Personalized Meal Classification Using Continuous Glucose Monitors Projna Paromitaa , Theodora Chasparia , Seyedhooman Sajjadia , Anurag Dasa , Bobak J. Mortazavia and Ricardo Gutierrez-Osunaa a Dept. of Computer Science & Engr., Texas A&M University Abstract Managing diabetes mellitus (DM) requires monitoring the glucose response to meals, also known as the postprandial glucose response (PPGR). The PPGR to a meal is significantly affected by the amount of carbohydrates, but other macronutrients (e.g., protein, fat, fiber) are also known to affect the PPGR. This suggests that the type of meal consumed can be automatically identified by analyzing the shape of the PPGR, as measured by a continuous glucose monitor (CGM). As a step towards this goal, this study proposes a metric-learning approach to learn personalized PPGR embeddings to account for the inherently large inter-individual variability in PPGRs. Metric learning is implemented with a Siamese neural network (SNN) that models the relative distance between meals consumed by a participant. Em- beddings learned with the SNN outperform features directly extracted from PPGRs, yielding 50% and 77% accuracy on the considered tertiary and binary meal classification tasks, respectively. Findings from this work would ultimately help in designing intelligent user interfaces for assisting patients with DM in dietary monitoring. Keywords Personalization, metric learning, inter-individual variability, continuous glucose monitors, postprandial glucose response 1. Introduction though new technologies based on food pho- Diabetes mellitus (DM) is a chronic progres- tography and wearable sensors are also be- sive metabolic disorder that requires signif- ing investigated, each with their own pros icant self-management, including nutrition, and cons in terms of accuracy and obtru- exercise, and medication [1]. An impor- siveness [2, 3, 4, 5, 6, 7]. For instance, di- tant aspect of daily DM management lies in etary monitoring through memory recall can controlling patients’ postprandial glucose re- be burdensome and is confounded by pa- sponses (PPGR), mostly by monitoring diet. tients’ bias [8]. Food photography can be Current forms of dietary monitoring rely less burdensome, therefore promoting user largely on manual input and memory recall, adherence [8, 9]. However, food photogra- phy methods still require manual patient in- put, which can be noisy and potentially re- Joint Proceedings of the ACM IUI 2021 Workshops, April 13–17, 2021, College Station, USA sult in missing values, especially in the case " projna.paromita@tamu.edu (P. Paromita); of smaller meals or snacks [9, 10]. The pho- chaspari@tamu.edu (T. Chaspari); tos collected via food photography also need hooman130@tamu.edu (S. Sajjadi); adas@tamu.edu (A. to be analyzed by an expert or an automated Das); bobakm@tamu.edu (B.J. Mortazavi); rgutier@tamu.edu (R. Gutierrez-Osuna) machine learning algorithm to obtain food  intake estimates [11]. © 2021 Copyright for this paper by its authors. Use permit- ted under Creative Commons License Attribution 4.0 Inter- Continuous glucose monitors (CGMs) use national (CC BY 4.0). CEUR http://ceur-ws.org CEUR Workshop Proceedings is common in type I diabetes, and is rapidly (CEUR-WS.org) Workshop ISSN 1613-0073 Proceedings increasing in type II diabetes, the most com- within each participant, rather than absolute mon form of the disease [12]. CGMs can be patterns, therefore taking into account indi- a valuable source of information to monitor vidual differences. Metric learning is imple- diet accurately and in a relatively unobtru- mented with a SNN that learns a PPGR trans- sive manner by measuring PPGRs, therefore formation to minimize the distance between they have the potential to accurately capture samples of the same meal consumed by an in- even small meals [13, 14]. CGMs can also dividual and maximize the distance between help researchers gain valuable insights into different meals. The proposed pairwise simi- the complex interplay between PPGRs and larity measure does not require us to estimate the macronutrient composition of a given the distribution of PPGRs for each type of meal, due to the fact that the shape of PP- meal, so it can it is suitable for small-sample- GRs depends on the macronutrient compo- size applications [22]. As an additional step sition of a meal [15, 16, 17, 18, 19]. For ex- toward personalization, we combine partici- ample, carbohydrate-rich meals generally re- pants’ anthropometric and metabolic infor- sult in PPGRs with high and narrow peaks, mation with the learned personalized em- whereas meals rich in fat cause wider peaks, beddings of the SNN. We evaluate our ap- and meals rich in protein depict lower and proach on a publicly-available dataset with moderately broader peaks [20]. However, three types of meals. Results indicate that the anthropometric and metabolic character- PPGR transformations learned from the pro- istics of a person also significantly affect PP- posed metric-learning approach outperform GRs, leading to large inter-individual vari- the conventionally used PPGR statistics for ability to identical meals [20, 21]. This paper the task of meal classification. Namely, in- proposes an algorithm to automatically iden- tegrating participants’ anthropometric and tify meals by analyzing PPGRs collected by metabolic characteristics into classifier fur- CGM devices. Such way of meal back track- ther increases its accuracy in certain cases. ing, along with the popularly used methods of meal tracking would help us to build a 2. Prior work fool-proof application for proper diet moni- Computational models for CGM signals have toring. primarily focused on predicting hyper- or We propose a metric learning approach hypo-glycemic episodes [13, 23]. Prior work to achieve personalized meal classification has further examined the prediction of the based on PPGRs. The objective is to learn PPGR given a specific meal [20, 24, 25, 26, 27]. a transformation that embeds a glucose re- Zeevi et. al. showed that personal informa- sponse while considering within-person dis- tion related to dietary habits, physical activ- tances of consumed meals. For this reason, ity, and gut microbiota can improve the pre- we design a metric learning approach, im- diction of PPGRs for specific meals, which plemented with a Siamese neural network has implications to personalized diet educa- (SNN) architectures, that 1) learns a trans- tion interventions [20]. However, the inverse formation that projects PPGRs for a partic- problem of predicting macronutrients from ipant that belong to similar meals to the PPGRs has not been thoroughly examined, same region in the new feature space, while 2 except for our own work [18, 19]. Anurag et also projecting samples from different meals al. proposed a sparse decomposition model to non-proximal regions of the new fea- for representing PPGRs, while Sajjadi et al. ture space. Our approach models pairwise explored the use of machine learning mod- distances between different types of meals els for predicting the amount of macronutri- ents in meals. Both works were evaluated in a set of 15 participants who consumed 9 pre- designed meals with promising results. Current methods for dietary monitoring primarily rely on computer vision algorithms that seek to detect the components of a meal by analyzing its photographs [4, 5, 6, 28]. Figure 1: Average PPGR per meal averaged However, food photography presents chal- across participants for 140-minute analysis win- lenges for vulnerable populations, such as el- dow (𝐴𝑊 ). derly adults, who may face challenges related to dexterity, coordination, and vision when which is a significant benefit compared to using mobile devices [29]. In addition to food other algorithms that tend to be more “data- photography, other works have proposed hungry." In addition, the resulting model con- wearable food monitoring technologies that tains a small number of parameters, therefore rely on smart utensils and wearable sen- it does not have large memory requirements sors [7, 30]. These approaches tend to pro- or computational cost, and could be imple- vide reliable measurements of the food quan- mented as part of edge computing technolo- tity, but they cannot estimate the macronu- gies, thus also promoting aspects of privacy trient composition. In addition, smart uten- preservation. sils might not always be readily available to the user, therefore potentially resulting in 3. Data Description and missing data. Pre-Processing Given the expanding use of CGMs in dia- We used a publicly available dataset of PP- betic populations, our method can contribute GRs contaning data from 30 participants (25- to providing a feasible alternative to current 65 years) [21]. Individual glucose measure- food monitoring methods. Overall, our con- ments were collected 5 minutes apart using tribution to the existing literature is as fol- a Dexcom G4 device. Each participant con- lows: (1) While most studies have focused on sumed three types of standardized meals at predicting one’s PPGR based on a given meal, most twice, resulting in at most six meal the inverse problem of estimating the con- samples per participant. The meals were stituents of a meal based on the PPGR is rela- isocaloric, but varied in the amount of pro- tively unexplored. Limited prior work from tein, fat, and fiber [21]. The three meals were: our group has investigated the problem of (1) cornflakes and milk (CF), which were low estimating macronutrient composition from in fiber and high in carbohydrates; (2) peanut PPGRs [31, 18, 19], but meal classification butter sandwich (PB), which had a high from PPGRs has not been yet examined; and amount of fat and protein; and (3) a PRO- (2) in contrast with the majority of work, BAR protein bar (Bar), which had moderate which relies on modeling class-wise distri- amounts of fat and protein. Fig. 1 illustrates butions of PPGRs [20, 24, 25], our work pro- the average PPGR across individuals for each poses a metric learning approach that mod- of the three meals, which indicates that there els the pairwise distance between different are marked differences in PPGRs across the meals consumed by each participant. The three types of meals. The data further include proposed metric learning approach does not participants’ anthropometric characteristics, require a large number of labelled samples, including age and body mass index (BMI), as well as metabolic characteristics, such as insulin, fasting blood glucose (FBG), an oral glucose tolerance test (OGTT), Hemoglobin A1c (HbA1C), high sensitive C-reactive pro- tein (hsCRP), and triglyceride to high-density lipoprotein (HDL) ratio (tri/HDL). Anthro- pometric and metabolic characteristics were combined with the PPGR features and com- Figure 2: PPGR signals and family of Gaussian kernels centered at five-time points and used to prised the input of the proposed machine compute the area under the curve (AUC) of the learning models as an additional step to per- PPGR. sonalization. Prior to feeding PPGRs to the SNN, we by subtracting its minimum value and divid- performed data-cleaning procedures, includ- ing by the difference between the maximum ing linear interpolation of missing samples and minimum value. Finally, we extracted and baseline correction. Baseline correction the area under the curve (AUC) at 3, 5, and 8 was done by subtracting the mean of the distinct time points, following Huo et.al. [31], first 6 data points prior to meal consumption, resulting in a total of 16 features. We experi- which we assumed served as the fasting glu- mented with these features extracted both in cose level. Following prior work [20, 21, 31], the normalized and non-normalized PPGRs. we also experimented with various analy- AUCs for 5-time points of the analysis win- sis windows (𝐴𝑊 ), including 140, 90, and dow are illustrated in Fig. 2. These AUC fea- 65 minutes. For the purpose of our experi- tures capture fasting glucose levels, as well as ments, we aim to model pairwise distances glucose rise and recovery patterns in various between the same meal and different meals. time resolutions. For this reason, our analysis included only those participants who consumed each meal 4.2. Personalized metric learning at least twice. This resulted in 20 participants of PPGR embeddings we could use to model pairwise distances be- We designed a metric learning algorithm tween Bar and CF and, and 19 participants that models the pairwise distance between for the Bar vs. PB meal and PB vs. CF meals. meals consumed by a participant (Fig. 3). The number of participants reduced to 11 The problem was formulated as both binary when we considered all three meals for ter- (i.e., CF/PB, PB/Bar, Bar/CF) and tertiary (i.e., tiary classification. When considering 𝐴𝑊 CF/PB/Bar) classification tasks (Section 3). of 65 and 90 minutes, the number of partici- Inputs to the model consisted of the 16 AUC pants became 14 and 16 for the tertiary com- features of PPGRs (Section 4.1). Let 𝑔𝐖 be bination, respectively. a function parameterized by 𝐖, which per- forms a transform of the original AUC space 4. Methodology 𝐱 and transformed embedding 𝑔𝐖 (𝐱). Also, let 𝑛𝑐 be the set of samples belonging to 4.1. Glucose response feature class 𝑐 from participant 𝑛. The parameters 𝐖 extraction are learned to minimize the distance of sam- We computed the maximum and minimum of ples from the same meal-type and maximize each participant’s baseline-corrected PPGR the distance of samples from different meals (Section 3). Then, we normalized the signal Figure 3: Schematic representation of the proposed personalized meal classification based on PPGRs. Metric learning learns PPGR embeddings based on pairwise meal distances within a participant. An- thropometric and metabolic characteristics are combined with the learned PPGR embeddings for the final meal classification. within a participant: We compare the proposed personalized 𝐖∗ = 𝑎𝑟𝑔 min max(∑ ∑ ∑ 𝑑 (𝑔𝐖𝟏 (𝐱), 𝑔𝐖𝟏 (𝐱′ )) glycemic embeddings learned by the SNN 𝐖𝟏 𝐖𝟐 𝑛 𝑐 𝐱,𝐱′ ∈𝑛𝑐 architecture to the raw AUC features that (1) +∑∑ ∑ 𝑑 (𝑔𝐖𝟐 (𝐱), 𝑔𝐖𝟐 (𝐱′ ))) comprised the input to two baseline mod- 𝑛 𝑐≠𝑐 ′ ′ els: a 1-layer feed-forward neural network 𝑐 ′ 𝐱 ∈ 𝑛 , 𝐱 ∈ 𝑛 𝑐 where 𝑑(⋅, ⋅) was the 𝑙2−norm. Pairwise (FNN) with 16 nodes and a logistic regres- metric learning was performed via a SNN sion model. The architecture of the FNN (Fig. 3). The SNN included 3 layers with 16 was selected to be equivalent to the 16-node neurons each (i.e., siamese network; Fig. 3). fully-connected layer following the output of The third layer of the SNN comprised of the SNN architecture. We anticipate that the a 16-dimensional output, which represented proposed model performs well in all scenar- the transformed PPGR input samples 𝑔𝐖 (𝐱) ios where learning the original distribution is and 𝑔𝐖 (𝐱′ ) in each branch. Following that, difficult due to the inherent inter- and intra- the 𝑙2−norm between 𝑔𝐖 (𝐱) and 𝑔𝐖 (𝐱′ ) was subject variability. However, because we do computed. The learned glycemic embedding not have this data, comparing against deeper 𝑔𝐖 (𝐱) was fed into a set of fully-connected networks with larger numbers of tunable pa- layers comprised of 16 neurons (i.e., feedfor- rameters would either underfit or overfit. ward neural network; Fig. 3), which learned All classification experiments were per- a feature transformation between the PPGR formed using a leave-one-subject-out cross- embedding and the type of meal. Glycemic validation. Simple classification accuracy embeddings were learned based on the orig- was averaged over 40 iterations to remove inal PPGR, as well as the normalized PPGR random effects from parameter initialization using min-max normalization. Bayesian hy- and dropout. The classes considered here perparameter optimization [32] was used to are balanced, therefore the chance accuracy optimize the dropout between layers (i.e., is approximately 33.33% for the tertiary and {0.0, 0.1, 0.2, 0.3, 0.4, 0.5}) and the 𝑙2−kernel 50% for the binary task. A paired t-test with regularization of each layer of the network the assumption of unequal variance among (i.e., {10 , 10 , 10 , 10 }). −1 −2 −3 −4 the two groups was used to calculate signifi- Table 1 4.3. Integrating anthropometric Accuracy (%) of tertiary meal classification, us- and metabolic ing the area under the curve (AUC) of PPGRs as characteristics to the PPGR an input to a feedforward neural network (FNN) and a logistic regression (LR) model, as well as the embedding AUC embeddings learned by the proposed metric We incorporated the anthropometric and learning approach. metabolic measurements of each participant Normalized PPGR to (1) the glycemic embedding 𝑔𝐖 (𝐱) learned Analysis window Metric learning FNN LR by the SNN (Section 4.2), followed by a fully- 140 minutes 50.00 47.84 50.00 connected layer (Fig. 3); and (2) the raw 90 minutes 46.14*** 42.71 42.85 AUC features followed by the FNN and logis- 65 minutes 45.89 45.2 45.83 *: 𝑝 <0.05 ; **: 𝑝 <0.01; ***:𝑝 <0.001 tic regression models. We report the corre- Non-normalized PPGR sponding classification results in order to see Analysis window Metric learning FNN LR whether the inclusion of these features fur- 140 minutes 44.86 47.84 50 ther improves the meal classification perfor- 90 minutes 45.59* 42.7 44.04 65 minutes 45.63 45.21 45.63 mance. Anthropometric and metabolic fea- *: 𝑝 <0.05 ; **: 𝑝 <0.01; ***:𝑝 <0.001 tures were included in five unique combina- tions according to their individual and group Table 2 characteristics. Combinations 1 and 2 con- Accuracy (%) of binary meal classification be- sisted of only anthropometric features (i.e., tween pairwise combinations of peanut butter age, BMI) and only metabolic measurements (PB), cornflakes (CF), and protein bar (Bar), us- (i.e., insulin, FBG, OGTT, HbA1C, hsCRP, ing the area under the curve (AUC) of PPGRs in Tri/HDL), respectively. Meanwhile, combi- a feedforward neural network (FNN) and a logis- nation 3 included all features from both com- tic regression (LR) model, as well as the AUC em- binations 1 and 2. In combination 4, we beddings learned by the proposed metric learning approach. The analysis window is 140 minutes. only considered features that were signifi- cantly correlated with the AUCs extracted Normalized PPGR from PPGR (i.e., age, insulin, OGTT, HbA1C, Task Metric learning FNN LR Tri/HDL). The final combination 5 consisted PB-CF 74.79*** 72.29 72.22 of three metabolic measurements (i.e., FBG, PB-Bar 69.68*** 65.19 68.42 OGTT, HbA1C), which were directly related Bar-CF 52.49 53.59 56.25 to the participant’s glucose level. *: 𝑝 <0.05 ; **: 𝑝 <0.01; ***:𝑝 <0.001 Non-normalized PPGR 5. Results Task Metric learning FNN LR We report the accuracy of the proposed met- PB-CF 66.36 66.54 69.73 ric learning and baseline models (Section 4.2). PB-Bar 61.92 60.39 64.47 Table 1 presents the tertiary classification Bar-CF 51.33 52.96 52.5 results over various analysis window (𝐴𝑊 ) *: 𝑝 <0.05 ; **: 𝑝 <0.01; ***:𝑝 <0.001 lengths, while Table 2 presents the binary classification results for 𝐴𝑊 = 140 minutes cant differences between the proposed metric for all meal combinations. We observe that learning and the two baselines. the metric learning approach performs sig- nificantly better than the 1-FNN and logistic regression baseline in the majority of cases for the tertiary task (Table 1) and in many Table 3 Accuracy (%) of binary meal classification between pairwise combinations of peanut butter (PB), corn- flakes (CF), and protein bar (Bar), combining the PPGR embeddings learned by the proposed metric learning approach with anthropometric and metabolic characteristics. The analysis window is 140 minutes. The normalized PPGR was used. Only PPGR & Anthropometric/Metabolic Combination Task PPGR 1 2 3 4 5 PB-CF 74.79 77.41*** 71.4 72.56 72.11 72.11 PB-Bar 69.68 68.1 61.6 61.44 60.85 65.54 Bar-CF 52.49 53.93 54.75* 55.01** 53.42 53.79 *: 𝑝 <0.05 ; **: 𝑝 <0.01; ***:𝑝 <0.001 Combination 1: [Age, BMI]; Combination 2: [Insulin, FBG, OGTT, HbA1C, hsCRP, Tri/HDL]; Combination 3: [Age, BMI, Insulin, FBG, OGTT, HbA1C, hsCRP, Tri/HDL]; Combination 4: [Age, Insulin, OGTT, HbA1C, Tri/HDL]; Combination 5: [FBG, OGTT, HbA1C] cases for the binary task (Table 2). This indi- nessed in other cases. cates that learning a personalized embedding learning through metric learning can bene- 6. Discussion fit meal classification performance, even af- Our results indicate that personalized PPGR ter normalizing the corresponding PPGR. We embeddings through metric learning can ef- further compare AUC features from the nor- fectively differentiate between meals. While malized and non-normalized PPGR through a the scope of our current work is limited due paired t-test with the assumption of unequal to the sparsity of datasets that include CGM variance. For tertiary classification with 140 signals with concurrent meal intake anno- minutes analysis window, we achieve accu- tation, we believe that the ability to predict racy up to 50% for the normalized signal, dietary intake has a broad range of appli- which is significantly higher (𝑝 < 0.001) than cations in the context of automated real-life the 44% accuracy from the non-normalized dietary monitoring and interventions. Par- signal (Table 1). Similarly, for the binary clas- ticularly, these can have valuable implica- sification, results demonstrate the effective- tions for improving the accuracy of auto- ness of normalizing PPGR (Table 2), reaching matic diet monitoring based on CGM devices 75% accuracy. for people with (pre)(diabetes. The model re- The combination of age and BMI with the quires small amount of data for personaliza- PPGR embeddings learned by the proposed tion and also includes a small number of pa- metric learning approach depicted the best rameters, therefore making it ideal for a light results. Also, combinations 2 and 3, which in- user interface (UI). Moreover, the continuous clude participants’ metabolic characteristics, PPGR collection through CGMs ensures that improve classification for the Bar-CF meal no meal–no matter how small–is overlooked, pair from 52.49% to 54.75% and 55.01%, re- therefore can potentially accommodate users spectively. Individual characteristics seem to with non-routine eating patterns. Overall, benefit classification tasks that are difficult to the model could be easily added to the exist- learn solely from the PPGR, such as the CF ing user platforms that are compatible with and Bar which depicted similar PPGR pat- CGM signals, therefore allowing patients to terns (Fig. 1), while no improvement is wit- monitor their PPGR patterns and better un- derstanding the effect of each meal on their acteristics appears to partially help in certain PPGR. These can be also beneficial for de- cases. As part of our future work, we plan veloping new technology-assisted dietary in- to explore the feasibility of this system in terventions, in which patients can visualize, classifying diverse real-life meals, which can understand, and internalize the interplay be- eventually contribute to effective dietary in- tween meal intake and PPGR, therefore pro- terventions. We also plan to collect data from moting positive behavior change [33]. 90 participants, which will provide us with Despite the encouraging results, this study the opportunity to evaluate our approach at presents various limitations. First, the stan- a broader scale with a wider array of repeated dardized meals considered here are similar meals. in terms of calorie intake and carbohydrate content, therefore rendering the PPGR sim- References ilar across meals (Fig 1). This might be a potential reason why the final classifica- [1] S. R. Shrivastava, P. S. Shrivastava, J. Ra- tion accuracies were modest. Second, our masamy, Role of self-care in manage- work is limited to binary and tertiary clas- ment of diabetes mellitus, Journal of di- sification of standardized meals, while sig- abetes & Metabolic disorders 12 (2013) nificantly more meal diversity exists in real- 14. world settings. We note that the dataset [2] P. Novak, B. K. Seljak, F. Novak, Design- that we have utilized is the only available ing visual interface for nutrition track- dataset in which a given meal is administered ing of patients with parkinson’s disease to each participant twice, a component es- (????). sential to our analysis. Third, the addition [3] S. Kim, T. Schap, M. Bosch, R. Ma- of anthropometric and metabolic character- ciejewski, E. J. Delp, D. S. Ebert, C. J. istics marginally improves the classification Boushey, Development of a mobile user performance. Given that the data were col- interface for image-based dietary as- lected by healthy individuals, we observed sessment, in: Proceedings of the 9th In- little variation in their metabolic character- ternational Conference on Mobile and istics, which may be a potential factor con- Ubiquitous Multimedia, 2010, pp. 1–7. tributing to the marginal increase in the sys- [4] A. Bedri, D. Li, R. Khurana, tem performance when such features were K. Bhuwalka, M. Goel, Fitbyte: added. Prior studies also found that the Automatic diet monitoring in uncon- HbA1C and FBG are the most highly cor- strained situations using multimodal related metabolic measurements with PPGR sensing on eyeglasses, in: Proceedings [21, 20], which is also reflected in our results. of the 2020 CHI Conference on Human Factors in Computing Systems, 2020, pp. 1–12. 7. Conclusion [5] H. Kalantarian, N. Alshurafa, M. Sar- We have shown that the personalized PPGR rafzadeh, A survey of diet monitoring embeddings learned with the proposed met- technology, IEEE Pervasive Computing ric approach outperform the original PPGR 16 (2017) 57–65. features for meal classification. PPGR nor- [6] H. Hassannejad, G. Matrella, malization significantly (p<0.05) improves P. Ciampolini, I. De Munari, M. Mor- performance, while adding individual char- donini, S. Cagnoni, Automatic diet monitoring: a review of computer vision and wearable sensor-based ment: consensus guidelines for contin- methods, International journal of uous glucose monitoring (cgm), Di- food sciences and nutrition 68 (2017) abetes technology & therapeutics 10 656–670. (2008) 232–246. [7] T. Vu, F. Lin, N. Alshurafa, W. Xu, Wear- [14] S. E. Berry, A. M. Valdes, D. A. Drew, able food intake monitoring technolo- F. Asnicar, M. Mazidi, J. Wolf, J. Capdev- gies: A comprehensive review, Com- ila, G. Hadjigeorgiou, R. Davies, puters 6 (2017) 4. H. Al Khatib, et al., Human postpran- [8] L. E. Burke, J. Wang, M. A. Sevick, Self- dial responses to food and potential for monitoring in weight loss: a system- precision nutrition, Nature medicine atic review of the literature, Journal of 26 (2020) 964–973. the American Dietetic Association 111 [15] J. M. Miles, A role for the glycemic in- (2011) 92–102. dex in preventing or treating diabetes?, [9] Å. Norman, K. Kjellenberg, D. Tor- The American journal of clinical nutri- res Aréchiga, M. Löf, E. Patterson, “ev- tion 87 (2008) 1–2. eryone can take photos.” feasibility and [16] R. R. Holman, S. K. Paul, M. A. Bethel, relative validity of phone photography- D. R. Matthews, H. A. W. Neil, 10-year based assessment of children’s diets–a follow-up of intensive glucose control mixed methods study, Nutrition Jour- in type 2 diabetes, New England journal nal 19 (2020) 1–14. of medicine 359 (2008) 1577–1589. [10] J. Most, P. M. Vallo, A. D. Altazan, [17] D. M. Nathan, D. R. Group, et al., L. A. Gilmore, E. F. Sutton, L. E. Cain, The diabetes control and complications J. H. Burton, C. K. Martin, L. M. Red- trial/epidemiology of diabetes interven- man, Food photography is not an accu- tions and complications study at 30 rate measure of energy intake in obese, years: overview, Diabetes care 37 (2014) pregnant women, The Journal of nutri- 9–16. tion 148 (2018) 658–663. [18] A. Das, B. Mortazavi, T. Chaspari, S. Saj- [11] M. Archundia Herrera, C. B. Chan, Nar- jadi, P. Paromita, L. Ruebush, N. Deutz, rative review of new methods for as- R. Gutierrez-Osuna, A sparse coding sessing food and energy intake, Nutri- approach to automatic diet monitoring ents 10 (2018) 1064. with continuous glucose monitors, in: [12] Continuous glucose monitoring (CGM) International Conference on Acoustics, systems/devices market size, share Speech, Signal Processing (ICASSP), trends analysis report by component ???? (transmitters, sensors, insulin pumps), [19] S. Sajjadi, B. Mortazavi, A. Das, T. Chas- by end use (hospitals, homecare), and pari, P. Paromita, L. Ruebush, N. Deutz, segment forecasts, 2018 - 2024, 2018. R. Gutierrez-Osuna, Towards the de- https://www.researchandmarkets. velopment of subject-independent in- com/reports/4613458/ verse metabolic models, in: Interna- continuous-glucose-monitoring-cgm. tional Conference on Acoustics, Speech, [13] I. B. Hirsch, D. Armstrong, R. M. Signal Processing (ICASSP), ???? Bergenstal, B. Buckingham, B. P. Childs, [20] D. Zeevi, T. Korem, N. Zmora, D. Israeli, W. L. Clarke, A. Peters, H. Wolpert, D. Rothschild, A. Weinberger, O. Ben- Clinical application of emerging sen- Yacov, D. Lador, T. Avnit-Sagi, M. Lotan- sor technologies in diabetes manage- Pompan, et al., Personalized nutrition by prediction of glycemic responses, proach to diabetic blood glucose predic- Cell 163 (2015) 1079–1094. tion, volume 3, 2017. [21] H. Hall, D. Perelman, A. Breschi, [28] C. K. Martin, T. Nicklas, B. Gunturk, P. Limcaoco, R. Kellogg, T. McLaughlin, J. B. Correa, H. R. Allen, C. Champagne, M. Snyder, Glucotypes reveal new pat- Measuring food intake with digital pho- terns of glucose dysregulation, PLoS bi- tography, Journal of Human Nutrition ology 16 (2018) e2005143. and Dietetics 27 (2014) 72–81. [22] S. Motiian, M. Piccirilli, D. A. Adjeroh, [29] K. A. Siek, Y. Rogers, K. H. Connelly, Fat G. Doretto, Unified deep supervised do- finger worries: how older and younger main adaptation and generalization, in: users physically interact with pdas, in: Proceedings of the IEEE International IFIP Conference on Human-Computer Conference on Computer Vision, 2017, Interaction, Springer, 2005, pp. 267–280. pp. 5715–5725. [30] Q. Huang, Z. Yang, Q. Zhang, Smart-u: [23] W. Gu, Y. Zhou, Z. Zhou, X. Liu, H. Zou, Smart utensils know what you eat, in: P. Zhang, C. J. Spanos, L. Zhang, Sug- IEEE INFOCOM 2018-IEEE Conference armate: Non-intrusive blood glucose on Computer Communications, IEEE, monitoring with smartphones, Pro- 2018, pp. 1439–1447. ceedings of the ACM on interactive, [31] Z. Huo, B. J. Mortazavi, T. Chaspari, mobile, wearable and ubiquitous tech- N. Deutz, L. Ruebush, R. Gutierrez- nologies 1 (2017) 1–27. Osuna, Predicting the meal macronutri- [24] J. M. Colmenar, S. M. Winkler, G. Kron- ent composition from continuous glu- berger, E. Maqueda, M. Botella, J. I. Hi- cose monitors, in: 2019 IEEE EMBS In- dalgo, Predicting glycemia in diabetic ternational Conference on Biomedical patients by evolutionary computation & Health Informatics (BHI), IEEE, 2019, and continuous glucose monitoring, in: pp. 1–4. Proceedings of the 2016 on Genetic and [32] J. Snoek, H. Larochelle, R. P. Adams, Evolutionary Computation Conference Practical bayesian optimization of ma- Companion, 2016, pp. 1393–1400. chine learning algorithms, in: Ad- [25] D. J. Albers, M. Levine, B. Gluckman, vances in neural information process- H. Ginsberg, G. Hripcsak, L. Mamyk- ing systems, 2012, pp. 2951–2959. ina, Personalized glucose forecasting [33] A. Kankanhalli, J. Shin, H. Oh, Mobile- for type 2 diabetes using data assimi- based interventions for dietary behav- lation, PLoS computational biology 13 ior change and health outcomes: scop- (2017) e1005232. ing review, JMIR mHealth and uHealth [26] J. M. Velasco, S. Winkler, J. I. Hidalgo, 7 (2019) e11312. O. Garnica, J. Lanchares, J. M. Colme- nar, E. Maqueda, M. Botella, J.-A. Ru- bio, Data-based identification of predic- tion models for glucose, in: Proceed- ings of the Companion Publication of the 2015 Annual Conference on Genetic and Evolutionary Computation, 2015, pp. 1327–1334. [27] H. N. Mhaskar, S. V. Pereverzyev, M. D. van der Walt, A deep learning ap-