Digital Twins in Neuroscience

The term digital twins appeared in a 1993 book by David Gelernter entitled Mirror Worlds: Or the Day Software Puts the Universe in A Shoebox. How It Will Happen and What It Will Mean. In one sentence, digital twins are “precise, virtual copies of machines or systems” (Tao and Qi, 2019). In more detail, they are “sophisticated computer models” that “mirror almost every facet of a product, process or service” and can be constantly updated by data collected from sensors in real time (Tao and Qi, 2019).

Such twins allow not only visualization but also experimentation and forecasting of future scenarios (Wickramasinghe et al., 2022) from micro to macro. They first emerged in engineering to test products, but they have now been piloted to sketch solutions and preempt problems in many contexts, from logistic management to global warming. Energy companies employ digital twins to track the operations of wind turbines, and NASA has been using digital copies (i.e., of spacecraft to monitor their status) since the 1960s (Dang et al., 2023). Singapore is the first country in the world to have a digital twin copy of itself, street by street. Dynamic bidirectional mapping is one of the key aspects of digital twins, which can collect real-world data and allow simulations of physical entities in real time (Wickramasinghe et al., 2022), along with their analytical and predictive capability (Katsoulakis et al., 2024).

Among the many areas in which they have been piloted is medicine, including cardiology, dermatology, geriatrics, infective disorders, internal medicine, oncology, orthopedics, and radiology specialities. We are now closer to Mirror Worlds becoming a neuro-reality more than ever, as digital twins have also entered the neuroscience realm (Fig. 1). Multiple sclerosis, one of the most common causes of neurological disability in young adults, can be an ideal disease to target because it is a chronic autoimmune, degenerative, and lifelong condition characterized by complexity, heterogeneous course, and multidimensionality (Voigt et al., 2021). Digital twins could be modeled on each patient and might offer a way to deal with such complexity while providing new insights into individualized management and patient participation (Voigt et al., 2021). Also, a digital twin conceptual framework might help build an AI algorithm to estimate the age of onset of disease-specific brain atrophy, which is a known clinical problem in multiple sclerosis, and can do so even when longitudinal data are sparse or lacking (Cen et al., 2023).

Figure 1.

A figurative representation of a digital twin in neuroscience, from biology to predictions via data collection and model.

Digital twins might support forecasting disease progression in patients with ischemic stroke. In one study, a model based on a variational autoencoder was applied, and simulated patient trajectories were indistinguishable from real patient data, with “similar feature means, standard deviations, inter-feature correlations, and covariance structures” (Allen et al., 2021). Digital twins can even help assess short-term adaptation and long-term recovery in stroke patients during rehabilitation. By adjusting the state and parameters, digital twins can offer a valuable window into the adaptation level and rehabilitation progress (Chen et al., 2023). They can support healthcare providers while making even more informed decisions. For common diseases, such as dementia, digital twins might ensure greater precision when it comes to patient care, considering the increased number of people being diagnosed with dementia every day globally and pressing healthcare issues, with an approach the authors highlighted as being “consistent with a value-based healthcare philosophy” (Wickramasinghe et al., 2022).

Conceptually, digital twins go toward patient-centered care and precision medicine (Voigt et al., 2021), may provide a cost-efficient method of conducting theoretical clinical trials (Allen et al., 2021), and support surgeons in planning and performing surgical operations. Patient-specific twins could integrate human physiology and immunology with patient-specific clinical data to produce real-time predictive simulations of immune responses and viral infection (Laubenbacher et al., 2021). Sophisticated twins are also self-improving: they monitor the divergence between predictions and observations and use these divergences to improve their accuracy (Laubenbacher et al., 2021). By combining continual forecasting and small-scale interventions, digital twins can potentially prevent problems, reduce the occurrences of critical failures (Laubenbacher et al., 2021), and promote early interventions. In theory, digital twins can be “treated” with drugs computationally, and the therapy with the best effect on the digital twin could be suggested for treating the real patient.

Digital twins can also mirror a department or an entire hospital. Creating digital twins of departments and hospitals might allow better planning, care coordination, and accurate management of resources, as twins can provide a broader and more comprehensive perspective on several processes (Erol et al., 2020). Among the potential benefits are more accurate predictions, better resource allocation, and error minimization.

From an economic perspective, the global market size for the Internet of Things in healthcare, which includes smart sensors and wearables that can collect and send data about entities and environments around them, is expected to grow from USD 72.5 billion in 2020 to USD 188.2 billion by 2025, at a compound annual growth rate of 21% (Erol et al., 2020). The market size for global healthcare digital twins was estimated at USD 572.4 million in 2022 and is forecasted to grow at a compound annual growth rate of 25.6% from 2023 to 2030 (https://www.grandviewresearch.com/industry-analysis/healthcare-digital-twins-market-report/toc).

Pragmatically, we still do not have an agreed definition of a digital twin nor shared guidelines on creating one. There is uncertainty around its definition: under this umbrella term, we can find anything from a simulation to a 3D model or even a “set of integrated models or software that pairs the digital world with physical assets” or a series of worlds where each approach has its norms (Tao and Qi, 2019). Not only are there many definitions of what a digital twin is, but it is a concept that has evolved over time and is still evolving (Katsoulakis et al., 2024). However, neuroscience has demonstrated that an agreement on the essential features of a digital twin in a very specific circumstance can be found. It is the case of acute stroke care in the neuro-critical care unit. By using the Delphi method, whose assumption is that the forecasts of a structured group of experts are more accurate than those of unstructured groups, 18 experts reached a consensus and brought together knowledge to establish rules for developing a twin model (Dang et al., 2023). After three rounds, 77.5% of statements reached consensus, 9.2% were excluded, and only 13.3% did not reach a consensus of the original 120 statements. The key features of digital twins defined through the Delphi process are also easily interpretable models compared with associative models of AI, which develop rules based on finding associations in datasets (Dang et al., 2023).

While there are advantages to testing neuroscientific scenarios and neurological decisions in silico, several issues exist as digital twins suffer from old and new problems. Some limitations are linked to data collection and management: Which data should be collected? How many pieces of data and how frequently? How can we handle large datasets quickly and efficiently? Missing or erroneous data can distort and bias decision-making, but sampling more data than the system can handle might lead to bottlenecks (Tao and Qi, 2019). Data should be of high quality and securely collected: the collection should adhere to rigorous procedures and ideally be standardized and safely stored (Voigt et al., 2021). The accessibility, integration, privacy, and sharing of data are other major issues (Wickramasinghe et al., 2022), as is the facility by which adding or removing data sources can be executed and who has the permission to do this (Laubenbacher et al., 2021) or the scattered ownership of data (Tao and Qi, 2019), all aspects that can dampen the enthusiasm of many. Data security across platforms and systems and the clarity of data visualization are areas for improvement, along with intellectual property-related aspects. The cost of realizing a digital twin is not a secondary aspect either. Plus, the digital divide is real and risks jeopardizing this innovation's applicability on a global scale.

There is a need to standardize digital twin methods and develop expert consensus rules for digital twin models across diseases in neuroscience. Any digital twin, individual or system, should be extensively validated versus reality before it is used as a basis for decisions. However, while some might like having detailed information about potential future diseases and making “informed” decisions, others might not want to know. Inequalities in the healthcare system might lead to the underrepresentation of some racial or patient groups, and this, in combination with AI bias and reproducibility issues, can be extremely detrimental at different levels.

As penned in Mirror Worlds by David Gelernter in 1993: “… in this sense, Mirror Worlds are a natural culmination of all these technologies of motion and communication; because they show you so much more, a huge sweep instead of a narrow puny medieval silver, and for precisely that reason, they show you better.” Discussions and collaborations between patients, clinicians, computer scientists, engineers, ethicists, and lawyers are urgently needed, as long as they are the real ones and not their digital counterparts.