Physiology-Guided Blood Glucose Predictive Model Using Minimal

Blood Glucose Dynamics

Sarala Ghimire

, Turgay Celik

, Martin Gerdes

and Christian W. Omlin

Department of Information and Communication Technologies, Centre for e-Health, University of Agder,

Grimstad, Norway

Department of Information and Communication Technologies, Centre for Artificial Intelligence Research (CAIR),

University of Agder, Grimstad, Norway

Keywords: Blood Glucose, Data-Driven, Hybrid, Mathematical Model, Prediction, Physiology.

Abstract: Modelling blood glucose and insulin dynamics using mathematical equations requires a deep understanding

of individual physiology and relying on numerous predefined parameters necessitating extensive clinical and

personal data, making direct use of these models for blood glucose prediction computationally intensive and

inaccurate. Though data-driven models are more efficient and require no individual physiology, they produce

predictions that are inconsistent with known glucose-insulin interactions. Thus, this study aims to investigate

the potentiality of physiological models integrated with data-driven approach for predicting blood glucose

level. It intends to extract simple physiological dynamics of blood glucose kinetics and incorporate them into

a data-driven model, with less reliance on detailed individual data. The result demonstrated that the model

integrating physiological modelling of insulin and meal absorption significantly improved the performance

particularly in larger window size that enabled the model to better capture longer-term trends and temporal

dependencies.

1 INTRODUCTION

Type 1 Diabetes is a chronic metabolic condition

characterized by disrupted blood glucose homeostasis

due to the lack of endogenous insulin production

(Georga et al., 2012). As a result, subjects with type

1 diabetes must rely on exogenous insulin to

compensate for this deficiency (Oviedo et al., 2017).

To maintain blood glucose levels within the desired

range and to reduce the risk of complications such as

hypoglycaemia or hyperglycaemia, one must follow

a dietary guideline, administer insulin in right

amount, and proper exercise regimen (Zhu et al.,

2022). One promising technology for managing

blood glucose is continuous glucose monitoring

(CGM), which tracks glucose levels in real time and

is often integrated into automated insulin delivery

systems (Della Cioppa et al., 2023). These automated

systems, known as closed-loop control systems or

artificial pancreas, use predictive models to forecast

future blood glucose levels and adjust insulin delivery

accordingly.

Thus, accurate prediction of blood glucose levels

is vital for management of blood glucose. The use of

mathematical equations that represent the body’s

blood glucose dynamics to make accurate predictions

require detailed knowledge of individual physiology,

such as a person’s glucose absorption rates, insulin

sensitivity, and how their body handles glucose

during activities like eating or fasting (Oviedo et al.,

2017) (Woldaregay et al., 2019). They also depend on

a large number of pre-set parameters, requiring

extensive clinical and personal data about the

individual. On the other hand, data-driven models

focus more on self-monitored historical data such as

past glucose levels, insulin doses, and food intake and

not require as much detailed physiological

information (Woldaregay et al., 2019). While this

approach is easier to implement as it requires less

knowledge about the individual’s specific

physiology, it has limitations. These models are often

less consistent with actual physiological processes

and are of a black box nature. Also, many tend to lose

accuracy over longer timeframes (Ghimire et al.,

2024).

Previous study (Pawar et al., 2021) suggests that

combining simplified physics-based models with

neural network architecture can lead to improved

984

Ghimire, S., Celik, T., Gerdes, M. and Omlin, C. W.

Physiology-Guided Blood Glucose Predictive Model Using Minimal Blood Glucose Dynamics.

DOI: 10.5220/0013395200003911

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 18th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2025) - Volume 2: HEALTHINF, pages 984-990

ISBN: 978-989-758-731-3; ISSN: 2184-4305

predictive accuracy and reduced uncertainty. In line

with this principle, this paper investigates a hybrid

modelling paradigm. The hybrid model integrates the

subcutaneous insulin absorption kinetics and the meal

absorption dynamics within the predictive model for

blood glucose levels, as opposed to relying on

complex and data intensive full mathematical models.

It combines the strengths of physiological models

(which reflect real glucose dynamics) with machine

learning (ML) techniques (which can handle large

amounts of historical data). The primary aim is to

obtain more accurate blood glucose predictions that

remain consistent with physiological principles and

can be interpreted. Additionally, it intends to provide

a better understanding of its potential strength for

future expansion.

The remainder of this article is organized as

follows. The existing works are discussed in section

2. The description of the proposed method is

presented in section 3. Section 4 covers the data sets

used, data processing, and experiments setup,

including the evaluation metrics. In Section 5,

experimental results and analysis are provided.

Finally, the article is concluded in Section 6.

2 RELATED WORKS

Several studies (Bertachi et al., 2018; Contreras et al.,

2018; Contreras et al., 2017; Erdős et al., 2023; Georga

et al., 2012; Mougiakakou et al., 2006; Sun et al., 2020)

have developed models, integrating mathematical

models of blood glucose dynamics with data-driven

models. Study (Georga et al., 2012) enhanced

performance of the Support Vector Regression (SVR)

model by incorporating additional features such as the

rate of glucose appearance from meal intake (Ra),

plasma insulin levels (Ip), and cumulative glucose

appearance over a specific period (SRa). These

features, obtained from meal and insulin models of

blood glucose dynamics, serve as inputs, provide a

straightforward way to incorporate domain knowledge

into learning models. In a similar approach, study

(Mougiakakou et al., 2006) included plasma insulin

and glucose appearance rate from meals as input

features through insulin kinetics and meal models,

respectively. Additional features such as insulin on

board i.e., the remaining active insulin, carbohydrates

on board i.e., remaining carbs yet to be converted to

glucose, glucose appearance rate from carbs, rate of

glucose appearance in the blood from guts, and activity

on board were employed in studies (Bertachi et al.,

2018; Contreras et al., 2018; Contreras et al., 2017; Sun

et al., 2020) to refine ML models. Additionally, study

(Erdős et al., 2023) proposed a hybrid method that

sequentially combines predictions from both

mathematical and ML models, where the ML model

predicts residuals based on phenotypic features, which

are then subtracted from the predictions made by

mathematical model. The results demonstrated that the

personalized physiological models consistently

outperformed the data-driven and the hybrid model

approaches. Personalized physiological models may

inherently capture individualized critical processes and

features that data-driven models struggle to

individualize or interpret effectively. Although these

studies have shown improvements in performance

through the integration of various physiological

knowledge, they fail to analyse the potential of

integration across different contexts, a gap that is

addressed in this study.

3 PROPOSED METHOD

To explore the potential of the physiological inputs, a

straightforward model architecture was designed,

consisting of two Long short-term memory (LSTM)

layers with 64 units each, followed by a fully

connected layer with 64 neurons as the final layer, as

illustrated in Figure 1. LSTM networks are a

specialized type of recurrent neural networks (RNNs)

designed to effectively capture long-term temporal

dependencies in sequential data (Aliberti et al., 2019),

comprising memory cells, input, forget, and output

gates that dynamically manages information flow,

preserving the essential patterns and insights over

long sequences, enhancing their effectiveness for

time series prediction (Afsaneh et al., 2022).

Figure 1: System architecture of the proposed hybrid

model.

Physiology-Guided Blood Glucose Predictive Model Using Minimal Blood Glucose Dynamics

985

Different configurations of this model are used to

assess their potential performance. Predictive model

incorporating physiological inputs (insulin plasma (IP)

and carb on board (COB) including glucose (G)) is

labelled as PM, while the model without physiological

input is referred to as W-PM, that includes bolus (Bo),

basal (B), and carb (C) including G. The models are

further divided into input mode and output mode: in

input mode, the final layer receives no additional

inputs, while in output mode, an additional input

feature, IP, is included in the final layer.

When bolus insulin is administered, it doesn’t

instantly lower blood glucose; rather, it enters the

bloodstream, peaks after a certain time, and gradually

dissipates (Lehmann & Deutsch, 1992). Based on

similar physiology, the assumption is that if ML

model incorporates insulin dynamics that provides a

temporal representation of insulin’s effect, reflecting

how insulin concentration changes over time as is

seen in the body’s physiological process, it allows for

a more nuanced understanding of how blood glucose

will respond. For a simple bolus-only ML model that

considers only the dose and the timing of insulin

administration, however, is challenging to capture

how insulin gradually influences blood glucose.

Figure 2 depicts the bolus administered dose (Marling

& Bunescu, 2020) and its timing curve along with the

insulin dynamics following the insulin dose, where

administered bolus is observed at a specific time step,

while the insulin dynamics profile exhibits a gradual

decay of insulin over time. It illustrates how insulin

levels rise and fall, giving a clear indication of how

glucose levels might influence. Additionally, giving a

second bolus dose before the previous dose has fully

dissipated results in cumulative insulin effects in the

bloodstream. By modelling plasma insulin, it is

possible to track these overlapping, enabling precise

predictions with increased insulin activity.

Figure 2: Insulin decay curve given by the physiological

insulin model.

Thus, based on this physiology, the proposed

model architecture incorporates the insulin dynamics

to provide a temporal representation of insulin’s

effect on how insulin concentration changes over

time, where we utilized the Lehmann model

(Lehmann & Deutsch, 1992) that aligns with the

physiological process of insulin absorption dynamics

as given in equation (1):





































 



 (1)

where,  is plasma insulin concentration, 



is the

first-order rate constant of insulin elimination, and 



is the volume of insulin distribution,  is the time

elapsed from the injection, 





is the time at which

50% of the dose, D, has been absorbed and  is a

preparation-specific parameter defining the insulin

absorption pattern of the different types of insulin.

The insulin dynamics that give plasma insulin is

based on rapid-acting insulin, which is known to have

an immediate effect on reducing blood glucose levels

with fast absorption and action. This makes it

particularly relevant for scenarios where quick

glucose adjustments are needed after a meal.

In addition to this, instead of using the discrete

carbohydrate data into the model, the carbohydrate on

board based on the meal intake is estimated to convert

it to a time action profile of absorption, similar to the

insulin dynamics as given by (2) (Dave et al., 2021).

It estimates the amount of carbs not yet appeared in

blood glucose. The curves in Figures 3 show how the

carbs decay over time.





 







 



 





 (2)

Figure 3: The carbohydrate decay curve is given by the

physiological meal model.

HEALTHINF 2025 - 18th International Conference on Health Informatics

986

Here, 



denote the time at which  amount of

carbohydrates taken. 



is the carbohydrate

absorption rate, which is 0.5g/min after initial delay

(



) of 15min and  is the current time. These time

action profiles of insulin and carbs are fed into the PM

model as additional inputs along with glucose in

initial or final layer as shown in Figure 1. The plasma

insulin in the output mode is included in the final

layer of the model as shown in Figure 1, to act as a

corrective mechanism for the model to correct the

prediction at the last minute. This architecture implies

that the model first makes the prediction using input

features and then applies correction based on the

amount of insulin present or absorbed.

4 EXPERIMENTS

4.1 Datasets

The models were trained and tested using OhioT1DM

(Marling & Bunescu, 2020) dataset that encompasses

eight weeks of data from twelve patients with type-I

diabetes. It involves seven males and five females

within the age range of 20 to 80 years. The dataset

contains blood glucose measurements recorded every

five minutes, comprising total 166532 samples, using

Medtronic Enlite CGM sensors, along with bolus and

basal insulin administered via Medtronic 530G

insulin pumps at regular intervals. Also, the dataset is

supplemented by other daily events such as meal and

exercise info reported by the patients via a

smartphone app or fitness band. The proposed study

explores the data on blood glucose levels, insulin and

carbohydrate intake. The OhioT1DM dataset was

originally provided with separate training and testing

data for each of the 12 patients, which were directly

used for model training and evaluation. The training

data was further divided into an 80% training subset

and a 20% validation subset. The model, trained on

the training subset, was evaluated on the test set,

which consisted of new, unseen data.

4.2 Experimental Setup

To address missing data and standardize it,

resampling and normalization were applied. The

dataset was resampled to align all signals to a uniform

5-minute sampling grid. Data with consecutive gaps

exceeding 30 minutes were excluded to prevent

inaccuracies in prediction trajectories. Missing values

were imputed using linear interpolation; however, no

imputation was performed on the test set.

Standardized normalization was applied to scale the

data to a mean of zero and a standard deviation of one.

Additionally, to investigate the relationship

between blood glucose levels and their prior values

and to find the right window size, the autocorrelation

function (ACF) (Semmlow, 2012) was utilized. As

ACF describes how well a signal correlates with

adjacent part of itself, the plot of the ACF on blood

glucose levels revealed a strong correlation between

the current glucose levels and glucose levels from the

previous day as shown in Figure 4. Based on this

analysis, a sliding window of 24 hours was used

instead of a 2-hour historical data window to predict

blood glucose levels 30 minutes in advance, as

opposed to the literatures (Cui et al., 2021; Ghimire

et al., 2024; Oviedo et al., 2017). The model was

trained for 100 epochs with 10 repetitive runs using

Huber loss (Tong, 2023) as the cost function. The

adaptive moment estimation (Adam) optimizer was

employed to minimize the loss function.

Optimization parameters included a learning rate

decay of 0.1, a decay patience of 10, and early

stopping patience of 30.

Figure 4: Auto-correlation function of blood glucose with

its previous values of 24hrs.

4.3 Evaluation Metrics

Model performance was assessed using the root mean

square error (RMSE), a standard metric that

quantifies the similarity between predicted and actual

blood glucose levels, defined as follows:

 













 







 







(3)

Here,  represents the number of test samples, 

denotes the actual blood glucose levels, 



is the

predicted glucose level, and PH stands for the

Physiology-Guided Blood Glucose Predictive Model Using Minimal Blood Glucose Dynamics

987

prediction horizon—the time frame for which blood

glucose levels are forecasted. A lower RMSE value

indicates better model performance in glucose

prediction. For validation, we employed a 30-minute

prediction horizon (PH), implies that the model

predicts blood glucose levels 30 minutes ahead.

5 EXPERIMENTAL RESULTS

AND DISCUSSION

Table 1 presents the results for the proposed model

with a 30-minute PH across various input feature

combinations, both with and without a physiological

model. The results indicate that the PM model

attained a lower RMSE than the W-PM model,

demonstrating the significance of the use of

physiological inputs. While adding insulin plasma in

the final layer did not reduce RMSE in either

approach, it enhanced the prediction certainty

compared to models without plasma insulin, as is also

illustrated in Figures 5 and 6 with highlights. No

significant difference in RMSE was observed

between input and output modes within the PM

model. Overall, incorporating physiological

dynamics led to a performance improvement of over

10%, and adding physiological dynamics in the final

layer contributed further to prediction certainty.

Table 1: Summary of the RMSE(Std) result for different

combinations of input features.

Predictive

Models

Input Mode (in features)

RMSE(Std)

W-PM

(24hrs)

Output (G, BO, C, B) (IP)

20.154 (1.32)

Input (G, BO, C, B)

20.142 (2.08)

PM (24hrs)

Output (G, CO) (IP)

17.606 (0.13)

Input (G, CO, IP)

17.526 (0.159)

W-PM (2hrs)

Input (G, BO, C, B)

19.235 (0.135)

PM (2hrs)

Input (G, CO, IP)

19.437 (0.233)

PM: Physiological Model, W-PM: Without Physiological

Model, G: Glucose level, BO: Bolus insulin, C:

Carbohydrate, B: Basal insulin, CO: Carbohydrate on

Board and IP: Plasma Insulin

From the results, it is evident that using

physiological models offers no significant

improvement when a shorter 2-hour window is

employed. The potential reason for this is the

continuous data sampled at 5-minute intervals using

physiological modelling for insulin and meal,

successfully created smoother transitions between

data points, allowing the models to see gradual

changes rather than abrupt shifts. With a 24-hour

window, this continuous representation enabled the

model to better capture longer-term trends and

temporal dependencies. In contrast, for the discrete

insulin and carbs data, model struggled to identify

patterns over longer windows due to the limited

number of data points within each window, which

constrained the model’s ability to learn these trends

effectively. However, with a 2-hour window, the

difference between the continuous and discrete is

minimal and there is enough overlap between the

neighbouring data points within the window, making

the abrupt changes in the discrete insulin and carbs

data less impactful on model performance.

Figure 5: Uncertainty prediction plot of input mode of W-

PM model.

Figure 6: Uncertainty prediction plot of output mode of PM

model.

To further explain the output of the predictive

model, Shapley additive explanation (SHAP), a

model agnostic explainability approach is utilized.

SHAP plots as shown in Figures 7 and 8 for both

models highlight that blood glucose is the most

influential feature in driving the model's decisions,

with carbs (or COB in the case of the PM model)

HEALTHINF 2025 - 18th International Conference on Health Informatics

988

being the second most important. Bolus and basal

have a minimal to no impact on the output

predictions, though bolus shows a slightly higher

influence. A higher SHAP value for COB compared

to carbs indicates that the PM model assigns

significant importance to COB in predicting future

blood glucose levels. This also suggests that COB has

a notable positive effect on predictions, aligning with

the physiological process where an increase in COB

leads to an increase in future blood glucose levels,

demonstrating the prediction to be consistent with the

physiology. This observation supports the superior

performance of the PM model, which benefits from

the inclusion of physiological inputs compared to the

W-PM model.

Figure 7: Summary plot of SHAP analysis of PM model.

Figure 8: Summary plot of SHAP analysis of W-PM model.

6 LIMITATIONS

Despite potential results, this study has some

limitations. The study modelled the physiological

process using only bolus insulin, excluding basal

insulin, which plays a crucial role in blood glucose

regulation. Also, as the study focused on a simplified

physiological model, incorporating more complex

physiological processes could provide deeper insights

and improve prediction accuracy. In addition, the

SHAP analysis highlighted which input features

contribute most to the model’s predictions, revealing

associations identified by the model, but cannot

provide the underlying causal mechanisms.

Therefore, for drawing causal conclusions additional

analysis such as counterfactual analysis can

potentially be beneficial.

7 CONCLUSIONS

This study introduced a straightforward integration of

a physiological model with a data-driven approach for

predicting blood glucose levels. The findings

demonstrated that this hybrid model, incorporating a

simple physiological model, has the potential to

enhance predictive performance, also increase

certainty when integrated into the final layer.

Moreover, the analysis highlights that incorporating

the physiological process of gradual change

demonstrates its significance with larger window

sizes. This inclusion notably improved model

performance by preserving temporal dynamics that

would otherwise be lost in discrete data lacking a

physiological basis. Additionally, blood glucose is

identified as the most influential contributor for

prediction output, followed by COB, demonstrating

the importance of physiological dynamics.

Given the study's demonstrated potential for

physiological integration, future work could focus on

incorporating complex physiological modelling into

the model to enhance its representation of

physiological processes, predictive accuracy and

explainability.

REFERENCES

Afsaneh, E., Sharifdini, A., Ghazzaghi, H., & Ghobadi, M.

Z. (2022). Recent applications of machine learning and

deep learning models in the prediction, diagnosis, and

management of diabetes: a comprehensive review.

Diabetology & Metabolic Syndrome, 14(1), 1-39.

Aliberti, A., Pupillo, I., Terna, S., Macii, E., Di Cataldo, S.,

Patti, E., & Acquaviva, A. (2019). A multi-patient data-

driven approach to blood glucose prediction. IEEE

ACCESS, 7, 69311-69325.

Bertachi, A., Biagi, L., Contreras, I., Luo, N., & Vehí, J.

(2018). Prediction of Blood Glucose Levels And

Nocturnal Hypoglycemia Using Physiological Models

and Artificial Neural Networks. KDH@ IJCAI,

Contreras, I., Bertachi, A., Biagi, L., Vehí, J., & Oviedo, S.

(2018). Using Grammatical Evolution to Generate

Short-term Blood Glucose Prediction Models. KDH@

IJCAI,

Contreras, I., Oviedo, S., Vettoretti, M., Visentin, R., &

Vehí, J. (2017). Personalized blood glucose prediction:

A hybrid approach using grammatical evolution and

physiological models. PLoS ONE, 12(11), e0187754.

Cui, R., Hettiarachchi, C., Nolan, C. J., Daskalaki, E., &

Suominen, H. (2021). Personalised short-term glucose

prediction via recurrent self-attention network. 2021

IEEE 34th International Symposium on Computer-

Based Medical Systems (CBMS),

Physiology-Guided Blood Glucose Predictive Model Using Minimal Blood Glucose Dynamics

989

Dave, D., DeSalvo, D. J., Haridas, B., McKay, S., Shenoy,

A., Koh, C. J., Lawley, M., & Erraguntla, M. (2021).

Feature-based machine learning model for real-time

hypoglycemia prediction. Journal of Diabetes Science

and Technology, 15(4), 842-855.

Della Cioppa, A., De Falco, I., Koutny, T., Scafuri, U., Ubl,

M., & Tarantino, E. (2023). Reducing high-risk glucose

forecasting errors by evolving interpretable models for

type 1 diabetes. Applied soft computing, 134, 110012.

Erdős, B., van Sloun, B., Goossens, G. H., O’Donovan, S.

D., de Galan, B. E., van Greevenbroek, M. M.,

Stehouwer, C. D., Schram, M. T., Blaak, E. E., &

Adriaens, M. E. (2023). Quantifying postprandial

glucose responses using a hybrid modeling approach:

Combining mechanistic and data-driven models in The

Maastricht Study. PLoS ONE, 18(7), e0285820.

Georga, E. I., Protopappas, V. C., Ardigo, D., Marina, M.,

Zavaroni, I., Polyzos, D., & Fotiadis, D. I. (2012).

Multivariate prediction of subcutaneous glucose

concentration in type 1 diabetes patients based on

support vector regression. IEEE Journal of Biomedical

and Health Informatics, 17(1), 71-81.

Ghimire, S., Celik, T., Gerdes, M., & Omlin, C. W. (2024).

Deep learning for blood glucose level prediction: How

well do models generalize across different data sets?

PLoS ONE, 19(9), e0310801.

Lehmann, E., & Deutsch, T. (1992). A physiological model

of glucose-insulin interaction in type 1 diabetes

mellitus. Journal of biomedical engineering, 14(3),

235-242.

Marling, C., & Bunescu, R. (2020). The OhioT1DM dataset

for blood glucose level prediction: Update 2020. CEUR

workshop proceedings,

Mougiakakou, S. G., Prountzou, A., Iliopoulou, D., Nikita,

K. S., Vazeou, A., & Bartsocas, C. S. (2006). Neural

network based glucose-insulin metabolism models for

children with type 1 diabetes. 2006 International

Conference of the IEEE Engineering in Medicine and

Biology Society,

Oviedo, S., Vehí, J., Calm, R., & Armengol, J. (2017). A

review of personalized blood glucose prediction

strategies for T1DM patients. International journal for

numerical methods in biomedical engineering, 33(6),

e2833.

Pawar, S., San, O., Aksoylu, B., Rasheed, A., & Kvamsdal,

T. (2021). Physics guided machine learning using

simplified theories. Physics of Fluids, 33(1).

Semmlow, J. (2012). Autocorrelation Function. Retrieved

October from

https://www.sciencedirect.com/topics/engineering/auto

correlation-function

Sun, X., Rashid, M. M., Sevil, M., Hobbs, N., Brandt, R.,

Askari, M.-R., Shahidehpour, A., & Cinar, A. (2020).

Prediction of Blood Glucose Levels for People with

Type 1 Diabetes using Latent-Variable-based Model.

KDH@ ECAI, 20, 115-119.

Tong, H. (2023). Functional linear regression with Huber

loss. Journal of Complexity, 74, 101696.

Woldaregay, A. Z., Årsand, E., Walderhaug, S., Albers, D.,

Mamykina, L., Botsis, T., & Hartvigsen, G. (2019).

Data-driven modeling and prediction of blood glucose

dynamics: Machine learning applications in type 1

diabetes. Artificial intelligence in medicine, 98, 109-

134.

Zhu, T., Li, K., Herrero, P., & Georgiou, P. (2022).

Personalized blood glucose prediction for type 1

diabetes using evidential deep learning and meta-

learning. IEEE Transactions on Biomedical

Engineering, 70(1), 193-204.

HEALTHINF 2025 - 18th International Conference on Health Informatics

990