Human Digital Twins and Machine Learning Applications in Precision Medicine and Surgery: Current State and Future Directions

Abstract

Human digital twins (“HDT”) are one-on-one digital replicas of human beings, organs such as the heart and lungs, or pathophysiological processes such as immune systems, where the digital replicas and the physical counterparts are tethered with each other. Critical to the HDT is a connector (“bridge”) that links the human and digital counterparts. Sensors on human bodies obtain real-time pathophysiological data and pass them through the bridge to the digital twin. The digital twin runs artificial intelligence/machine learning (“AI/ML”) algorithms on this input and the resulting output is passed via the bridge to the connected human being. This combination of a connected human being, a digital counterpart and the bridge is unique to HDTs distinguishing them from simulations, clones, and digital assistants. HDTs are the prime drivers of precision medicine and personalised care. While the most common clinical uses of HDTs are as yet in cardiology and surgery, as this technology will evolve, new uses of HDT will be explored and will bring about a paradigm shift in medical care. In this chapter we have discussed the technology of HDTs, principles, methods of construction, and use of HDTs. We also discuss key limitations and human ethics related to the HDTs.

1. Introduction

Clinical care is complex and information-driven and the information shared between the clinical providers and the patient/client is asymmetric. At the bedside and clinic, the patient brings in highly personalised information about their physical complaints, their past history that they feel as significant to bring to the attention of the provider clinician, biochemical or radiological parameters obtained at various time points in their lives, often intermittent, while the doctor brings to the examination room, table, or bedside their expertise with years of experience and knowledge gained through rigorous training and continuous study. The information content of the clinician is knowledge they have accrued based on their training and such studies consist of reviews of trials, carefully conducted controlled trials with limited number of participants or studies and reviews of population based epidemiological studies, case series and case studies, and descriptions of pathophysiological mechanisms, and anatomical and structural narratives. All of these are based on at one hand a single patient or hundreds of individuals or simulations, or models. However, at the bedside and clinical care setting, the doctor needs to tailor their interventions to a specific patient. This individual may not necessarily match with the profile of the target population of the studies they have studied and trained themselves. As a result, in traditional medical practice, clinicians conventionally use a one-size-fits-all approach for treatment and diagnosis of their patients/clients. Albeit, every human being is different, this one-size-fits-all approach is at odds with the intended objective of delivering personalised and precision-driven care. Yet a paradigm shift is in the offing that we will cover in this chapter.

The human genome project has brought about a transformation in traditional ways of thinking about health risks, enabling personalised medicine, and has contributed to our understanding of the genetic determinants of health [1]. When linked with social determinants of health and gene-environment interactions, the study of genomics and epidemiology opens up potentials of personalised prevention and precision medicine. Yet the challenge for the clinician/provider is to care for that one patient/client for whom an intervention must be tailored with several unknowns, and how might this be achieved. One way might be to think in counterfactuals [2, 3].

As a thought experiment, imagine you are a clinician and you can access a replica of your patient or client, one on whom you can plan and study outcomes of different interventions and plan “what if” scenarios. Not only that, your best management plan from this replica can be passed back to the patient to update the health state of the patient. This might be unrealistic at one time, but today, a combination of big data, internet of things, and AI/ML makes it possible to digitally recreate such a possibility. This is the concept of the human digital twin (HDT) [4].

When we think of twins, we usually think of biological replicates, behavioural genetics or heritability of traits, the linkage and contrast between nature and nurture in explaining phenotypes. Twins refer to pairs of humans who share either 100% or 50% of their genes, depending on whether they are identical (monozygotic) or fraternal (dizygotic) twins, and they have similar phenotypes. As such, we expect that as their genetic make-ups and physical appearances are completely or partly similar to each other, relative differences in their genes will enable us to settle the debate about nature-nurture contributions to their variation in physical traits. With the advent of whole genome sequencing, genome wide association studies and polygenic risk scores, twin studies and genomics have revolutionised medicine and health care in general.

Even though twin studies help us to understand the nature-nurture debate, for obvious reasons, we cannot “manipulate” or conduct an intervention on one of the pair of twins to study or reflect changes in the co-twin. For example, if an identical twin were to suffer from cardiac arrest while flying and the co-twin were on ground, no intervention to treat cardiac arrest on the co-twin would work for treating cardiac arrest in the other (and you may wonder why), because twins live separate lives. They do not share any physical connection such that changes in one are transmitted to the other. This is where a “digital twin” is unique.

Now imagine we have created a digital replica of the pancreas. Although here we use this as a didactic example, in real life, Facchinetti et al. in 2013 have developed such a tool [5]. Further imagine that we have obtained data about the structure and function of human pancreas from anatomical and physiological sources and descriptors, and using MRI and CT scans of one particular individual, and we have created a three-dimensional functional digital pancreas specific for an individual in silico. On this digital pancreas, we have trained a machine learning model that can sense and mimic release of pancreatic hormones and we have “connected” this digital pancreas with the blood vessels of the real human being whom we have modelled for the purpose of controlling an insulin pump. Such a digital twin would enable a clinician to control insulin flow from continuous and dynamic blood flow and information transfer between the digital twin and the physical twin. So, the utility of a digital twin of an organ is to provide a 24-hour, continuous monitoring of its state, maintain constant communication, and update its status. Other models of HDTs are possible as well as we will see later in this chapter.

Boulos and Zhang have provided an illustrative example of how digital twins work, for example, if the algorithms are tasked with choosing from more than one treatment and selecting the best option out of them [6]. As an illustrative example, imagine an individual patient has an option of three or more treatment choices for a particular disease. We can create, not one, but several (in this case say at least three) identical copies of the digital twin based on the patient profile (“physical twin”) and then test various management options. Here, the physical individual data (genomic, metabolic, sensor data) are connected to the internet and streamed to the digital model. In the digital twins, we are able to test different combinations of treatments and study which of them is likely to yield the best result based on the data that are already fed from the physical twin. Then this treatment is selected for the physical twin. The fact that the best treatment is trained on the digital twin and then transferred to the physical twin represents the reverse direction of the bidirectional pathway that characterises the digital twin.

To sum up, HDTs are a subclass of digital twin (DT) technology used in engineering with other objects such as aircrafts, buildings, machines, and even cities to mirror and connect physical objects and their digital replications. DT is an ensemble of three entities: (a) physical objects; (b) in silico replication of the physical object, and (c) interconnection between the physical and digital object, referred to as the bridge (Figure 1).

Figure 1.

Human digital twin, a conceptual diagram.

An HDT is not the same as a simulation, a clone, or even for that matter, a digital assistant. For example, simulators are used to train doctors about diagnoses and for clinical skills development. But simulators are not connected to a real human being in any way, nor are they meant to bring about a state transition or bring about a change in the human being they simulate. Digital twins are also not clones that mimic or mirror organs. In both situations, the presence of the connector or bridge that allows bidirectional data transmissions differentiates an HDT from either a simulation or a clone.

The rest of the chapter is organised in three sections. We begin with a brief history of digital twins. We then describe the principles of building HDTs. In the final section we discuss key ethical issues associated with HDTs. We end this chapter with a summary of the main ideas and a set of suggested readings for the interested reader.

1.1 Method of the review of literature for this chapter

The method used to write this book chapter was that of a narrative review. Key studies, books, and peer reviewed publications on the topic of “human digital twins” in healthcare were identified by searching the following databases and keywords for the past 5 years (1st January 2019 to 31st December 2023 both dates inclusive, all articles and resources published in English language were included, and those that did not have an English language translation or were not available in English language were excluded from the scope of this review, and no translations were attempted; all studies that were either reviews or meta-reviews or reviews of reviews were included in the narrative review). The databases, dates searched and keywords included were as follows:

• Pubmed/Medline, “human”, “digital” “twin OR twins”, “healthcare”, “medicine” “surgery” (words in title or abstract, date filters between 1/1/2019 and 31/12/2023, English Language, full text or free full text available).

• Google Scholar, “human”, “digital”, “twin OR twins”, “healthcare”, “medicine”, “surgery” (these words were searched in the title or text word in papers and resources in Google Scholar, same date and language filters as above).

• Semantic Scholar, “human”, “digital”, “twin OR twins”, “healthcare”, “medicine”, “surgery” (Semantic Scholar does not have specific filters for title and abstracts, hence, all resources returned as a result of the search were included in the analysis, subject to date and language filters).

Following retrieval of the studies, these were first scanned for their titles and abstracts or summaries to identify if these studies would fit the scope of the review for the book chapter. The resources were then obtained in full text and summarised to derive main themes as described in the rest of the chapter. Further references were obtained from the text included in the full text of the papers and resources and after scanning the reference lists of the retrieved articles/papers/books/book chapters/resources.

2. History, definition and classification of human digital twins

One of the first applications of a DT was with mirror technologies in the 1970s in the Apollo 13 mission. Apollo 13 was a crewed spacecraft mission from NASA launched to land on the moon but the moon landing was abandoned after it was detected that oxygen tanks in a submodule had leaked. A mirror technology was used to rescue astronauts during the Apollo 13 mission; when the air tanks exploded in the spacecraft, engineers on earth at NASA exploited the simulated environment to model and test possible solutions, and successfully devised an improvised air purifier. From earth, engineers instructed the astronauts in the spacecraft how to build the air purifier with materials available in the spacecraft. Besides, using simulations in the mirror environment, engineers on earth finally found a way to bring back the crew of Apollo 13 to earth alive [7]. However, this was not strictly a digital twin although the technology was applied to save lives of astronauts stuck on board. As another example, Kerckhove has written about a video on the knowledge navigator by John Sculley, the then CEO of Apple in 1987, where the video describes a scenario in the day of a university professor. The professor in that video uses a digital assistant that comes to life and provides assistance via videoconferencing in real time, but then a digital twin is not a digital assistant either [8]. A digital assistant can provide real time advice only when accessed but it neither “mimics” the human nor “mirrors” the human’s life, nor is it connected to the human at all times. In 2002, Michael Grieves and John Vickers first proposed the concept of “digital twin” for product lifecycle management as a set of “virtual information constructs” emulating structure, function, and behaviour of physical assets that are dynamically updated with data from their physical twins throughout the life cycle. This refers to a digital representation of an active unique product (real device, object, machine, service, or intangible asset) or unique product-service system (a system consisting of a product and a related service) that comprises its selected characteristics, properties, conditions, and behaviours by means of models, information, and data within a single or even across multiple life cycle phases [9].

According to Pascual, HDTs apply the concept of replica, that is part of DTs, to humans, thus creating digital replicas of either a human or part of a human [10]. HDTs are built by collecting data from various sources, including cameras worn on the body, sensors, wearables, medical devices, and medical records and then these are used to build a digital replica. The replica is then programmatically converted to a model using machine learning (“DT model”). The DT model can then be used to predict disease progression, assess impact of treatments, help in pre-operative preparations, and indeed conduct different types of clinically relevant optimisation procedures.

Although day to day clinical applications, prediction of the courses of illnesses, and optimisation of treatment are useful applications of HDTs, Sun et al. contend that HDTs hold a great prospect for precision medicine [11]. Precision medicine refers to the practice of medicine that takes into account a person’s unique individual characteristics, rather than using the conventional “one-size-fits-all” approach to medical treatment. This combines genomics and big data to optimise treatments. But because medicine is so diverse, this also means that for their use in precision medicine, HDTs need to be discipline specific. It is even proposed that the definition of HDTs need to be discipline-specific. For example, a cardiac digital twin should be differently defined than an oncologic digital twin [12]. Even though the concept of HDTs is intuitive in the sense that a physical system, a digital reconstruction of that physical system, and a channel to connect the two are three components, in reality, the configuration of these three entities depend on what we want the HDTs to do for us. For example, while a cardiac digital twin is designed to mimic the human heart and its electrophysiological states, an oncologic HDT might only mimic a cancerous cell, so in reality, the two HDTs are quite different. This gets even more complex as we attempt to design human beings in their natural or social environments, where, rather than a digital reconstruction of a set of specific anatomical features or pathophysiological attributes, our aim is to represent an “entire” human being as one would be in a natural environment, albeit in a virtual environment. The point being that, context and purpose of various HDTs are very different.

That said, in general, we define two different classes of HDTs. An HDT instance is one where we design an HDT for a simple object that aims to perform one particular function, e.g., a cell or a heart. When we combine several different HDTs together to construct a unified set of functions, we have an HDT aggregate. An example of an HDT aggregate might be that of an HDT that integrates heart and lungs to model the joint functions and structure of the two organ systems.

Another way of classifying HDTs is based on how the physical and digital twins are connected. Chakshu has described HDTs as active, semi-active, and passive. Active digital twins are those where the physical twins and digital twins are interconnected and there is real-time transfer of data between the two. Semi-active digital twins are those where the digital and the physical twins are connected to each other but there is no real-time transfer of data between the two entities but data are stored and forwarded to the digital twin after some time lag. The data in this case usually consists of time series data and the data analysis and updating of the physical twin is not real time, but rather, after the data are collected, stored, and then forwarded after a lag of time from the physical to the digital twin. The digital twin then analyses the stored data, and at that point, passes it back so that the physical twin’s state is updated. The passive digital twins are those where there is no real time update of the physical twin but uses a human-in-the-loop (see below) to update the physical and the digital twins. Although there is no real-time status update of the digital twin, the channel eventually results in some form of modification. Many human digital twins would possibly fall under this category as if, the modelling outputs are then deferred back to the clinician for action, then while no real time update of the human health state takes place, nevertheless, it is up to the clinician to take action based on the model outputs. We will later see in this chapter that this is sometimes ethically admissible. In the next section, we discuss the steps to develop an HDT.

3. Principles of developing human digital twins

3.1 Methodology of developing human digital twins

During development of an HDT, the developer needs to keep in mind the purpose or goal of the twin. Here we provide a simple scheme of the architecture of a HDT (Figure 2).

Figure 2.

Schema of a HDT. The red arrows flow from the physical twin as raw data to the physical twin. The blue arrows flow from the physical twin to the bridge and from the bridge to the digital twin and from the digital twin to the AI/ML tools to signify flow of processed data. The green arrows that flow from AI/ML tools to the bridge and from the bridge to the physical twins signify flow of processed information that are aimed to modify the state of the health of physical twin.

Even though the basic idea of the digital twins remains the same, the exact instantiation and type of the digital twin developed will depend on several factors. These include:

1. The purpose of the twin: for example a twin that will mimic and mirror the electrophysiology of the heart will be very different from that of another twin that will mimic and twin the knee joint and movement of the lower limb.

2. Type of information that will be obtained from the physical twin. This will determine in turn how the digital twin will process this information. For example, if the physical twin has access to the electronic health records and narrative data such as life stories but no other information, then the digital twin will be different compared with a digital twin where the physical twin has access to connected devices, wearables, and mobile devices that record life (referred to as “lifelogging”).

3. Type of the connector, or how information is passed. As Chakshu has stated, depending on whether the model will receive real time data as opposed to data that are stored and forwarded over time, the HDT will be active, semi-active, or passive.

The goal of the first step in building an HDT is to digitise physical assets to their corresponding digital versions. In this step, either a human cell, or an organ, or in some cases, representation of an entire human being is considered to be digitised. The structure of the human cell or the organ is derived from prior literature and features that are specific to the human being are obtained from a range of pathophysiological investigations, biochemical parameters, and radiological data that are unique to the person whose features will need to be converted to digital form. A range of digital tools can be used for this purpose. Unity and Amazon IoT are popular general purpose digital twin models that can be adapted for converting humans and organs to their digital counterparts [13, 14]. OpenCMISS-Zinc (‘Zinc’) is a cross-platform software library for building complete modelling and visualisation applications, from rich model representation to high quality OpenGL graphics rendering made by the Auckland Bioengineering Institute and OpenSim are popular choices for realistic organ modelling. Other approaches of model building include finite element modelling [15, 16, 17].

Cheng states that creation of fine-grained human models in silico is time and resource-consuming, and a way to keep the effort tractable and affordable is to use standard models as templates [18]. The templates are then used to build model instances through various sculpting processes. Another way is to use deep learning techniques to produce 3D human images from the two dimensional images using convolutional neural networks [19]. In this context, deep phenotyping, or fine grained individual data obtained from a combination of sensor data, genomics, and electronic health records is instrumental in creating fine grained human or human organ or cellular representation in silico [20].

In the second step, developers build the connector or the bridge that spans the physical and the digital twins. Here, developers need to base their development on what and how to obtain the data that will be passed from the physical twin to the digital twin and back. If data from the physical twin were to be obtained from the sensors embedded in the physical twin (PT), then using application programming interfaces for the sensors, developers would need to ingest the data into the digital twin (DT) and then populate the database within the digital twin. On the other hand, if the data were to be fed intermittently into the DT by the PT themselves, such as intermittently connecting a mobile device and activating the data transfer, that might indicate development of a different architecture. In clinical settings where patients can have embedded medical devices and sensors, use of Internet of Medical Things (IoMT) is a useful strategy to ingest data from the PT to the DT [21]. The bridge is a two-way bridge, so the developers will need to consider how the results and outputs of the AI/ML algorithms will need to be transmitted from the DT to the PT. This process can either utilise wifi, or bluetooth connectors, or passed via the clinician or the caregiver.

When a human being is literally ‘in the middle’ of the linkage between PT and DT, such as a clinician who will make the decisions about the output of the algorithms, this situation is referred to as HITL (“human in the loop”). Here, a team of humans (clinicians and decision makers) interact with the end users (PT) to predict their future statuses and provide suggestions for improving their wellness in general. The system, or the connector, exchanges data through the Internet and operates in a “semi-automatic” way based on the heuristics and decisions of the experts in the middle.

In the final step of the construction of the HDT, physics are added to the models and various machine and deep learning algorithms are used to process the input data from the PT. In the words of Professor Peter Hunter of the University of Auckland in New Zealand, “Biology is clearly extremely complex but it too has to obey the laws of physics and chemistry, so there is no fundamental reason why we could not build a predictive model of the anatomy and physiology of a human body capable of being personalised and used for disease prevention or treatment.” [22]. We briefly introduce here a few principles of deep learning that can be instrumental for building the predictive models that Professor Hunter has described in his description of HDTs. While a full description of the various deep learning algorithms is beyond the scope of this chapter, the following is an indicative description that may help in understanding the basic idea of the role of deep learning and machine learning algorithms as they are applied in HDTs.

Model construction is at the core of DT in medicine, combining human anatomy and digital technology through image processing, digital collection processing, mathematical modelling and other technologies. This includes mechanics modelling, mechanical network modelling using a system of ordinary differential equations, and statistical modelling as predictive modelling of individual components.

Machine learning (“ML”) refers to the process where a system is fed input data, output data, and the machine then figures out the algorithms that connect the input with the output data (Figure 3).

Figure 3.

A machine learning schema. A machine learning figures out the optimum algorithm that connects input data with output data.

In machine learning, analysts provide the machine with input and output data, and based on the input and output data, the machine decides on the best algorithm to match the input with the output data (Figure 3). Machine Learning can be supervised, self-supervised, or unsupervised. In supervised machine learning, the analysts provide the “machine” both with input (termed “features”) and output data (termed “target”) in the form of labelled target variables and “ground truths” against which the machine learning model learns, and the machine then determines the algorithm. Machine Learning algorithms can predict values of continuous variables, or classify or categorise patients or individuals into clusters, extract information from text data (natural language processing), and generate new patterns from sparse information (“generative learning”). In self-supervised and unsupervised machine learning algorithms, the machine learning process is not provided with labelled data, instead the machine identifies clusters and classifies the data into separate “bins” (Figure 4).

Figure 4.

A schematic diagram showing the relationships between artificial intelligence, machine learning, and deep learning. Note that artificial intelligence (“AI”) is a superset of machine learning, and machine learning is a superset of deep learning as approaches to solve real world problems.

Deep learning is a subset of machine learning modelled after neuronal transfer of information (hence termed as “neural network”), where nodes and edges connect to form a network. A network can be dense or shallow depending on the number of “layers” between the input and the output layers (Figure 5).

Figure 5.

A schematic diagram of a deep learning model showing artificial neural network (ANN). X = feature, w = weight of the edge that connects the input nodes to the hidden nodes, y = target. At each node a sum of the weighted X variables, i.e., X multiplied with w, are then subjected to an activation function to determine the value of the downstream node. The network iteratively adjusts the weights of the edges of the network to minimise the difference between the predicted values at the output layer and the “ground truths” or true value.

The network “learns” using a range of forward and backpropagation algorithms [23]. As with machine learning, deep learning can be supervised, semi-supervised, or unsupervised. The exact deep learning algorithm to run depends on the goal of the deep learning and input sources. If the task is to build predictive models of continuous variables such as glycated haemoglobin (HbA1c), a measure of control of blood sugar level averaged over a period of 3 months, an artificial neural network (ANN) is selected to output a predictor that has a continuous variable. On the other hand, if the purpose is to do image processing tasks, where images are either clustered or used for prediction of patterns of images such as clinical diagnoses from an X-rays or CT scans, use of mathematical concepts referred to as “convolutions” or convolutional neural networks (CNN) are used. Here, the complexity of the images are reduced using matrices and filters. Where input data are fed in the form of time series, such signal processing, electrocardiogram, or electroencephalogram, recurrent neural networks such as long short term memory (LSTM) networks are used to decode and analyse the data. For a comprehensive review of different deep learning algorithms, see Goodfellow et al. [23].

In summary, in order to develop an HDT, one needs to start with a well-articulated objective that the DT is designed for and will achieve. This will drive the design of the digital twin and the data needs. A combination of big data, whether obtained in real-time or intermittently, and the type of data will determine the choice of the algorithms that will be run on the digital end of the digital twin itself. Finally, the connector needs to be built either over Internet of things (IoT) or connected devices that can transmit data from the PT to the DT via bluetooth, local wifi connection or over the Internet. Alternatively, the user can input data directly to the DT database, which then returns the results to the user either directly or via providers (HITL). A number of approaches and designs are possible in building an HDT.

For example, Bjelland et al. [24] describe their building an HDT of the human lumbar spine for the prediction of the biomechanical properties of a real lumbar spine for different human postures. They obtained realtime motion posture and spatial position of the human body using human-motion-capture technology and calculated the lumbar posture using a wearable virtual-reality device and sensor data. Finally, they modelled the HDT of the lumbar spine using finite element modelling [24].

Barricelli et al. have described development of SmartFit, an HDT for sports and human performance for athletes using IoT [7]. In building the HDT, they captured, for each athlete, a set of measurements describing the athlete’s behaviour over a period of time during which they collected using wearable sensors the athlete’s heartbeat, number of steps, physical activity, sleep, and food or mood logging. The HDT allowed the coaches to define heuristics that would be triggered during specific contexts. In turn, these heuristics enabled notifying the athlete whose behaviour needed to be corrected in real time [7, 25].

Chase et al. studied use of HDTs in intensive care units (ICUs) where they found use of bidirectional flow of information through connectors were well utilised for control of glycaemia, and identified that cardiovascular care, sedation and mechanical ventilation were candidates for HDT applications in ICU settings [26].

In summary, for both building and interpretation of studies on HDTs, it is essential to evaluate the purpose of the HDT but also several aspects of the modelling strategies. In particular, the process of ingesting the data and establishment of the connector between the digital twin and the physical twin. As the devil lies in the details, while these have the potential to significantly impact advanced surgical and medical care processes, these are fraught with limitations and risks, in particular, ethical issues associated with HDTs need to be carefully considered, to which we turn now [27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37].

4. Limitations and ethical issues with human digital twins

The promises and limitations of HDTs pertain to personalised or precision medicine. In personalised or precision medicine, an individual’s genomic, pathophysiological, socioeconomic and other data are taken into account at individual levels [2]. This unified approach, where individual data are taken into account rather than as an aggregate, is possible because of deep phenotyping. The machine learning and deep learning models, in turn, are trained and validated on population based or large data sets of individuals. Model testing uses individual-specific data to either predict a value for that individual or to classify that particular individual into a cluster. These predicted values and clustering are then used to determine the next course of action. This next course of action could be a treatment or it could be used for classifying the human being into a predetermined category.

While this process can be conducted in isolation using any machine learning procedure, these are integrated in the context of HDTs where the physical and the digital twins are interconnected. In HDTs, the connector ensures an inflow and outflow of data either continuously or intermittently between the physical twin and the digital twin. In doing this, the HDT can end up being a near automated mode of controlling health decisions that can be then devolved to a machine rather than to a human. As an illustrative example, consider the case of an automated blood glucose monitor and insulin pump system where the HDT monitors the state of blood glucose and adjusts the flow of insulin according to the flow and algorithm to maintain an optimum level. As HDT technology will be widely adopted, it will have the potential to lead to an ubiquitous tendency of using engineering solutions to address health issues, often, with minimum need for human intervention. This then raises a dilemma in medical care or clinical practice where traditionally, we expect a direct contact between the doctor/clinician and the patient/client to be our mainstay of care if not the norm. From this perspective of clinical care, with the emergence of HDTs in clinical care, we need to be mindful of at least four considerations as limitations of HDT and can potentially pose ethical challenges [38].

First and foremost, HDTs are essentially based on an engineering paradigm derived from digital twins for machines such as aircrafts and turbines, but human beings are not inert mechanical devices or agents. Humans live within social contexts and have complex biochemical, biophysical, biomechanical needs and uniqueness that must be considered. Braun has questioned the meaning and representativeness of “human” in HDT, and about correspondence between the individual who is represented in the digital form or avatar and the physical person. Braun has argued from the point of “agency” that an individual who is represented in the HDT should be able to control what aspects of the digital self can be manipulated for machine learning or modelling. According to Braun, using the cases of predictive modelling for instance, any implication of the findings and predictive modelling must be conducted in collaboration with the owner of the physical body with full consent [39].

Secondly, as engineering approaches to solve human health issues in clinical care have been traditional in clinical care (for example use of heart valves, and orthopaedic prostheses), it is perhaps natural that cardiology and orthopaedics are the two disciplines that have been the early adopters of HDT. Even so, this raises an interesting issue around how we define what is normal with respect to HDT. Bruynseels et al. have argued that even while adopting an engineering solution, we need to be mindful of the granularity of the models and how we define the normal [40]. As HDTs are predicated on DTs where it is possible to define the digital artefacts precisely, in HDTs, and particularly, models in medicine are partial and “coarse grained” because we do not have complete information on every variable at an individual level.

This also brings to question the meaning of what is normal and the benchmarks against which treatments and preventive health will need to be assessed if HDTs go mainstream. Usually, when we think of illnesses, particularly from the perspective of biochemical markers (for example, we consider a HbA1c of 65 mmol/mol as deviating away for normal for blood sugar control and as a marker for cardiovascular risk and clinicians would aim to bring down a person’s glycated haemoglobin level to about 48 mmol/mol for optimum control or thereabouts), such measurements are based on population based studies conducted on hundreds of thousands of individuals and normative values are derived from such studies. But the patient or the client at hand is a unique individual and the values may have varied and remained at a level that is “normal” for that one person “over a period of time” that is not captured in population based studies. Our current position of how we define normal for an individual is based on population based studies because we do not have sufficient fine grained information collected longitudinally for an individual over the individual’s lifetime. With the advent of HDT, this scenario may change. If and when that happens, and as normal is “redefined” at a personal level, Bruynseels argue that this will in turn lead to a redefinition of health across life span and this distinction between health states and what can be “enhanced” for individuals with digital twin technology across an individual’s life span calls for a moral debate, particularly when “enhancement” is defined in terms of increased “abilities” of an individual.

Thirdly, HDT technology has the potential to further worsen the already existing problems of inequitable access to healthcare and risks of breaching security of information. Popa et al. have considered a socio-ethical scan of the digital twin technology in healthcare, based on their qualitative research on how stakeholders of HDT have viewed the advantages and limitations of this technology in society [13]. An emergent theme from their research is that the participants were concerned about the leakage of breach of their personal information as HDTs are essentially personal longitudinal data troves, so leakage or “hacking” of that information will jeopardise personal security. A second and related issue was that of the resulting inequitable access to care. Equitability of access to care denotes where socioeconomic variables explain how one accesses health services or utilises them, rather than their biological or assessed clinical status. Popa et al. have argued that HDT will have the risk of worsening the already fragile inequitable access to health care both within countries as well across the world between those regions that can afford the technology and costs of establishment of HDT frameworks and those, particularly in the global south, that may not be possible for them to afford this technology [13].

Finally, but not the least, the benefits and risks of HDT technology will not be the same for all age groups, in particular, for children. Drummond and Coulet have stated from the perspective of paediatric care that, considering smartphones as near ubiquitous tools for collection of digital data, penetration of smartphone usage are still low among the children and young adult population; they estimated that about 75% of the young population worldwide have access to smartphones but the issue goes beyond the extent of access to smartphone [41]. Drummond and Coulet have argued that with children and young people and HDTs, use of smartphones might be what they termed as a “double-edged sword”. On the one hand, access to smartphones or ubiquitous access to smartphones would be beneficial to the children who will need the use of such devices to constantly update their digital twins and would then make way for more robust application of HDT technology for healthcare management. On the other hand, widespread use of smartphones among children would also lead to an increase in the risks of addiction. In the end, smartphone devices may provide some benefits but they are also associated with worsening behavioural risks. As an alternative to smartphones, wearable technology, rather than smartphones alone, might be considered to collect continuous data. This would then, as the authors argue, lead to other unanticipated issues such as those pertaining to the size of the devices. As children grow, they are likely to outgrow the size of the devices implanted or made available, and there would also be issues around calibration of the devices as children’s biological parameters change as they become more mature. Besides, children are at higher risks than adults to lose or damage the connected devices, or even “eat” or ingest them. Besides, in case of discrepancy between a child’s self-report of clinical illness and that reported by the device on body, it might be possible that the adults, on whom a child is dependent, could consider that the digital device is more accurate than the children themselves, leading to further complications and risk of mistreatments, or subject to automation bias. For all these reasons, age appropriate provision of devices that are capable of interfacing with HDT technology is likely to be challenging [41].

In summary, HDTs will likely bring about a paradigm shift in the way medical care is delivered, but they are also at risk of inheriting the various risks and disadvantages that an overly engineering approach to solve biological problems will ensue. Besides, as HDTs rely on a high degree of connectivity (either with bluetooth, wireless signals or provision of Internet and preferably “always-on” availability of internet connection), these will have their own associated issues, for different age groups as well. For example, children would have unique issues with HDTs and devices of their own. HDTs are also likely to worsen the North-South divide in terms of achieving health statuses, since disparity of access to technology will be a limiting factor as to which section of the population and which countries will have better access to HDT technology. We also do not know what health impacts will constantly worn and implantable devices that will interface with the HDT technology may have. For all these reasons, while we are optimistic that HDT will indeed usher in a paradigm shift in medicine and surgery, we will also need to keep a close eye on the limitations and issues around human ethics.

5. Conclusion

In conclusion, human digital twin technology (HDTs) are about to bring about a paradigm shift in medicine and surgery by enabling a shift towards personalised medicine or precision medicine and prevention, using an engineering approach that has already transformed several other industries such as aviation and manufacturing. Much of the requirements that would enable this technology to be ported to medicine and surgery are already present, and emergent applications in cardiovascular care, intensive care, and surgery (cardiothoracic surgery, orthopaedics, skull base surgery) are already in place. Besides, using genomics, deep phenotyping, AI/ML, generative learning, and combining with longitudinally collected personal level data over a person’s lifetime, it will usher in a paradigm shift about what is “considered” to be normal from a population health perspective as opposed to “normal” in a person’s lifetime. In turn, HDTs would enable them to address several “wicked” healthcare issues, such as the problems and challenges of real-time monitoring of a person’s individual health status, dynamic, on the spot analysis and precise treatment for diseases, which have been traditionally challenging with current approaches. HDTs would also make it possible to select precise treatment options for the individual patient by using computer algorithm-based methods and principles in bioinformatics, a holy grail for precision medicine. Besides, HDTs would enable patients to exercise a greater degree of autonomy and, when successfully applied over a large section of the population, will achieve equitable treatment for patients irrespective of race or gender.

At the same time, concerns abound their implementation and the devil is in the details. Success of HDTs depends to a large extent on the accuracy of simulation and the models that are used. As machine learning and in particular, deep learning and generative learning algorithm tools continue to evolve, the computational complexities will evolve as well, making older approaches redundant. What effect this will have on those individuals with older devices and sensors and those with limited resources to keep up with technology is uncertain. Notwithstanding remarkable improvement in models and algorithms, technical hurdles to creating high-fidelity models remain, including the multi-scale and multi-physics nature of the models, and the difficulty in linking sub-process models across scales and physics, and the scarcity of experimental data. As we noted, HDTs will have to deal with issues around privacy and maintenance of individual data and breaches, and what ethical connotations they will have, so checks and balances will need to evolve over time.

Overall, the future prospects with HDTs are optimistic. A major contribution of HDT would be the ability to track an individual’s life journey, using data collected by wearable sensors and the lifestyle registered by the individual for the transition from clinical medicine to preventive medicine. In future, individuals will have a full-lifecycle DT body, where data are collected from birth to form a virtual twin, which will grow with the child and serve as a life-long health record or medical experiment object. This is unprecedented in the history of human medicine but that possibility is closer now than ever. For health systems, using the DT body and epidemiological big data, the healthcare systems will be able to perform real-time monitoring of the patient’s health status and predict the risk of disease at a fine grain personal level. However, optimism grounded in technological advancement should also be tempered with the ground realities of issues around humans being humans and about issues around privacy, morality, ethics, and equitable access to health care, all of which deserve serious considerations and deliberations.

Acknowledgments

The author also acknowledges the Faculty of Health, University of Canterbury, for supporting the work presented here. The author did not receive any funding from any source for writing this manuscript.

Conflict of interest

The author declares no conflict of interest.