively on a CGM-based/CGM-only approach for the remote monitoring of IA.To fully automate the logging of IA times in a reliable manner, most systems would have to be even more accurate than they currently are.
If the goal is to further automate IA timing and effectively log macronutrient intake, approaches would have to incorporate specific algorithms for this task.Several publications explored whether estimating macronutrients from CGM data is possible.For instance, Samadi et al. estimated the carbohydrate content of meals (84).Results were promising, with 64.1% of the detected IA events having an absolute carbohydrate estimation error of less than 25 g (84).
Similarly, if the goal is to administer JITAIs to impact acute IA, detection times would have to decrease further.However, as mentioned above, the lag time-caused suboptimal detection times might have to be considered an inherent limitation of CGM-based approaches.Only if future studies succeed at further reducing detection times will the application of CGM-based approaches for dietary JITAIs aiming to alter IA in the moment in the truest sense of the word become possible.This is especially true for cases in which meals are followed by only small and/or delayed postprandial glucose excursions (e.g., after high-fat meals) or when meals contain only a small amount of carbohydrates, as (timely) detection appears difficult here (66,74).It would also be necessary to explicitly test the detection performance in cases of such challenging IA (e.g., ketogenic diets).Empirical data on such cases might enable the prediction of in which settings CGM-based approaches can be used for successfully detecting IA (e.g., only in contexts where at least moderate amounts of carbohydrates are consumed).
We advise that future studies use different approaches on the same dataset, providing comprehensive CGM and objective IA data, and then compare their performance using the abovementioned metrics.A starting point could be to compare the CGM-only approaches highlighted in Table 2.Such a fair and standardized comparison could further illuminate the currently most promising approach(es).
While CGM-only approaches are highly attractive because they only necessitate one single sensor (i.e., the CGM), multi-sensor solutions also hold great potential and should thus be further investigated.Specifically, combining the strengths of different sensors (e.g., CGM and wristbands) may yield superior results as compared to relying on only one sensor, although this remains to be determined empirically.
Lastly, similar to others (10), we strongly advise that researchers use interdisciplinary collaborations to develop new CGM-based dietary monitoring tools to combine technological and biological/ nutritional expertise.Interdisciplinary collaborations should ensure that the resulting tools are useful and optimized from both perspectives.

Conclusion
Based on an exhaustive and systematic literature search, this scoping review shows that it is possible to automatically detect IA using CGM-based approaches.Despite methodological issues and substantial overall heterogeneity among publications, CGM-based dietary monitoring might complement clinical and research practice.

(−
AccuracyXGB-N: 0.82 (0.05).RF-S was deemed to have the best overall performance of the eight models.

Dassau et al. ( 62 )
scheme to check for concordance in meal detections by the abovementioned methods.Importantly, insulin meal boluses were purposefully delayed by 1 h and thus did not confound the postprandial BG changes.

(
al. (64) e Preliminary results from a randomized three-way experiment on the safety and efficacy of CL insulin delivery with or without a meal detection module (an adaptive model-based meal detection algorithm) versus conventional pump therapy after a missed insulin bolus.The data stem from the same study as Palisaitis et al. (65).4 adolescents T1DM Per participant, three 9-h inpatient visits with one uncovered lunch meal with 60 g of CHO per visit were conducted ➔ 4 participants x FPs; time until meal detection = 35 [30-40] min; glucose increase at meal detection time = 2.89 ± 1.72 mmoL/L NIA (Continued) Brummer et al.

Faccioli
multicenter clinical trial on the feasibility of a long-term automated insulin delivery system were used for a retrospective evaluation of a super-twisting-based meal detector.14 days of SAP therapy under freeliving conditions preceded the main phase of the study; these data were used for the evaluation of the meal detector.Due to the use of SAP therapy, manual meal insulin boluses were given, and the results need to be interpreted taking this into consideration.30 (17/13); median age 44 years, algorithm with an additional automatic switching signal generator (SSG), i.e., a meal detection module.Importantly, in the clinical trial meal announcement was used, so the results need to be interpreted in light of this potential bias.

(
estimated meal onset time -actual meal onset time) = 9.0 min and 25 min delay time Insulin pump MiniMed 640G; CGM sampling time = 5 min, but up-sampled to 1 min to increase the detection sensitivity Of note, insulin boluses were used as another algorithm input.The results need to be interpreted considering this.
ROC); (2) area under the precision-recall curve (AUC-PR) ➔ cold-start: k = 110 min: AUC-ROC = 0.891; AUC-PR = 0.803; non-cold-start: Northridge, CA) placed on the abdomen; Equivital Sensor Electronics Module (Equivital I: Hidalgo Ltd., Cambridge, United Kingdom) combined with a heat-sensitive transmitter (pill) to assess heart rate, heat flux, core body temperature; To match the calorimeter data, the Equivital and CGM data were resampled using cubic spline interpolation k = minimal window size Palisaitis et al. (65) e Randomized, three-way, crossover trial comparing an AP system equipped with a meal detection algorithm (AP + MDA) with the AP alone and conventional pump therapy (CSII) in controlling blood glucose levels after a meal without accompanying insulin bolus.The data is from the same study as El Fathi et al. (64).11; 14.9 ± 1.3 years T1DM One self-selected lunch meal which was standardized between interventions for each participant ➔ Mixed meals with 55-65 g of CHO served 4 h after the start of the intervention Inpatient setting or at home with a member of the research staff with three 9

FIGURE 1 PRISMA
FIGURE 1PRISMA Flow diagram (88).a In some cases, more than one reason led to the exclusion of a publication; here, only the primary reason is listed for each publication.b In two cases, two publications of the same study were included.CGM, continuous glucose monitoring; IA, ingestive activity.

TABLE 1
Characteristics of included publications.

TABLE 1 (
Continued) included in the review, ordered by the first author's name.Some publications also include further information (e.g., in silico validation data, CHO content estimation), which are not discussed here.a Only performance measures relevant for our research questions are reported herein.
(65)ications b Bertrand et al.(60)andBertrand et al. (61)are reports of the same study.cTheauthors of this paper report many different performance metrics for each algorithm; for space-saving reasons we only report the lowest and highest values for each performance metric here.These values refer to the models that included CGM data.dTheauthors of this paper report many different performance metrics for each algorithm; for space-saving reasons we only report the lowest and highest means (and the according standard deviations) for each performance metric here.eElFathietal.(64)and Palisaitis et al.(65)are reports of the same study.f We calculated sensitivity, specificity, and accuracy from the total number of TPs, TNs, FPs, and FNs, respectively, and thus, the values refer to the total sensitivity, specificity, and accuracy across the sample.AP, artificial pancreas; AUC, area under the curve; BG, blood glucose; CGM, continuous glucose monitoring; CHO, carbohydrates; CI, confidence interval; CL, closed-loop; CSII, continuous subcutaneous insulin infusion; FN(s), false negative(s); FP(s), false positive(s); LDA, linear discriminant analysis; M, mean; MCC, Matthew's correlation coefficient; NIA, no information were available; SAP, sensor-augmented pump; SD, standard deviation; T1DM, type 1 diabetes mellitus; T2DM, type 2 diabetes mellitus; TN(s), true negative(s); TP(s), true positive(s).

TABLE 2
Critical appraisal of included publications.