The emergence of Digital Twin technology, coupled with Big Data paradigms, has enabled advanced analytics for real-time monitoring, simulation, and decision-making across multiple domains. In this paper, we address the emerging challenge of multidimensional risk analysis and prediction in Digital Twin environments, by proposing an innovative Multidimensional Big Data Analytics framework that integrates Digital Twin modeling, Cloud-based infrastructures, and OLAP-driven multidimensional analysis to assess and predict risks effectively. Our approach is designed to support complex risk scenarios, such as those in healthcare, by leveraging distributed data processing, real-time analytics, and predictive modeling. Through extensive experimental evaluations over real-life healthcare datasets, we demonstrate the efficiency and scalability of our framework in performing multidimensional risk analytics and prediction.
Digital twins (e.g., [
alfredo.cuzzocrea@unical.it (A. Cuzzocrea); ismail.benlaredj@unical.it (I. Benlaredj) : 0000-0002-7104-6415 (A. Cuzzocrea); 0009-0003-1138-8039 (I. Benlaredj)
© 2025 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
In summary, the research on digital twins for big data applications is multifaceted, encompassing data management, modeling and simulation, real-time analytics, and interoperability. These efforts are driving the evolution of digital twins from conceptual models to practical tools that enhance efficiency, reliability, and decision-making across various industries.
In this paper, we focus on a specific, innovative research challenge:risk analysis over digital twins (e.g., [12,13]). The final goal of this new idea is to design and develop data analytics methodologies for characterizing, analyzing and, moreover, making predictions about the risk of organizations (e.g., healthcare systems) by inspecting the data and knowledge kept in digital twins. Without any loss of generality, big data technologies have largely been used for risk management activities, in various organizations (e.g., [14,15]). Given the strict interaction with data analytics methodologies and ML, several methodologies have been applied for analyzing organizational risk using digital twin data, particularly in sectors like healthcare. For instance, techniques such as regression analysis, decision trees, and neural networks are used to predict potential risks by analyzing historical and real-time data from digital twins [14]. For instance, in healthcare, predictive analytics can forecast patient deterioration or equipment failure. Similarly, time series analysis analyzes data points collected or recorded at specific time intervals to identify trends, seasonal patterns, and potential anomalies that could indicate risks [15]. Tools like dashboards and heat maps help in visualizing complex data from digital twins, making it easier to identify patterns and outliers that could signify risks [16]. Descriptive statistics such as mean, median, and standard deviation are used as well to summarize and describe the main features of a dataset, providing insights into the normal operating conditions and deviations that might indicate risks [17]. Indeed, leveraging digital twin data through various data analytics methodologies allows organizations to proactively identify, assess, and mitigate risks. These methodologies enhance decision-making, improve operational efficiency, and ultimately contribute to better risk management practices.
In line with this relevant research trend, this paper proposes the application of the innovative multidimensional big data analytics paradigm [18-20] to the specific case of risk analysis, via proposing an innovative model, and the related framework, for supporting multidimensional risk analysis and prediction over digital twins.
The remaining of this paper is structured as follows. Section 2 offers a comprehensive overview of state-of-the-art proposals related to the context of multidimensional risk analysis and prediction over digital twins. Section 3 summarizes our proposed multidimensional big data analytics paradigm as applied to the specific research context. Section 4 describes the framework that implements our methodology for multidimensional risk analysis and prediction. In Section 5, we introduce and describe the real-life datasets used in our experimental assessment. Section 6 presents the first part of our experimental campaign, which consists of multidimensional risk analytics. Similarly, Section 7 presents the second part, which consists of the multidimensional risk prediction. Finally, Section 8 provides conclusions and possible future work.
In this Section, we provide a comprehensive overview of state-of-the-art proposals in the context of multidimensional risk analysis and prediction over digital twins.
[21] contributes significantly to the intersection of computational intelligence, data privacy and Digital Twins (DT) technology for Cyber-Physical Systems (CPS) in smart cities. The approach centers on using Differential Privacy Frequent Subgraph-Big Multigraph (DPFS-BM) for preserving data privacy in DT, which maps the complex physical spaces of CPS into virtual environments. Authors propose a novel Differential Privacy-AlexNet (DP-AlexNet) model for analyzing and predicting Big Graphic Data (BGD) in DT, solving the gradient dispersion problem typically encountered in deep networks. Through comparative analysis with various neural networks (e.g., CNN, RNN, etc.), the model demonstrates superior performance in terms of error rates, training, and testing time.
In the field of precision medicine forCardio-Vascular Disease (CVD), the concept of digital twins
has gained attention for its potential to improve patient-specific diagnosis and treatment. By creating
a virtual representation of a patient that integrates real-time updates from various clinical, imaging,
molecular and so forth, the DT can optimize treatment decisions and predict disease progression. In
[
[
In [
[
In [
[
[
[
Finally, [
From the analysis of the active literature, the strident interest for our investigated topic clearly emerges.
In this Section, we propose a brief summary about the multidimensional big data analytics paradigm we proposed recently [18-20], and how it is applied to the multidimensional risk analysis and prediction problem.
Basically, the proposed methodology aims at applying the well-known multidimensional and
OLAP analysis paradigm [
designing and implementing a proper big multidimensional analysis model over the target heterogeneous data sources, driven by specific analysis goals;
designing and implementing a proper (distributed) MapReduce-based big multidimensional data storage model, based on a reference big data platform (e.g., Apache Hadoop); designing and implementing a proper collection of (distributed) functional-based multidimensional big data analytics procedures, driven by specific analytics goals; designing and implementing a proper collection of (distributed) multidimensional big data visualization tools, driven by the target knowledge fruition .
It should be noticed here that, as regards research challenges, the most annoying problem is
represented by the issue of creating and developing ad-hoc big data management and processing
algorithms, which support the analytics and prediction phases, by strictly taking into account
performance aspects, which play a relevant role in emerging big data research (e.g., [
When applied to the big data scenarios represented by digital twins, the most significant aspect for the methodology we propose is represented by the issue of appropriately capturing the streaming and pervasive nature of such kind of data sources, by integrating a proper functional component. In addition, another important aspect for the case of digital twins is represented by the need for big data prediction tools, due to specific operational settings of these environments.
Finally, still in the specific context we investigate, the core multidimensional model is “shaped” as a multidimensional risk model, meaning that the measures to analyze are connected to the evaluation of the risk in the specific setting (e.g., healthcare) and the core multidimensional predictive model is “shaped” as a multidimensional risk prediction model, both tailored to digital twins.
In this Section, we introduce our proposed innovative framework designed to support multidimensional risk analysis and prediction over digital twins, particularly in the healthcare domain. Notably, this framework is distinguished by its integration of digital twin modeling, Cloudbased infrastructure, and advanced analytics, which enable risk assessment and prediction. Figure 1 shows a comprehensive overview of this reference framework.
As shown in Figure 1, the framework is composed of several key layers, each playing a crucial role in the overall functioning of the framework. These layers are described below and will be discussed in detail in the following.
Cyberphysical World: this layer focuses on gathering data from various sensors, such as heart sensors, which provide critical information about heart disease analysis;
Cloud-Based Infrastructure: this infrastructure consists of two core components: first,Digital Twin Modeling component, which involves interconnected Cloud nodes with Apache Kafka instances that handle streaming data produced by digital twins; second, Multidimensional Risk Modeling component is responsible for creating data cubes in each Cloud node in order to be efficiently processed and analyzed;
Multidimensional Risk Analysis: this layer performs Cloud-based OLAP-shaped multidimensional risk analysis, enabling the exploration of heart disease risks from multiple dimensions;
Multidimensional Risk Analytics: this layer focuses on supporting multidimensional risk analytics derived from the analysis, making it easier to interpret and understand the multidimensional risk data;
Multidimensional Risk Prediction: this final layer is responsible for providing predictive insights based on multidimensional risk analysis.
In the following, we examine these layers and components with greater detail, by highlighting
their specific roles within the proposed framework. While our proposed framework is general enough
to capture any innovative big data setting scenario, here we specialize our description for the very
relevant case represented by the digital twin healthcare scenario (e.g., [
1. Cyberphysical World: this layer is the initial layer of our framework, where real-time healthcare data from the physical environment are captured. In the context of cardiovascular healthcare, it includes patient health metrics such as heart rate, blood pressure, Electrocardiogram (ECG) readings, and other cardiovascular signals. The aim of this layer is to ensure that an accurate dataset is available for analysis. The data gathered here serves as input to the following layers in the framework. This layer is crucial because the accuracy and quality of data directly impact the effectiveness of the analytical models and predictions of heart diseases; 2. Cloud-Based Infrastructure: this layer is crucial due to its importance in providing an environment for processing and managing large amounts of data collected from the cyberphysical world layer. The infrastructure is composed of two primary components, (i) Digital Twin Modeling component; (ii) Multidimensional Risk Modeling component, each playing a pivotal role in the data analysis process.
a. Digital Twin Modeling: this component is composed of several interconnected Cloud nodes,
each equipped with Apache Kafka instances to manage the data streams of the collected
data from heart sensors. Later the collected data is transmitted in CSV format and fed into
these Cloud nodes through Apache Kafka;
b. Multidimensional Risk Modeling: the second critical part of the Cloud-based infrastructure
is the Multidimensional Risk Modeling component, which transforms data into
multidimensional models (OLAP data cubes – e.g., [
Finally, the proposed framework operates in a continuous cycle, starting with real-time data collection from the cyberphysical world, which is then mirrored virtually via digital twin modeling. These data flow through Cloud-based infrastructure for risk modeling and culminate in visualized predictions and risk assessments. The seamless integration of these technological components allows healthcare providers to monitor patient conditions in real-time, predict future risks, and take preventive actions accordingly. This comprehensive solution effectively addresses the challenges faced by modern healthcare systems.
In this Section, we introduce the following three healthcare-related datasets: HeartStudy Dataset [
HeartStudy [
DiabetesHealthcare [
LungCancer [
In this Section, we are going to experiment with three different datasets related to healthcare, with particular attention to the multidimensional risk analytics supported by our proposed framework. Accordingly, we create dashboards for each dataset using Apache Superset in order to present key analytics and visualizations that help uncover critical patterns, trends, and actionable insights, contributing to data-driven decision-making in healthcare.
As described in Section 4, the multidimensional risk analytics is generated by our multidimensional OLAP model embedded in the framework, which adheres to the underlying multidimensional big data analytics paradigm [18-20].
First, we consider the HeartStudy dataset. The constructed HeartStudy dashboard showcases a series of analytical plots related to the target dataset. These visualizations provide insights into various key metrics and patterns, enabling a deeper understanding of the dataset attributes and trends. Figure 2 shows the HeartStudy dashboard, which provides an overview of important smoking and health patterns. The dashboard focuses the attention on the relationship between smoking cigarettes and CHD risk. It demonstrates 37.9K cigarettes per day in the overall dataset, and the most represented age in the dataset was from 40 to 50. Males smoke more than females. The cigarette smoking rate for people aged 40 to late 40s is the highest. Both males and females have roughly the same average glucose level. At the age of 32, smoking rates reached their peak and remained consistent thereafter.
Second, we consider the DiabetesHealthcare dataset. Figure 3 shows the DiabetesHealthcare dashboard, which is underlined by a few of the following health metric indicators. The AVG(glucose) maintains its stability in the range between 80 to 160. The histogram of SUM(pregnancies) shows that women aged 40 to 45 have the highest number of pregnancies. The area plot of AVG( bloodpressure) shows stability in the range of 50 to 90. The pie plot of AVG(skinthickness) shows similar averages across all ages. The AVG(insulin) is 178.78. Finally, two line plots represent how the MAX(BMI) will be in the range of 40 to 65, and MIN(BMI) is in the range of 0 to 30.
Finally, the LungCancer dataset is considered, being the corresponding LungCancer dashboard shown in Figure 4. Here are some of the insights derived from this dashboard. Two pie plots reveal that females have slightly higher average anxiety levels, while the COUNT(lung_cancer) is relatively the same for both genders. An histogram of COUNT(smoking) shows that males smoke more than females, with the most smokers in the late 50s to early 60s age group. Another histogram of AVG(allergy) indicates stability, with males slightly more affected. A line plot of AVG(anxiety) shows that both genders have similar anxiety levels, and the total COUNT(smoking) is 1.01K.
In this Section, we focus on the second aspect of our experimental assessment and analysis, i.e. the multidimensional risk prediction. Our model includes all steps from data processing, analysis, and generating predictions to the final results, and advanced techniques including data aggregation, multidimensional data analysis, and reporting with suitable plots. The very relevant innovation of our risk prediction approach is, again, the multidimensionality of the analysis, meaning that our approach completely captures the multidimensional nature of the surrounding analysis problem, in order to finally enhance theexpressive power of the overall (analytics) process.
In our experimental campaign, we have developed Credit Scoring [
In the following, due to space limitation, we present our experimental results as restricted to the multidimensional risk prediction over the HeartStudy dataset only, but we conducted the prediction accuracy analysis over all the three datasets, with similar results. 7.1.
In our Credit Scoring model [
Here, we present the results of applying our Credit Scoring model to the first of the three healthcare datasets previously described (i.e., HeartStudy dataset). By focusing on this initial dataset, we provide a closer look at the predictions, combining and aggregating various attributes. This approach uncovered valuable insights and helped identify relationships between key indicators.
Through the application of our model on the HeartStudy dataset, we generate predictions for the target variable, which is a 10-year risk of future CHD, we provide detailed analysis and informative reports. These outputs are visually represented in 3 D plots.
Figure 5 displays the prediction counts for CHD. In order to capture the multidimensional modelling, in Figure 5 ( a ) we select three attributes, including the target variable CHD. Numerical variables, such as heartRate and cigsPerDay, have been grouped into intervals (mostly four intervals). The first two attributes, heartRate and gender, have been combined resulting in a combination of intervals, while the third attribute, cigsPerDay, has been presented independently. The results have been aggregated based on the three selected attributes, with the prediction counts split into two categories: CHD and No CHD. Overall, this originates a multidimensional prediction model over digital twin data extracted from the target dataset, according to the three-dimensional range cigsPerDay, heartRate, gender.
Similarly, Figure 5 ( b ) shows the prediction counts for CHD over the range glucose, diaBP,
sysBP. For this range, we selected the three attributes, glucose, diaBP and sysBP, including the target
variable CHD. All these attributes were grouped into four intervals. The first two attributes,sysBP and
diaBP, have been combined resulting in a combination of bins, while the third attribute, glucose, has
been used independently. It is easy to notice, here, that this approach can easily scale to higher
numbers of dimensions, since multiple dimensions can be easily grouped/projected into different
subgroups of dimensions (e.g., [
As shown in Figure 5 ( b ), the predictions demonstrate that patients with high values of the three attributes (glucose > 110, diaBP > 90, and sysBP > 180) have a serious risk of developing CHD. We observe the risk of having CHD decreases with glucose less than 110; however, we notice that high diaBP and sysBP values still put some patients at risk of developing CHD even with a low level of glucose.
Figure 6 shows the prediction counts for CHD over the range currentSmoker, heartRate, prevalentHip. In this case, we selected three attributes, including the target variable CHD. All these attributes were grouped into four intervals. The first two attributes, (i.e., currentSmoker and prevalentHip), have been combined while the third attribute, heartRate, has been represented independently.
As a result, the predictions show that patients who are smokers and who have prevalent hypertension are at risk of developing CHD, especially patients with irregular heart rates ( heartRate > 90). On the other hand, predictions show that non-smoker patients without prevalent hypertension have a low risk of CHD. 7.2.
In our Deep Learning Network model [
With prepared data, we define the network architecture, including the parameters of layers, functions, and other key components. Afterward, we fine-tune the training parameters, such as Learning Rate, Batch Size and Epochs Number. Once these steps are completed, we proceed to train the network using the prepared data. For efficiency, a pre-trained model, named residualTrainedNetworkMacro, is provided, which allows us to bypass the training process and directly explore the results with other versions of the dataset. Moreover, the same advanced visualization techniques are employed for more detailed reports.
Through the application of our model on the HeartStudy dataset, we generate predictions for the target variable, which is the 10-year risk of future coronary heart disease CHD, thus furnishing detailed analyses and informative reports. These outputs are then visually represented in ad-hoc plots. The prediction model discriminates between the two same categories exploited for the Credit Scoring model, i.e. CHD and No CHD (see Section 7.1).
Figure 7 ( a ) displays the prediction counts for CHD over the range ⟨ cigsPerDay , heartRate , gender ⟩. By analyzing the results, the predictions indicate that female smokers have a higher likelihood of developing CHD compared to males. We observe that smokers with cigsPerDay < 5 are the biggest group, therefore, we observe that the latter is the group having the highest chance of getting a CHD, despite males and females having some numbers of getting CHD across all cigsPerDay interval cases.
Figure 7 ( b ) shows the prediction counts for CHD over the range ⟨ glucose , diaBP , sysBP ⟩. The predictions reveal that patients with high levels of all three attributes ( glucose > 110, diaBP > 90, and sysBP > 180) are at a significantly increased risk of developing CHD. While the risk decreases when glucose levels fall below 110; however, we observe that patients with high diaBP and sysBP values remain at risk, even if their glucose levels are low.
Figure 8 shows the prediction counts for CHD over the range ⟨ age , education , totChol ⟩. The predictions reveal that patients with lower education levels (degrees 1 and 2), as well as higher cholesterol levels (totChol > 200), tend to be connected with an increase in developing CHD, particularly among the older patients (age > 50). Contrarily, for patients under 40, the counts for CHD are relatively low across all education and cholesterol levels. It should be noticed here that the latter situation is very close to the real and effective today’s status about health guidelines for large communities, at various scales such as the province scale and the regional scale.
(a) (b) Figure 7: Prediction Counts for ⟨ cigsPerDay , heartRate , gender ⟩ ( a ) ⟨ glucose , sysBP , diaBP ⟩ ( b ) (Deep Learning Network Model) and
This paper has explored the convergence of two innovative and very relevant scientific areas: digital twins and big data technologies. In particular, we focused on the issue of supporting multidimensional risk analysis and prediction over digital twin data, by nicely applying an emerging state-of-the-art analytics paradigm, the so-called multidimensional big data analytics.
The result is represented by a comprehensive framework, for which we provided anatomy and main functionalities. In addition to this, we conducted an extensive experimental evaluation and analysis of the performance of the proposed framework in both the different analytics and prediction phases, over state-of-the-art datasets populating the target healthcare digital twin scenario.
Future work is mainly oriented to further extend our contributions with specific aspects of the emerging big data trend, such as performance optimization and outlier management, and, also, studying possible integration with emerging flexible security paradigms (e.g., [46,47]).
This work was partially supported by project SERICS (PE00000014) under the MUR National Recovery and Resilience Plan funded by the European Union - NextGenerationEU. The authors are grateful to Ilies Merzouk for his contribution to the experimental analysis of this research.
The authors have not employed any Generative AI tools. [4] F. Tao, J. Cheng, Q. Qi, M. Zhang, H. Zhang, F. Sui, “Digital Twin-Driven Product Design, Manufacturing and Service with Big Data”. The International Journal of Advanced Manufacturing Technology 94, pp. 3563-3576, 2018 [5] Q. Qi, F. Tao, “Digital Twin and Big Data Towards Smart Manufacturing and Industry 4.0: 360
Degree Comparison”. IEEE Access 6, pp. 3585-3593, 2018 [6] L. Nie, X. Wang, Q. Zhao, Z. Shang, L. Feng, G. Li, “Digital Twin for Transportation Big Data: A Reinforcement Learning-Based Network Traffic Prediction Approach”. IEEE Transactions on Intelligent Transportation Systems 25(1), pp. 896-906, 2023 [7] L. Nele, G. Mattera, E.W. Yap, M. Vozza, S. Vespoli, “Towards the Application of Machine Learning in Digital Twin Technology: A Multi-Scale Review”. Discover Applied Sciences 6, art. 502, 2024. [8] S.M. Zayed, G. Attiya, A. El-Sayed, A. Sayed, E.E. Hemdan, “An Efficient Fault Diagnosis Framework for Digital Twins Using Optimized Machine Learning Models in Smart Industrial Control Systems”. International Journal of Computational Intelligence Systems 16(1), art. 69, 2023 [9] P. Calvo-Bascones, A. Voisin, P. Do, M.A. Sanz-Bobi, “A Collaborative Network of Digital Twins for Anomaly Detection Applications of Complex Systems. Snitch Digital Twin Concept”.
Computers in Industry 144, art. 103767, 2023 [10] W.Y. Wang, J. Yin, Z. Chai, X. Chen, W. Zhao, J. Lu, F. Sun, Q. Jia, X. Gao, B. Tang, X. Hui, H.
Song, F. Xue, Z. K. Liu, J. Li, “Big Data-Assisted Digital Twins for the Smart Design and Manufacturing of Advanced Materials: From Atoms to Products”. Journal of Materials Informatics 2(1), pp. 1-27, 2022 [11] S. M. Zayed, G.M. Attiya, A. El Sayed. E.E. Hemdan, “A Review Study on Digital Twins with Artificial Intelligence and Internet of Things: Concepts, Opportunities, Challenges, Tools and Future Scope”. Multimedia Tools and Applications 82(30), pp. 47081-47107, 2023 [12] Y. Liang, Z. Yin, L. Nie, Y. Ba, “Shared Steering Control with Predictive Risk Field Enabled by
Digital Twin”. IEEE Transactions on Intelligent Vehicles 8(5), pp. 3256-3269, 2023 [13] L. Lin, H. Bao, N. Dinh, “Uncertainty Quantification and Software Risk Analysis for Digital Twins in the Nearly Autonomous Management and Control Systems: A Review”. CoRR abs/2103.03680 2021 [14] RiskDataControl.com, The Role of Data Analytics in Risk Management. Available at https://www.riskdatacontrol.com/role-data-analytics-risk-management/ [15] M. El Khatib, A. Ankit, I. Al Ameeri, H. Al Zaabi, R. Al Marqab, H.M. Alzoubi, M. Alshurideh, “The Role and Impact of Big Data in Organizational Risk Management”. In: The Effect of Information Technology on Business and Marketing Intelligence Systems. Studies in Computational Intelligence 1056, Springer, pp. 2139-2153, 2023 [16] R. Gul, M. A. S. Al-Faryan, “From Insights to Impact: Leveraging Data Analytics for Data-Driven Decision-Making and Productivity in Banking Sector”. Humanities and Social Sciences Communications 10(1), pp. 1-8, 2023 [17] G. Dicuonzo, F. Donofrio, G. Galeone, “Big Data and Artificial Intelligence to Support Risk Management: A Systematic Literature Review”. In: Chiucchi, M.S., Lombardi, R., Mancini, D. (eds) Intellectual Capital, Smart Technologies and Digitalization. SIDREA Series in Accounting and Business Administration, Springer, pp. 119-130, 2021 [18] A. Cuzzocrea, “OLAPing Big Social Data: Multidimensional Big Data Analytics over Big Social Data Repositories”. In: 4th International Conference on Cloud and Big Data Computing, pp. 15-19, 2020 [19] A. Cuzzocrea, “Multidimensional Big Data Analytics over Big Web Knowledge Bases: Models, Issues, Research Trends, and a Reference Architecture”. In: 8th IEEE International Conference on Multimedia Big Data, pp. 1-6, 2022 [20] A. Cuzzocrea, “A Reference Architecture for Supporting Multidimensional Big Data Analytics over Big Web Knowledge Bases: Definitions, Implementation, Case Studies”. International Journal of Semantic Computing 17(4), pp. 545-568, 2023 [21] Z. Lv, D. Chen, H. Feng, A.K. Singh, W. Wei, H. Lv, “Computational Intelligence in Security of Digital Twins Big Graphic Data in Cyber-physical Systems of Smart Cities”. ACM Transactions on Management Information Systems 13(4), art. 39, 2022 [45] S. Bayraci, O. Susuz, “A Deep Neural Network (DNN) based Classification Model in Application to Loan Default Prediction”. Theoretical and Applied Economics 4(621), pp. 75-84, 2019 [46] M.J.H. Faruk, H. Shahriar, M. Valero, F.L. Barsha, S. Sobhan, M.A. Khan, M.E. Whitman, A.
Cuzzocrea, D.C. Lo, A. Rahman, F. Wu, “Malware Detection and Prevention using Artificial Intelligence Techniques”. In: 9th IEEE International Conference on Big Data, pp. 5369-5377, 2021 [47] M. Masum, H. Shahriar, H. Haddad, M.J.H. Faruk, M. Valero, M.A. Khan, M.A. Rahman, M.I.
Adnan, A. Cuzzocrea, F. Wu, “Bayesian Hyperparameter Optimization for Deep Neural NetworkBased Network Intrusion Detection”. In: 9th IEEE International Conference on Big Data, pp. 54135419, 2021