Graphical Modeling-Based Digital Twin from Biomedical Data for Spinal Cord Injury Individuals

Graphical modeling (GM) was originally proposed as a statistical modeling framework that uses probabilistic graphs to encode conditional dependences and independence among variables of interest [1]. It is straightforward for stakeholders to understand the interaction among different factors and reasons using the structure and relations. Due to the strong ability of GM to represent complex probability distributions, it has been developed and adapted to various fields including earth science [2], biology [3], and healthcare [4] to dissect the structure and sparsity of targeted phenomenon to improve the understanding of mechanism and decision-making process. In digital healthcare and medicine with a high-stake nature, it is imperative to enable a transparent and accessible modeling paradigm such that both modeling technicians and medical experts can comprehend, trust, and improve the digital models for decision-making. Given the transparent structural representation of GM, we aim to develop an explainable Artificial intelligence (xAI) framework based on CGM (Causal Graphical Models, a special type of GM where edges represent causal relations) to cope with the challenges of real-world medical data and improve both the capacity and capability of medical AI. As a validation scenario of our approach, we are working with the Swiss Paraplegic Centre (SPZ) to improve the long-term prognosis of individuals with spinal cord injury (SCI). Secondary health conditions such as pulmonary infections, pressure injuries (PI), autonomic dysreflexia (AD), and cardiovascular dysregulation are highly prevalent among the SCI population, imposing a tremendous risk on the patient’s health and the healthcare system [5]. Through the medical records and multimodal-sensory monitoring system, we expect to end up with the digital twins of SCI individuals that can fuse different data sources, interact with clinical operations and expertise, and guide medical treatment. Your Tasks: 1. Literature Review on Digital Twins and Graphical Models: You will study literature related to graphical models with applications, digital twins with applications, and preferably the combination of both fields (GMDT). 2. Data Exploration and Analysis: You will explore multimodal datasets collected for PI and AD. The PI data record the conditions of hundreds of SCI individuals during their hospital stay. There are mixed-type variables including demographic features, lab testing values, and health conditions observed during their stay. Additionally, you will have access to multivariate time series extracted from bio-signals on SCI people under intervention experiments. 3. Methodology Development and Deployment: You will help define what data to use, design the experiments, develop, and validate both the modeling foundation and implementation pipeline for the GMDT. A graphical user interface is expected to be built based on the current framework we are developing in the lab. 4. Presentation and Documentation: You will prepare a high-quality manuscript for publication with a clear and engaging presentation of the methodology and validation results. References: [1] Glymour, C., Zhang, K. & Spirtes, P. Review of Causal Discovery Methods Based on Graphical Models. Front. Genet. 10, 524 (2019). [2] Runge, J. et al. Inferring causation from time series in Earth system sciences. Nat. Commun. 10, 2553 (2019). [3] Friedman, N. Inferring Cellular Networks Using Probabilistic Graphical Models. Science 303, 799–805 (2004). [4] Richens, J. G., Lee, C. M. & Johri, S. Improving the accuracy of medical diagnosis with causal machine learning. Nat. Commun. 11, 3923 (2020). [5] Haisma, J. et al. Complications following spinal cord injury: occurrence and risk factors in a longitudinal study during and after inpatient rehabilitation. J. Rehabil. Med. 39 5, 393–8 (2007).

Graphical modeling (GM) was originally proposed as a statistical modeling framework that uses probabilistic graphs to encode conditional dependences and independence among variables of interest [1]. It is straightforward for stakeholders to understand the interaction among different factors and reasons using the structure and relations. Due to the strong ability of GM to represent complex probability distributions, it has been developed and adapted to various fields including earth science [2], biology [3], and healthcare [4] to dissect the structure and sparsity of targeted phenomenon to improve the understanding of mechanism and decision-making process.
In digital healthcare and medicine with a high-stake nature, it is imperative to enable a transparent and accessible modeling paradigm such that both modeling technicians and medical experts can comprehend, trust, and improve the digital models for decision-making. Given the transparent structural representation of GM, we aim to develop an explainable Artificial intelligence (xAI) framework based on CGM (Causal Graphical Models, a special type of GM where edges represent causal relations) to cope with the challenges of real-world medical data and improve both the capacity and capability of medical AI. As a validation scenario of our approach, we are working with the Swiss Paraplegic Centre (SPZ) to improve the long-term prognosis of individuals with spinal cord injury (SCI). Secondary health conditions such as pulmonary infections, pressure injuries (PI), autonomic dysreflexia (AD), and cardiovascular dysregulation are highly prevalent among the SCI population, imposing a tremendous risk on the patient’s health and the healthcare system [5]. Through the medical records and multimodal-sensory monitoring system, we expect to end up with the digital twins of SCI individuals that can fuse different data sources, interact with clinical operations and expertise, and guide medical treatment.

Your Tasks:

1. Literature Review on Digital Twins and Graphical Models: You will study literature related to graphical models with applications, digital twins with applications, and preferably the combination of both fields (GMDT).

2. Data Exploration and Analysis: You will explore multimodal datasets collected for PI and AD. The PI data record the conditions of hundreds of SCI individuals during their hospital stay. There are mixed-type variables including demographic features, lab testing values, and health conditions observed during their stay. Additionally, you will have access to multivariate time series extracted from bio-signals on SCI people under intervention experiments.

3. Methodology Development and Deployment: You will help define what data to use, design the experiments, develop, and validate both the modeling foundation and implementation pipeline for the GMDT. A graphical user interface is expected to be built based on the current framework we are developing in the lab.

4. Presentation and Documentation: You will prepare a high-quality manuscript for publication with a clear and engaging presentation of the methodology and validation results.


References:

[1] Glymour, C., Zhang, K. & Spirtes, P. Review of Causal Discovery Methods Based on Graphical Models. Front. Genet. 10, 524 (2019).

[2] Runge, J. et al. Inferring causation from time series in Earth system sciences. Nat. Commun. 10, 2553 (2019).

[3] Friedman, N. Inferring Cellular Networks Using Probabilistic Graphical Models. Science 303, 799–805 (2004).

[4] Richens, J. G., Lee, C. M. & Johri, S. Improving the accuracy of medical diagnosis with causal machine learning. Nat. Commun. 11, 3923 (2020).

[5] Haisma, J. et al. Complications following spinal cord injury: occurrence and risk factors in a longitudinal study during and after inpatient rehabilitation. J. Rehabil. Med. 39 5, 393–8 (2007).

We envision a GM-based digital twin (GMDT) that can represent the hierarchical relations of different levels of abstraction in terms of demographical features, functioning abilities, daily activities, and health conditions for SCI individuals (a simple illustration of the idea is shown in the figure). This GMDT will be built based on the findings and results of previous studies together with specialized clinical trials and rehabilitation. With the structural equations (conditional distributions) induced by the GM, downstream tasks such as prediction, causal inference, and counterfactual reasoning can be conducted. The assimilation and evolution between the physical twin (SCI individual) and the digital twin will be implemented under the framework of GM similar to a recent approach [6]. [6] Kapteyn, M. G., Pretorius, J. V. R. & Willcox, K. E. A probabilistic graphical model foundation for enabling predictive digital twins at scale. Nat. Comput. Sci. 1, 337–347 (2021).

We envision a GM-based digital twin (GMDT) that can represent the hierarchical relations of different levels of abstraction in terms of demographical features, functioning abilities, daily activities, and health conditions for SCI individuals (a simple illustration of the idea is shown in the figure). This GMDT will be built based on the findings and results of previous studies together with specialized clinical trials and rehabilitation. With the structural equations (conditional distributions) induced by the GM, downstream tasks such as prediction, causal inference, and counterfactual reasoning can be conducted. The assimilation and evolution between the physical twin (SCI individual) and the digital twin will be implemented under the framework of GM similar to a recent approach [6].

[6] Kapteyn, M. G., Pretorius, J. V. R. & Willcox, K. E. A probabilistic graphical model foundation for enabling predictive digital twins at scale. Nat. Comput. Sci. 1, 337–347 (2021).

Host: Dr. Diego Paez (SCAI-Lab, ETHZ | SPZ) Supervision: Dr. Diego Paez & Yanke Li (SCAI-Lab, ETHZ | SPZ) Please send your CV and the latest transcript of records to Yanke Li (yanke.li@hest.ethz.ch)

Host: Dr. Diego Paez (SCAI-Lab, ETHZ | SPZ)
Supervision: Dr. Diego Paez & Yanke Li (SCAI-Lab, ETHZ | SPZ)
Please send your CV and the latest transcript of records to Yanke Li (yanke.li@hest.ethz.ch)