Personalized Prediction of Change in Fasting Blood Glucose Following Basal Insulin Adjustment in People With Type 2 Diabetes: A Proof-of-Concept Study

Abstract

Aims:

For people with type 2 diabetes treated with basal insulin, suboptimal glycemic control due to clinical inertia is a common issue. Determining the optimal basal insulin dose can be difficult, as it varies between individuals. Thus, insulin titration can be slow and cautious which may lead to treatment fatigue and non-adherence. A model that predicts changes in fasting blood glucose (FBG) after adjusting basal insulin dose may lead to more optimal titration, reducing some of these challenges.

Objective:

To predict the change in FBG following adjustment of basal insulin in people with type 2 diabetes using a machine learning framework.

Methods:

A multiple linear regression model was developed based on 786 adults with type 2 diabetes. Data were divided into training (80%) and testing (20%) sets using a ranking approach. Forward feature selection and fivefold cross-validation were used to select features.

Results:

Participants had a mean age of approximately 59 years, a mean duration of diabetes of 12 years, and a mean HbA_1c at screening of 65 mmol/mol (8.1%). Chosen features were FBG at week 2, basal insulin dose adjustment from week 2 to 7, trial site, hemoglobin level, and alkaline phosphatase level. The model achieved a relative absolute error of 0.67, a Pearson correlation coefficient of 0.74, and a coefficient of determination of 0.55.

Conclusions:

A model using FBG, insulin doses, and blood samples can predict a five-week change in FBG after adjusting the basal insulin dose in people with type 2 diabetes. Implementation of such a model can potentially help optimize titration and improve glycemic control.

Keywords

basal insulin titration dose guidance glycemic control machine learning multiple linear regression type 2 diabetes mellitus

Introduction

Suboptimal glycemic control among people with type 2 diabetes treated with basal insulin is a well-known and documented issue leading to persistent periods of hyperglycemia increasing the risk of severe micro- and macrovascular complications (e.g., blindness and cardiovascular diseases).^1-4,5 The reasons for suboptimal glycemic control are manifold and often caused by clinical inertia-induced behavior among people with type 2 diabetes and clinicians.¹ Furthermore, finding the optimal basal insulin dose, referred to as insulin titration, is complicated, as the optimal dose differs from person to person.^6,7 The individual optimal dose will, moreover, likely change over time due to variations in pancreatic insulin secretion, insulin sensitivity, lifestyle changes, stress levels, and so on.^1,3,7,8 Hence, identifying the optimal basal insulin dose is a pursuit that requires time and experience from clinicians together with patience and understanding from the person with type 2 diabetes undergoing the treatment.^1,2,6,9 These circumstances can lead to prolonged and conservative titration, potentially causing treatment fatigue and non-adherence among people with type 2 diabetes leading to suboptimal glycemic control despite the well-established efficiency of treatment with basal insulin from clinical trials.^2,6,9

Prior work has mainly focused on paper-based titration algorithms similar to the standard of care in outpatient and telemonitoring settings.¹⁰ A number of studies have also investigated the effect of utilizing apps and other digital devices in regard to basal insulin titration.^11-14 A few studies have explored mathematical models’ abilities within basal insulin titration in people with type 2 diabetes. These studies have mainly focused on utilizing control theory methods.^8,15,16 A recent study by Chen et al¹⁷ investigated machine learning approaches’ capabilities in basal insulin titration. To the knowledge of the authors of this work, this is the first study to utilize machine learning methods for basal insulin titration in people with type 2 diabetes. The study investigated six machine learning methods’ ability to predict the prescribed basal insulin dose based on data from electronic health records of 819 people with type 2 diabetes which showed promising results.¹⁷ However, it remains unclear whether the insulin dose prescribed was optimal for the subjects since the results reported by Chen et al¹⁷ offer limited insights into the glycemic control of the 819 people with type 2 diabetes upon which the models were developed. A model able to foresee the effect of a specific basal insulin dose on the fasting blood glucose (FBG) in people with type 2 diabetes during a typical titration period may ease some of the challenges related to titration. This could help identify the optimal basal insulin dose timelier by enabling more efficient and personalized titration. Thus, this study aimed to develop a model able to predict the effect of a change in basal insulin dose on FBG in people with type 2 diabetes using a machine learning framework.

Methods

Data Source

The data used to develop the present machine learning model originated from a 26-week clinical trial (ClinicalTrials.gov: NCT01819129) which investigated the efficacy and safety of fast-acting insulin aspart compared with insulin aspart in adults (>18 years) with type 2 diabetes treated with insulin glargine and metformin.¹⁸ Other antidiabetic treatment was disconnected. Before randomization eligible subjects underwent a 7-week run-in period where insulin glargine was actively titrated according to a standard of care titration scheme to the target of 72.0 to 90.0 mg/dL. Subjects enrolled in the study had an HbA_1c of 53 to 80 mmol/mol (7.0%-9.5%) at screening. Only data collected during the run-in period were used to develop the prediction model since participants were only treated with basal insulin analogs combined with metformin in this period.

Demographic information and blood and urine samples were collected at the screening visit. These were used as potential features in the machine learning model (see Table 1). During the run-in period, self-reported FBG was measured with a glucometer, and self-reported basal insulin doses were collected on three consecutive days each week. Although the clinical trial provided a glucometer that was factory-calibrated to plasma glucose values, there are still concerns about the accuracy of the glucose measurements due to their self-reported nature.¹⁸ However, it is worth noting that efforts were made to address these concerns by providing standardized equipment. Self-reported FBG was used to titrate participants’ insulin doses weekly. Furthermore, self-reported hypoglycemic events were collected.

Table 1.

Features Extracted From the Data.

Features
Sex
Age
Race
Trial site
Duration of diabetes
Median of insulin doses at week 1
Median of insulin doses at week 2
Median of insulin doses at week 7
Mean of FBG at week 1
Mean of FBG at week 2
Change in median of basal insulin dose from week 1 to 2
Change in mean of FBG from week 1 to 2
Ratio between the median of basal insulin doses at week 1 and 2
Ratio between the median of basal insulin doses at week 2 and 7
Total number of self-reported hypoglycemic events during the observation window
Blood samples collected at screening (Albumin in serum, alkaline phosphatase in serum, alanine aminotransferase in serum, serum glutamic oxaloacetic transaminase, basophils, bilirubin in serum, creatinine in serum, calculated creatinine clearance rate in serum, eosinophils, HbA1c, hematocrit, hemoglobin, potassium in serum, lymphocytes, monocytes, neutrophils, thrombocytes, protein in serum, erythrocyte sedimentation rate, sodium in serum, and leukocytes)
Urine samples collected at screening (Albumin, creatinine, ketones, protein, and erythrocyte sedimentation rate)

Abbreviation: FBG, fasting blood glucose.

Feature Extraction

The first step in the machine learning framework was to extract relevant features. The run-in period was split into an observation and a prediction window (see Figure 1). The first week of the run-in period was allocated to extract features that could give insights into the participants’ insulin sensitivity and pancreatic insulin production since c-peptide and similar tests were unavailable from the blood samples. Therefore, the prediction time was the beginning of the second week of the run-in period with a prediction window of approximately five weeks.

Figure 1.

Splitting of data into an observation window from where features were extracted and a prediction window which marks the time between prediction time and prediction target. The numbers 1 to 7 represent the days of the week.

The prediction target was the change in self-reported FBG from week 2 to 7. An overview of all features extracted from the data is seen in Table 1.

In total, 41 features were extracted and formed the pool of potential input variables for the prediction model.

Preprocessing of Data

The second step in the machine learning framework was to preprocess the data before model development. First, categorical features were encoded, so a numeric number represented different levels in the categorical features. Nominal categorical features (sex, race, and trial site) were hot one encoded, whereas ordinal categorical features (ketones in urine, protein in urine, and erythrocyte sedimentation rate in urine) were ordinal encoded. Second, data points below or above the limit of detection in collected blood and urine samples were handled using single imputation. Data points below the limit were replaced with the limit minus one and those above the limit were replaced with the limit plus one. Third, features with >20% missing data were excluded and after that participants with missing data were excluded (see Figure 2).

Figure 2.

The sorting of participants and features that occurred during preprocessing of the data prior to model development. n refers to the number of participants, and m refers to the number of features.

Finally, data were divided into 80% for model training and 20% for model testing stratified based on the change in FBG using a ranking approach (see Figure 3). The training data set was normalized before model development.

Figure 3.

The ranked approach to stratify participants into training and test data sets and create five subsets for cross-validation. The gradient of the blue color indicated the degree of change in fasting blood glucose from weeks 2 to 7 for the participants. The lighter the color, the more positive change, and the darker the color, the more negative.

Model Training and Validation

The third step of the machine learning framework was to train and validate a model. Using forward selection, fivefold cross-validation was used to select the most informative features while training a multiple linear regression model optimized on relative absolute error (RAE). The addition of features was continued until the model performance converged. Python version 3.8.8 was used to develop the model. An illustration of the process of training the prediction model is seen in Figure 4.

Figure 4.

The combined feature selection and model training process. In step 1, all features are tested separately, and the feature resulting in the lowest average relative absolute error (RAE) is selected. In step 2, the remaining features are tested in combination with selected features and added to the pool of selected features if a reduction in average RAE is seen. This process continues until performance converges. In step 3, a multiple linear regression model is trained on all selected features using fivefold cross-validation.

To validate the performance of the developed prediction model, the mean of RAE, mean of Pearson correlation coefficient (r) with associated two-sided P value, and coefficient of determination (R²) across the validation folds in the fivefold cross-validation were reported together with confidence intervals. Furthermore, a scatterplot of observed versus predicted data points for the training data set was constructed using the validation folds in the fivefold cross-validation. As part of the results, the model equation with normalized beta coefficients, representing the impact of a one-unit increase of a feature on the mean change in FBG when the remaining features are fixed, was computed.

Test of the Model

The prediction model was tested using the allocated test data set to determine the model’s performance. RAE, r with associated two-sided P value, and R² were reported. A scatterplot of observed versus predicted data points for the test data set was also constructed.

Results

Table 2 shows the baseline characteristics of the 786 participants divided into training and test data sets. The participants had a mean age of approximately 59 years, a mean duration of diabetes of 12 years, and a mean HbA1c at screening of 65 mmol/mol (8.1%) in both data sets. Thus, no apparent difference between the participants in the training and test data set was seen from visual inspection of Table 2.

Table 2.

Baseline Characteristics of Participants in the Training and Test Data Sets.

Features	Training data setn = 628	Test data setn = 158
Age (years)	59.7 ± 9.4	59.1 ± 9.2
Sex, female (%)	49.4	52.5
Race (%)
White	81.7	81.0
Asian	11.5	14.6
Other	0.075	0.046
Duration of diabetes (years)	12.2 ± 6.6	12.2 ± 6.8
Albumin in serum (g/L)	44.6 ± 2.7	44.7 ± 2.5
Calculated creatinine clearance, serum (mL/min)	121.6 ± 41.6	127.7 ± 50.8
Basal insulin dose at week 1 (units)	35.6 ± 23.7	35.7 ± 23.4
Basal insulin dose at week 7 (units)	49.8 ± 25.7	50.6 ± 24.6
FBG at week 1 (mg/dL)	140.4 ± 39.6	140.4 ± 39.6
FBG at week 7 (mg/dL)	111.6 ± 25.2	115.2 ± 27.0
HbA1c at screening (mmol/mol)	65 ± 6.5	65 ± 7.6
HbA1c at screening (%)	8.1 ± 0.6	8.1 ± 0.7
Total change in basal insulin dose from weeks 2 to 7 (units)	10.5 ± 10.0	11.1 ± 10.7

Abbreviation: FBG, fasting blood glucose.

Five features were selected based on the forward selection: FBG at week 2, total change in basal insulin dose, trial site, hemoglobin level in blood at screening, and alkaline phosphatase level in serum at screening. Training the machine learning-enabled model resulted in the following model equation:

\begin{array}{l} Δ {FBG}_{wk2-7} = 1.22 - 12.45 . {FBG}_{wk 2} \\ - 1.86 . Δ {basal insulin dose}_{wk2-7} \\ + 0.33 trial site + 0.86 hemoglobin level + 0.92 \\ alkaline phosphatase level + 6.86 {FBG}_{wk 2} \times Δ {insulin}_{wk2-7} \end{array}

Table 3 shows the result of validation and test of the model and Figure 5 shows the observed versus predicted change in FBG obtained during validation and test of the model, revealing a similar trend along the diagonal line (y=x). Thus, no apparent signs of overfitting..

Table 3.

Results of Validation and Test of the Development Model on Training and Test Data Sets, Respectively.

Metric	Validation, mean [95% CI]	Test
RAE	0.64 [0.61-0.67]	0.67
r	0.77* [.73-.80]	0.74*
R ²	0.59 [.53-.64]	0.55

Abbreviations: CI, confidence interval; REA, relative absolute error.

P < .001.

Figure 5.

Observed versus predicted change in fasting blood glucose obtained during validation and test of the present model.

Discussion

The present study demonstrated that it is feasible to predict the effect of a change in basal insulin dose on the FBG in people with type 2 diabetes using a machine learning-enabled linear model.

Other than the study by Chen et al,¹⁷ utilization of a machine learning approach has to the best of our knowledge not been investigated in regard to basal insulin titration in people with type 2 diabetes. The results both from Chen et al¹⁷ and the present study elucidate the potential of models developed using a machine learning approach to optimize insulin titration in type 2 diabetes. Of the five features selected as input to the present model, two aligned with Features selected by models in Chen et al.¹⁷: FBG level when the adjustment of insulin dose was made and size of the dose adjustment. The remaining three features selected by the present model were not available in the study by Chen et al.¹⁷ The selection of trial site as a feature could indicate some variations in the titration across different trial sites. A model including a trial site would be difficult to implement in a real-world setting as it would require the model to be retrained and when implemented at a new site or calibrated to each site. Investigating why trial site was selected and considering the best way to include that information for the purpose of future implementation is thus relevant to investigate and consider. At first glance, the two selected blood samples’ hemoglobin level, and alkaline phosphatase level may not seem related to glycemic control. However, both have been found to be correlated to diabetes-relevant factors which could explain the selection of these two blood samples. Hemoglobin level has been found to be positively correlated with insulin resistance in different types of tissue in people with diabetes, while abnormal hemoglobin levels mayindicate kidney or renal disfunction.^19-21 Elevated alkaline phosphatase levels have been indicated as a marker of nonalcoholic fatty liver disease, which also affects insulin resistance and hepatic glucose production.^22-24 Therefore, the two blood samples may give the model valuable insights into the participants’ disease complexity. This elucidates the need to include blood samples and information that directly reflect participants’ disease complexity such as c-peptide and whether diabetic complications are present in future work since such information seems to improve the model’s performance.

Machine learning approaches have been more widely applied to optimizing insulin doses within the area of type 1 diabetes where the focus has been optimization of bolus insulin doses,²⁵ insulin infusion rates,^26,27 or short-term prediction of glucose levels aimed as input to automated insulin delivery to optimize dosing advise.^28-30 Results from these studies further support the evidence that machine learning approaches can be utilized in optimizing insulin treatment. However, results from type 1 diabetes are not directly transferableto people with type 2 diabetes. Therefore, further research in utilizing machine learning to optimize basal insulin titration of people with type 2 diabetes is needed to explore the potential.

The present model’s performance showed a lack of ability to correctly predict positive changes in FBG following a change in insulin dose. This is evident based on inspection of Figure 3 where it is seen that observed positive changes in FBG are mainly predicted as changes of 0 mg/dL or less by the model. This could point to the fact that the model may not be able to capture patterns in the data resulting in a positive change in FBG. This could result from the lack of information on disease progression in the individual participant (e.g., pancreatic insulin production and insulin resistance). Despite the efforts to include such indicators as potential features through extracting knowledge about changes in insulin doses and FBG from week 1 to 2, this information might not give sufficient insights into pancreatic insulin production or insulin resistance. Moreover, the inability to correctly predict positive changes in FBG could also result from too narrow a distribution of changes in FBG represented in the data since most participants experience a negative change in FBG during the titration period. Finally, factors such as BMI, change in lifestyle, stress level, sickness, and so on that are known to also impact glycemic control were not included in this analysis due to the lack of availability which may impact the model’s performance negatively.^31,32 However, this information is often difficult to obtain and challenging to implement, as it is rarely routinely collected. Thus, increasing the number of participants in the data set upon which the model is developed and including information about pancreatic insulin production, insulin sensitivity, and other factors known to affect glycemic control such as BMI, change in lifestyle, and etc. could improve the performance and generalizability of the model.

The study was limited since both FBG and insulin doses were self-reported and, therefore, prone to issues with validity.^33,34 Especially insulin doses taken by people with type 2 diabetes are likely noise-prone.^35,36 As non-adherence is a common problem within type 2 diabetes, there are most likely mismatches between the self-reported doses and the doses taken by the participants.³³ This might have blurred the effect of changes in insulin dose on FBG since some of the reported doses might not have been what the participants took. Therefore, developing the model using insulin data collected by smart insulin pens to give more accurate information on the insulin doses affecting the FBG could be of interest. In the same line of thought, using continuous glucose monitors to capture FBG could also improve the validity of the FBG data.

A model able to predict the effect of a change in basal insulin dose on FBG in people with type 2 diabetes could be used as a decision-support tool for clinicians when titrating insulin. Thus, it is visible to clinicians which patients react promptly to changes in insulin dose and which patients require more extensive changes to gain the wanted change in FBG. Previous studies have mainly investigated the use of different approaches (e.g., digital devices, telemonitoring, etc.) for adjusting basal insulin in individuals with type 2 diabetes by recommending a specific insulin dose.^11-14 By flipping the titration algorithm and predicting the expected impact on FBG rather than the insulin dose, clinicians gain valuable knowledge on how a particular individual will respond to changes in insulin. This personalized insight enables tailored treatment adjustments to meet individualized FBG targets. However, practical factors must be considered before implementation of such a model. More refinement and thorough testing are necessary before implementation in a real-world setting as the model, at the current stage, cannot account for e.g., the use of different basal insulin with varying action profiles and concomitant medication. In a real-world setting, patients will receive various combinations of medications and different types of basal insulin which effect glycemic control differently. Thus, a titration model must also consider such factors if it is to be used in a real-world setting. This work is not at a stage of maturity where it is appropriate to implement. However, it elucidates the potential of using a machine learning-enabled model to predict the effect on FBG when adjusting basal insulin to overcome some of the clinical inertia challenges.

Conclusion

It was feasible to develop a machine learning-enabled linear model using self-reported FBG, self-reported insulin doses, and blood samples taken at screening to predict the five-week change in FBG following a change in basal insulin dose in people with type 2 diabetes. Such a model may have the potential to facilitate optimized titration potentially reducing non-adherence and treatment fatigue leading to better glycemic control. However, prior to implementation in the real world, the model must undergo further refinement and rigorous testing. This is imperative to ensure that the model can effectively be used to manage glycemic control in people with type 2 diabetes accounting for diverse factors such as the distinct action profiles of basal insulin, concomitant medication, stress levels, and so on.

Footnotes

Acknowledgements

The authors thank Novo Nordisk A/S for providing access to the data used to develop the machine learning model.

Abbreviations

FBG, fasting blood glucose; RAE, relative absolute error; r, Pearson correlation coefficient; R², coefficient of determination.

Declaration of Conflicting Interests

The author(s) declared the following potential conflicts of interest with respect to the research, authorship, and/or publication of this article: Author P.V. is the head of research at the Steno Diabetes Center North Denmark, funded by the Novo Nordisk Foundation. Author M.H.J holds Novo Nordisk shares.

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

ORCID iDs

Camilla Heisel Nyholm Thomsen

Thomas Kronborg

Stine Hangaard

Morten Hasselstrøm Jensen

References

Khunti

Gomes

Pocock

, et al. Therapeutic inertia in the treatment of hyperglycaemia in patients with type 2 diabetes: a systematic review. Diabetes Obes Metab. 2018;20(2):427-437. doi:10.1111/dom.13088.

Reach

Pechtner

Gentilella

Corcos

Ceriello

. Clinical inertia and its impact on treatment intensification in people with type 2 diabetes mellitus. Diabetes Metab. 2017;43(6):501-511. doi:10.1016/j.diabet.2017.06.003.

Chun

Strong

Urquhart

. Insulin initiation and titration in patients with type 2 diabetes. Diabetes Spectr. 2019;32:104-111. doi:10.2337/ds18-0005.

Mocarski

Yeaw

Divino

, et al. Slow titration and delayed intensification of basal insulin among patients with type 2 diabetes. J Manag Care Spec Pharm. 2018;24(4):390-400. doi:10.18553/jmcp.2017.17218.

Stratton

Adler

Neil

HAW

, et al. Association of glycaemia with macrovascular and microvascular complications of type 2 diabetes (UKPDS 35): prospective observational study. BMJ. 2000;321:405-412. doi:10.1136/bmj.321.7258.405.

Khunti

Giorgino

Berard

Mauricio

Harris

. The importance of the initial period of basal insulin titration in people with diabetes. Diabetes Obes Metab. 2020;22(5):722-733. doi:10.1111/dom.13946.

Petersen

Shulman

. Mechanisms of insulin action and insulin resistance. Physiol Rev. 2018;98:2133-2223. doi:10.1152/physrev.00063.2017.

Krishnamoorthy

Boiroux

Aradóttir

Engell

Jørgensen

. A model-free approach to automatic dose guidance in long acting insulin treatment of type 2 diabetes. IEEE Control Systems Letters. 2021;5:2030-2035. doi:10.1109/LCSYS.2020.3047493.

Forst

Choudhary

Schneider

Linetzky

Pozzilli

. A practical approach to the clinical challenges in initiation of basal insulin therapy in people with type 2 diabetes. Diabetes Metab Res Rev. 2021;37(6):e3418. doi:10.1002/dmrr.3418.

10.

Thomsen

CHN

Hangaard

Kronborg

Vestergaard

Hejlesen

Jensen

. Time for using machine learning for dose guidance in titration of people with type 2 diabetes? A systematic review of basal insulin dose guidance. J Diabetes Sci Technol. 2022:19322968221145964. doi:10.1177/19322968221145964.

11.

Bajaj

Venn

Aronson

. Randomized trial of long-acting insulin glargine titration web tool (LTHome) versus enhanced usual therapy of glargine titration (INNOVATE Trial). Diabetes Technol Ther. 2016;18(10):610-615. doi:10.1089/dia.2016.0182.

12.

Tamez-Pérez

Deceased Author Cantú-Santos

Gutierrez-González

González-Facio

Romero-Ibarguengoitia

. Effect of digital-tool-supported basal insulin titration algorithm in reaching glycemic control in patients with type 2 diabetes in Mexico. J Diabetes Sci Technol. 2021:19322968211034532. doi:10.1177/19322968211034533.

13.

Davies

Bain

Charpentier

, et al. A randomized controlled, treat-to-target study evaluating the efficacy and safety of insulin glargine 300 U/mL (Gla-300) administered using either device-supported or routine titration in people with type 2 diabetes. J Diabetes Sci Technol. 2019;13(5):881-889. doi:10.1177/1932296818821706.

14.

Nevoret

Gervaise

Delemer

, et al. The effectiveness of an app (insulia) in recommending basal insulin doses for french patients with type 2 diabetes mellitus: longitudinal observational study. JMIR Diabetes. 2023;8:e44277. doi:10.2196/44277.

15.

Aradóttir

Bengtsson

Jensen

, et al. Feasibility of a new approach to initiate insulin in type 2 diabetes. J Diabetes Sci Technol. 2021;15(2):339-345. doi:10.1177/1932296819900240.

16.

Aradóttir

Boiroux

Bengtsson

, et al. Model predictive control for dose guidance in long acting insulin treatment of type 2 diabetes. IFAC J Syst Contr. 2019;9:100067. doi:10.1016/j.ifacsc.2019.100067.

17.

Chen

Zhao

, et al. Real-time artificial intelligence assisted insulin dosage titration system for glucose control in type 2 diabetic patients: a proof of concept study. Curr Med. 2023;2:2. doi:10.1007/s44194-023-00020-7.

18.

Bowering

Case

Harvey

, et al. Faster aspart versus insulin aspart as part of a basal-bolus regimen in inadequately controlled type 2 diabetes: the onset 2 trial. Diabetes Care. 2017;40(7):951-957. doi:10.2337/dc16-1770.

19.

Chen

S-H

C-Z

Lin

J-D

, et al. The relationships between hemoglobin and diabetes factors (insulin resistance, glucose effectiveness, first- and second-phase insulin secretion) in Old Chinese. International Journal of Gerontology. 2018;12:303-309. doi:10.1016/j.ijge.2018.05.007.

20.

Song

H-Y

Wei

C-M

Zhou

W-X

H-F

Wan

Q-J

. Association between admission hemoglobin level and prognosis in patients with type 2 diabetes mellitus. World J Diabetes. 2021;12:1917-1927. doi:10.4239/wjd.v12.i11.1917.

21.

Ueda

Ishimura

Shoji

, et al. Factors affecting progression of renal failure in patients with type 2 diabetes. Diabetes Care. 2003;26(5):1530-1534. doi:10.2337/diacare.26.5.1530.

22.

Dharmalingam

Yamasandhi

. Nonalcoholic fatty liver disease and type 2 diabetes mellitus. Indian J Endocrinol Metab. 2018;22:421-428. doi:10.4103/ijem.IJEM_585_17.

23.

Kocabay

Telci

Tutuncu

, et al. Alkaline phosphatase: can it be considered as an indicator of liver fibrosis in non-alcoholic steatohepatitis with type 2 diabetes. Bratisl Lek Listy. 2011;112(11):626-629.

24.

Harris

. Elevated liver function tests in type 2 diabetes. Clinical Diabetes. 2005;23:115-119. doi:10.2337/diaclin.23.3.115.

25.

Cappon

Vettoretti

Marturano

Facchinetti

Sparacino

. A Neural-network-based approach to personalize insulin bolus calculation using continuous glucose monitoring. J Diabetes Sci Technol. 2018;12(2):265-272. doi:10.1177/1932296818759558.

26.

Daskalaki

Diem

Mougiakakou

. Model-free machine learning in biomedicine: feasibility study in type 1 diabetes. PLOS ONE. 2016;11(7):e0158722. doi:10.1371/journal.pone.0158722.

27.

Sun

Jankovic

Budzinski

, et al. A dual mode adaptive basal-bolus advisor based on reinforcement learning. IEEE J Biomed Health Inform. 2019;23(6):2633-2641. doi:10.1109/JBHI.2018.2887067.

28.

van Doorn

WPTM

Foreman

Schaper

, et al. Machine learning-based glucose prediction with use of continuous glucose and physical activity monitoring data: the Maastricht Study. PLOS ONE. 2021;16(6):e0253125. doi:10.1371/journal.pone.0253125.

29.

Martínez-Delgado

Munoz-Organero

Queipo-Alvarez

. Using absorption models for insulin and carbohydrates and deep leaning to improve glucose level predictions. Sensors (Basel). 2021;21:5273. doi:10.3390/s21165273.

30.

Hidalgo

Colmenar

Kronberger

Winkler

Garnica

Lanchares

. Data Based Prediction of Blood Glucose Concentrations Using Evolutionary Methods. J Med Syst. 2017;41:142. doi:10.1007/s10916-017-0788-2.

31.

Chatterjee

Khunti

Davies

. Type 2 diabetes. Lancet. 2017;389:2239-2251. doi:10.1016/S0140-6736(17)30058-2.

32.

Haghighatpanah

Nejad

ASM

Haghighatpanah

Thunga

Mallayasamy

. Factors that correlate with poor glycemic control in type 2 diabetes mellitus patients with complications. Osong Public Health Res Perspect. 2018;9(4):167-174. doi:10.24171/j.phrp.2018.9.4.05.

33.

Edwards

Wang

, et al. Use of connected pen as a diagnostic tool to evaluate missed bolus dosing behavior in people with type 1 and type 2 diabetes. Diabetes Technol Ther. 2022;24(1):61-66. doi:10.1089/dia.2021.0239.

34.

Blackburn

Swidrovich

Lemstra

. Non-adherence in type 2 diabetes: practical considerations for interpreting the literature. Patient Prefer Adherence. 2013;7:183-189. doi:10.2147/PPA.S30613.

35.

Osborn

Gonzalez

. Measuring insulin adherence among adults with type 2 diabetes. J Behav Med. 2016;39:633-641. doi:10.1007/s10865-016-9741-y.

36.

Stolpe

Kroes

Webb

Wisniewski

. A systematic review of insulin adherence measures in patients with diabetes. J Manag Care Spec Pharm. 2016;22(11):1224-1246. doi:10.18553/jmcp.2016.22.11.1224.