AI and Digital Twins Transforming Healthcare IoT

Vikas Sharma Kapil Sharma Akshi Kumar

2024 14th International Conference on Cloud Computing, Data Science & Engineering (Confluence) | 979-8-3503-4483-7/24/$31.00 ©2024 IEEE | DOI: 10.1109/CONFLUENCE60223.2024.10463366

vikassharma12387@gmail.com kapil@ieee.org akshi.kumar@gold.ac.uk

Department of Information Professor, Department of Associate Professor, Department
Technology Information Technology of Computing, Goldsmiths,
Delhi Technological University, Delhi Technological University, University of London, United
Delhi, India Delhi, India Kingdom

Abstract: In this age of digital and smart healthcare, insufficient data [4], [5]. IoT is a key development
cutting-edge technologies are being used to improve because it makes it easier to combine healthcare with
operations, patient well-being, life expectancy, and consumer gadgets and to gather and share data
healthcare costs. Digital Twins (DT) have the potential instantly. This integration helps people live longer and
to significantly change these new technologies. DTs healthier lives. ECG, blood pressure, heart rate, and
could revolutionise digital healthcare delivery with glucose levels can all be tracked using a combination
extraordinary creativity. A digital representation of a of app-based tracking and in-device sensors. These
physical asset that is always its digital twin due to real- devices upload data to the cloud for analysis and
time data processing. This paper proposes and builds a comparison in an effort to improve recording accuracy.
DT-based intelligent healthcare system that is aware of Alerts are sent out to healthcare providers when
its environment. This approach is a great advance for abnormalities are detected, dramatically improving
digital healthcare and could improve service delivery. both quality and efficiency [5]–[9].
Our most notable contribution is a machine learning-
based electrocardiogram (ECG) classifier model for The concept of a "digital twin" for a physical thing is a
cardiac diagnostics and early problem detection. Our top technical trend according to the IEEE Computer
cardiac models predict some situations with Society [10]. The digital twin technology in healthcare
exceptional accuracy when applied to different ways. marketing, monitoring, and administration will have
These findings highlight the potential for Digital Twins profound effects. Our breakthrough utilises the Internet
in healthcare to create intelligent, comprehensive, and of Things, data analytics, and AI to create digital twins
scalable Health-Systems that improve patient- of actual patients [11]. Healthcare is advanced,
physician communication. Our ECG classifier also sets professionals are given more agency, and practical
a precedent for using Artificial Intelligence (AI) and examples are used in this new paradigm. We developed
Machine Learning (ML) to continually monitor wide an ECG classifier for use in monitoring heart disease
range of human body data and identify outliers. ECG patients. This machine learning model, developed in
data processing has improved significantly using real time using ECG data, outperformed the
neural network-based algorithms over classic machine alternatives [12], [13]. The potential of digital twins in
learning methods. In conclusion, our work integrates healthcare is discussed, along with relevant studies,
digital twins with cutting-edge AI and machine framework architecture, system implementation, and
learning to revolutionise healthcare. Future healthcare results. Problems and potential solutions are discussed
will be predictive and improve lives. [14].

Keywords: Digital Twin, Healthcare, Internet of To improve the functionality and performance of
Things (IoT), Artificial Intelligence (AI) machine learning models, digital twins (DT) are deeply
integrated into them. A digital twin, as used in machine
I. Introduction learning, is a virtual version of a genuine item or
system integrated into models to enable a thorough and
Technology has had an impact on every sector of the dynamic knowledge of real-world things. A virtual
economy, including production, agriculture, education, counterpart comparable to the physical counterpart's
and healthcare. Smartphones, smart buildings, and behaviour, traits, and interactions must be created to
healthcare wearables are all revolutionised by the achieve this integration. Real-time information may be
Internet of Things' (IoT) [1] ability to easily connect used for training and optimising machine learning
them [2], [3]. Improvements in data accuracy, data models through continuous data interchange between
sharing, and pandemic preparedness have all resulted the digital twin and the physical system. Improved
from the widespread adoption of IoT devices in the accuracy, efficiency, and flexibility are fostered by the
healthcare industry. The Institute of Medicine in the models' capacity to adapt and develop in response to
United States estimates that
979-8-3503-4483-7/24/$31.00 400,000
©2024 IEEEindividuals lose 6 changes in the natural environment because of this
their use
Authorized licensed lives each
limited year dueINSTITUTE
to: NATIONAL to medical errors caused
OF TECHNOLOGY by symbiotic
SURATHKAL. Downloaded interaction.
on May 11,2025In fields where
at 04:57:07 real-time
UTC from insights
IEEE Xplore. Restrictions apply.
 and predictive skills are critical, such as predictive II. Related Work
maintenance, healthcare monitoring, and industrial
operations, using digital twins in machine learning is All of these studies demonstrate how DT research has
particularly beneficial. developed through time; the majority of articles
published today focus on theoretical framework and
The paper presents the introduction of digital twin (DT) models. The problem of the DT in smart healthcare
technology in healthcare with AI and different ML system has been discussed in a number of studies, but
models. The rest of the part is organized as follows. The no practical solutions or validation have yet been
literature review of DT for recent updates is offered. The most pertinent work has been condensed
represented in Section 2. Section 3 describes system in Table 1.
implementation with MIT-BIH dataset of healthcare
and shown result.

Table 1
LITERATURE REVIEW

S. No. Ref. Technology Limitations

Used
1. Benedictis et al. DT, 3D x Cannot be implemented in real-time scenarios
[15] Sensors, AI
2. Elayan et al. [16] DT, IoT x Linking real-time data
x Incorporating the models into the suggested
framework
3. Pilati et al. [17] DT, IoT x Can’t be implemented in real-time
x Adoption of the Technology
4. Esteban et al. [18] DT, GAN x Data organization especially images
x Discriminator and Generator training time
9. Erol et al. [19] DT, IoT, AI x Limited empirical data
x Lack of long-term implementation data
10. Suzuki et al. [20] Medical x Japanese adults only
imaging and x Small sample size
statistical x Results may not generalise to other groups
analysis
11. Golse et al. [21] DT x Small sample size
x Specific context
x Need for additional confirmation before wider
use
13. Laubenbacher et al. DT x Complexity of the immune system
[22] x Challenges inaccurate representation of
individual variability
14. Aubert et al. [23] DT, 3D x Limited generalizability
x A very limited number of cases studied
16. Grosman et al. [24] DT x Limited sample size
x Specific to Type 1 diabetes patients
x Real-world applicability may vary

III. System Implementation Laboratory collected these 48 half-hour snippets of

In this section, Convolutional Neural Networks, Multi- two-channel ambulatory ECG recordings from 47
layer Perceptron’s, Logistic Regression, Long-Short patients between 1975 and 1979. The five categories of
Term Memory Networks, and Support Vector the dataset are as follows: A for normal, B for
Classifiers were the techniques used. supraventricular, C for premature ventricular
contraction, D for a combined ventricular and normal
A. Dataset
rhythm, and E for an irregular heart rate.
This 2024 14th
research International
paper's datasetConference on Cloud
was derived from Computing,
the Data Science & Engineering (Confluence)
Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY SURATHKAL. Downloaded on May 11,2025 at 04:57:07 UTC from IEEE Xplore.
7
Restrictions apply.
MIT-BIH Arrhythmia Database. The BIH Arrhythmia
 B. Result and Discussion whereas Figure 2 shows the performance of the CNN
The best accuracy for the dataset required the model. Over multiple epochs, both accuracy and loss
construction of five distinct models; this part presents improved. This suggests that the models' forecasts are
the parameters, performance, collected results, and becoming less certain. To that end, the LSTM model
evaluation for each model that was used. Python, the with the lowest validation loss was preserved at epoch
Sklearn package, Tensorflow, and Keras were used in 5, and the CNN model with the lowest validation loss
the experiment. Pandas, Numpy, and Matplotlib were was saved at epoch 4. LSTM validation accuracy is
also utilised to aid in data preprocessing and results shown to be higher than CNN validation accuracy,
analysis. while CNN validation loss is found to be lower. They
are so close in performance that the difference between
Figure 1 shows the Loss and Accuracy performance of
them is nearly negligible.
the LSTM sequential model over training epochs,

(a) LSTM model Loss performance (b) LSTM model Accuracy performance

Fig 1: LSTM model Loss and Accuracy performance

(a) CNN model Loss performance (b) CNN model Accuracy performance

Fig 2: CNN model Loss and Accuracy performance

C. Evaluation (2) Confusion Matrix

Due to the asymmetry of the testing data, the
(1) Accuracy precision reported may not always be reliable.
As indicated in Equation 1, the proportion of The confusion matrices for each model can be
True Positives and True Negatives (TP and calculated to guarantee optimal performance.
TN) out of all data samples (TP, TN, FP, and The TP, FP, TN, and FN values were obtained
FN) that are correctly predicted can be after obtaining the Confusion Matrix. After
calculated. obtaining confusion matrix values, The
்௉ା்ே classification report for each model's overall
‫ ݕܿܽݎݑܿܿܣ‬ൌ  ்௉ାே௉ାி௉ାிே (1)
performance is displayed in Table 2.

8 2024 14th International Conference on Cloud Computing, Data Science & Engineering (Confluence)
Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY SURATHKAL. Downloaded on May 11,2025 at 04:57:07 UTC from IEEE Xplore. Restrictions apply.
 Table 2
PRECISION, RECALL, AND F1 SCORE VALUES FOR DIFFERENT ML MODELS

Classes/ LSTM CNN MLP SVC LR

Metrics
Precision

Precision

Precision

Precision

precision
F1-score

F1-score

F1-score

F1-score

F1-score
recall

recall

recall

recall

recall
A 0.9 0.9 0.9 0.9 0.9 0.9 0.9 0.9 0.9 0.9 0.7 0.8 0.9 0.6 0.7
9 7 8 9 7 7 9 6 8 7 4 4 7 5 8
B 0.6 0.8 0.7 0.5 0.8 0.6 0.5 0.8 0.6 0.2 0.6 0.3 0.1 0.6 0.2
0 3 8 6 8 2 2 8 5 9 5 7 4
C 0.9 0.9 0.9 0.9 0.9 0.9 0.8 0.9 0.8 0.3 0.7 0.4 0.2 0.7 0.4
2 4 2 2 5 2 7 3 3 9 7 9 2 2
D 0.6 0.8 0.7 0.5 0.8 0.6 0.6 0.8 0.7 0.1 0.8 0.1 0.0 0.8 0.1
0 7 0 7 8 7 3 1 8 7 8 8 5
E 0.9 0.9 0.9 0.9 0.9 0.9 0.9 0.9 0.9 0.8 0.9 0.8 0.7 0.9 0.8
9 9 9 7 9 7 5 8 6 1 5 4 1 2
Macro 0.8 0.9 0.8 0.8 0.9 0.8 0.7 0.9 0.8 0.5 0.7 0.5 0.4 0.7 0.4
avg 2 2 5 1 3 4 9 4 9 5 5 6 8
Weighte 0.9 0.9 0.9 0.9 0.9 0.9 0.9 0.9 0.9 0.8 0.7 0.8 0.8 0.6 0.7
d avg 8 6 6 7 7 5 5 6 6 9 6 8 8 4
LSTM CNN MLP SVC LR

In terms of Precision, Recall, and F1-score, Neural- score) and the class of interest. The choice of the
Network-Based algorithms outperformed other models "better" model may vary depending on the application's
when the Macro and Weighted average outcomes for priorities. Here's a brief analysis:
each measure were taken into account. The outcomes
also showed that the LSTM sequential model had the 1. Class A:
highest precision, recall, and F1-score scores of all the x LSTM, CNN, and MLP have high precision
evaluated models (0.98, 0.96, and 0.96 Weighted (0.99), recall (0.97), and F1-score (0.98).
Average for Precision, Recall, and F1-score). x SVC and LR have slightly lower precision (0.97)
and F1-score (0.84).
The table presents a performance evaluation of various
2. Class B:
machine learning models (LSTM, CNN, MLP, SVC,
and LR) across different classes (A, B, C, D, E) based x CNN has the highest precision (0.58) and F1-
on precision, recall, and F1-score metrics. Precision score (0.68).
measures the accuracy of positive predictions, recall x SVC has the highest recall (0.86).
assesses the ability to capture all relevant instances, and x LSTM, MLP, and LR also perform reasonably
F1-score combines both precision and recall into a well.
single metric. Each model's performance is assessed for 3. Class C:
each class, and the values in the table represent the
corresponding precision, recall, and F1-score for that x All models show good performance, with high
precision (0.87-0.92), recall (0.8-0.95), and F1-
class. The "Macro avg" row provides the average
score (0.8-0.92).
metrics across all classes without considering class
x LSTM, CNN, and MLP have slightly higher F1-
imbalance, while the "Weighted avg" row considers the scores.
class distribution. The higher the values, the better the 4. Class D:
model's performance. This table serves as a
comprehensive comparison of the models' x LSTM, CNN, and MLP have higher precision,
effectiveness in distinguishing and classifying recall, and F1-score.
instances across different categories. x SVC and LR have lower precision and F1-score.
5. Class E:
To determine which model is better, we need to
consider the specific metric (precision, recall, or F1-
2024 14th International Conference on Cloud Computing, Data Science & Engineering (Confluence)
Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY SURATHKAL. Downloaded on May 11,2025 at 04:57:07 UTC from IEEE Xplore. Restrictions apply.
9
 x All models perform well, with high precision study introduces context-aware healthcare using
(0.95-0.99), recall (0.96-0.99), and F1-score Digital Twins and other cutting-edge technology,
(0.96-0.99). which could improve healthcare delivery and patient
x LSTM, CNN, and MLP have slightly higher F1- outcomes.
scores.
Considering the "Macro avg" and "Weighted avg" References
rows, LSTM tends to have higher average performance
[1] Z. Zhang, F. Wen, Z. Sun, X. Guo, T. He, and
across all classes, followed closely by CNN and MLP.
C. Lee, “Arti fi cial Intelligence-Enabled
However, the choice of the best model depends on the Sensing Technologies in the 5G / Internet of
specific priorities of the application. If precision is Things Era : From Virtual Reality / Augmented
more critical, models with higher precision values Reality to the Digital Twin,” 2022, doi:
might be preferred, while if recall is a priority, models 10.1002/aisy.202100228.
with higher recall values would be favoured. The trade-
off between precision and recall is reflected in the F1- [2] V. Sharma, K. Sharma, and A. Kumar, “From
score, making it a commonly used balanced metric for Theory to Practice: A Systematic Review of
Digital Twin Implementations Across Industry
model evaluation.
4.0,” in 2023 14th International Conference on
D. Emerging Issues with Digital Twins Computing Communication and Networking
Technologies (ICCCNT), IEEE, 2023, pp. 1–7.
First, there is a lack of faith in data-transmitting
gadgets. [3] V. Sharma and S. Ramamoorthy, “A Review
on Secure Data access through Multi-Keyword
x There is an immediate demand for competent
Searching in Cloud Storage,” in 2021 Third
and honest experts in the sector.
International Conference on Intelligent
x Concerns about the reliability of Machine Communication Technologies and Virtual
Learning models. Mobile Networks (ICICV), IEEE, 2021, pp.
x Consensus, education, and technological 70–73.
progress are essential for establishing confidence.
[4] M. A. I. Mozumder, M. M. Sheeraz, A. Athar,
Data Security and Privacy S. Aich, and H. C. Kim, “Overview:
x Safeguarding Your Data from Hackers and Technology Roadmap of the Future Trend of
Other Online Dangers. Metaverse based on IoT, Blockchain, AI
x Managing private information safely. Technique, and Medical Domain Metaverse
x Integration of Internet-of-Things devices is a Activity,” Int. Conf. Adv. Commun. Technol.
complex process. ICACT, vol. 2022-Febru, pp. 256–261, 2022,
x Adherence to data protection laws. doi: 10.23919/ICACT53585.2022.9728808.
[5] D. M. Botín-Sanabria, S. Mihaita, R. E.
Search for Standardisation Peimbert-García, M. A. Ramírez-Moreno, R.
x Security, privacy, and the ability to sync data A. Ramírez-Mendoza, and J. de J. Lozoya-
are all affected. Santos, “Digital Twin Technology Challenges
x Digital twin adoption can be sped up with the and Applications: A Comprehensive Review,”
use of standardized procedures. Remote Sens., vol. 14, no. 6, pp. 1–25, 2022,
doi: 10.3390/rs14061335.
IV. Conclusion
This paper provides a digital twin architecture for [6] N. Khan, I. U. Haq, F. U. M. Ullah, S. U. Khan,
health care systems that are aware of their environment. and M. Y. Lee, “Cl-net: Convlstm-based
A patient's digital twin employs an ECG-based heart hybrid architecture for batteries’ state of health
rhythms classifier to monitor health and discover and power consumption forecasting,”
Mathematics, vol. 9, no. 24, 2021, doi:
anomalies. The efficiency and precision of five
10.3390/math9243326.
methods were compared. Digital Twins in healthcare
improve health, lifespan, and cost while creating [7] H. Hassani, X. Huang, and S. MacFeely,
intelligent, holistic Health-Ecosystems. It's supposed to “Impactful Digital Twin in the Healthcare
fix medical difficulties. ECG classifiers inspire Revolution,” Big Data Cogn. Comput., vol. 6,
Machine Learning and AI for continuous health no. 3, pp. 1–17, 2022, doi:
monitoring, which increases quality of life. ECG 10.3390/bdcc6030083.
analysis is better with Neural-Network-based methods [8] M. Kherbache, M. Maimour, and E. Rondeau,
than with classical
2024 ML.
14th Use real-time
International data, increase
Conference on use Computing, “Digital
Cloud Data Science
twin&atEngineering
network for(Confluence)
fromthe
IEEE IIoT
Xplore. using
10
Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY SURATHKAL. Downloaded on May 11,2025 04:57:07 UTC Restrictions apply.
cases to include additional body metrics, and integrate eclipse ditto and hono,” IFAC-PapersOnLine,
more systems are framework development goals. This vol. 55, no. 8, pp. 37–42, 2022.
 [9] S. Jamil, M. Ur Rahman, and Fawad, “A 2021, doi: 10.1109/JIOT.2021.3051158.
Comprehensive Survey of Digital Twins and
Federated Learning for Industrial Internet of [17] F. Pilati, R. Tronconi, G. Nollo, S. S. Heragu,
Things (IIoT), Internet of Learning for and F. Zerzer, “Digital twin of covid‐19 mass
Industrial Internet of Things (IIoT), Internet of vaccination centers,” Sustain., vol. 13, no. 13,
Vehicles (IoV) and Internetofof Drones pp. 1–26, 2021, doi: 10.3390/su13137396.
(IoD),” Appl. Syst. Innov., vol. 5, no. 3, 2022, [18] E. Piacentino and C. Angulo, “Generating fake
doi: 10.3390/asi5030056. data using GANs for anonymizing healthcare
[10] S. R. Newrzella, D. W. Franklin, and S. Haider, data,” in International Work-Conference on
“5-Dimension Cross-Industry Digital Twin Bioinformatics and Biomedical Engineering,
Applications Model and Analysis of Digital Springer, 2020, pp. 406–417.
Twin Classification Terms and Models,” IEEE [19] T. Erol, A. F. Mendi, and D. Doğan, “Digital
Access, vol. 9, pp. 131306–131321, 2021, doi: transformation revolution with digital twin
10.1109/ACCESS.2021.3115055. technology,” in 2020 4th international
[11] Y. Wu, K. Zhang, and Y. Zhang, “Digital Twin symposium on multidisciplinary studies and
Networks: A Survey,” IEEE Internet Things J., innovative technologies (ISMSIT), IEEE, 2020,
vol. 8, no. 18, pp. 13789–13804, 2021, doi: pp. 1–7.
10.1109/JIOT.2021.3079510. [20] T. Suzuki, H. Takao, S. Rapaka, and S.
[12] T. Erol, A. F. Mendi, and D. Dogan, “The Fujimura “Rupture Risk of Small Unruptured
Digital Twin Revolution in Healthcare,” 4th Intracranial Aneurysms in Japanese Adults,”
Int. Symp. Multidiscip. Stud. Innov. Technol. Stroke, pp. 641–643, 2020, doi:
ISMSIT 2020 - Proc., 2020, doi: 10.1161/STROKEAHA.119.027664.
10.1109/ISMSIT50672.2020.9255249. [21] N. Golse, F. Joly, P. Combari, and M. Lewin,
[13] J. Zhang and Y. Tai, “Secure medical digital “Predicting the risk of post-hepatectomy portal
twin via human-centric interaction and cyber hypertension using a digital twin: A clinical
vulnerability resilience,” Conn. Sci., vol. 34, proof of concept,” J. Hepatol., vol. 74, no. 3,
no. 1, pp. 895–910, 2022, doi: pp. 661–669, 2021.
10.1080/09540091.2021.2013443. [22] R. Laubenbacher, A. Niarakis, and T. Helikar,
[14] T. Sun, X. He, and Z. Li, “Digital twin in “Building digital twins of the human immune
healthcare: Recent updates and challenges,” system: toward a roadmap,” npj Digit. Med.,
Digit. Heal., vol. 9, 2023, doi: vol. 5, no. 1, p. 64, 2022.
10.1177/20552076221149651. [23] A. Haleem, M. Javaid, R. Pratap Singh, and R.
[15] A. De Benedictis, N. Mazzocca, A. Somma, Suman, “Medical 4.0 technologies for
and C. Strigaro, “Digital Twins in Healthcare: healthcare: Features, capabilities, and
an architectural proposal and its application in applications,” Internet Things Cyber-Physical
a social distancing case study,” IEEE J. Syst., vol. 2, no. April, pp. 12–30, 2022, doi:
Biomed. Heal. Informatics, vol. XX, no. Xx, 10.1016/j.iotcps.2022.04.001.
pp. 1–12, 2022, doi: [24] B. Grosman, A. Roy, D. Wu, and N. Parikh,
10.1109/JBHI.2022.3205506. “1006-P: personalized hybrid closed-loop
[16] H. Elayan, M. Aloqaily, and M. Guizani, therapy using a digital twin in patients with
“Digital Twin for Intelligent Context-Aware Type 1 diabetes: at-home data,” Diabetes, vol.
IoT Healthcare Systems,” IEEE Internet 69, no. Supplement_1, 2020.
Things J., vol. 8, no. 23, pp. 16749–16757,

2024 14th International Conference on Cloud Computing, Data Science & Engineering (Confluence) 11
Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY SURATHKAL. Downloaded on May 11,2025 at 04:57:07 UTC from IEEE Xplore. Restrictions apply.