Perspective Chapter: Making the Shift to Personalized Preventive Medicine with Human Digital Twins

Abstract

Human digital twins (HDTs) have the potential to support a paradigm shift from one-size-fits-all sick care to highly personalized preventive healthcare. By providing important context for complex disease processes and enhancing our understanding of the dynamic interactions that lead to non-communicable diseases (NCDs), HDTs are poised to offer researchers, care providers, and public health agencies the toolset they need to predictively diagnose and treat NCDs with highly customized interventions. The precision health knowledge gained from HDTs can help patients understand their NCD risks, public health authorities support care pathways that effectively prevent or delay the onset of chronic diseases, and care providers prescribe interventions based on an individual’s unique biological, behavioral, and environmental characteristics. This chapter presents key human digital twin concepts and model performance evaluation criteria. Digital twin applications in preventive medicine research, clinical care, and public health are presented while acknowledging the associated challenges, including model robustness and ethical concerns surrounding the use of digital twins to model humans.

1. Introduction

Our current approach to healthcare is unsustainable. Favoring one-size-fits-all reactive sick care above preventive healthcare creates societal and economic burdens. Globally, NCDs, including heart disease, stroke, cancer, diabetes, and chronic lung disease, are the leading causes of death and disability. According to the World Health Organization (WHO), NCDs are collectively responsible for 74% of all deaths worldwide, resulting in 15 million people dying between the ages of 30–69 years each year [1].

Even though an estimated 80% of NCDs are preventable [2], the global NCD burden is expected to increase by 17% over the next 10 years [3]. The coronavirus (COVID-19) pandemic has amplified the global burden of NCDs. Based on available data from multiple countries, it is evident that people with underlying NCDs such as hypertension, diabetes, cardiovascular disease, chronic respiratory disease, and cancer have a higher risk of severe COVID-19 disease and are more likely to die from it [4].

Individuals, families, businesses, governments, and economies, directly and indirectly, bear the brunt of NCDs through high treatment costs and productivity losses caused by premature mortality, early labor force exits, absenteeism, and lowered work capacity [5]. The cost of NCDs to healthcare systems currently accounts for almost half of the general hospital expenditures in most developed countries [6, 7]. Over the last two decades, healthcare spending on diseases like cancer [8] and diabetes [9] has increased faster than disease incidences.

Despite global commitments to control NCDs, preventive care is underfunded, and nations lack the policies and interventions to effectively prevent and manage chronic diseases. In the U.S., Europe, and other wealthy countries, preventive care averaged 2.4% of total public and private healthcare expenditure [10]. As with any medical expenditure, the value of prevention is judged based on the benefit it provides, measured by improved quality of life, productivity, or both. To improve preventive medicine’s cost-to-benefit ratio, control programs should focus on pre-onset of disease and pre-pathogenesis activities. Spending on secondary and tertiary prevention occurs too late—after destructive physical and mental effects have impacted the patient’s productivity and quality of life. Once an NCD is diagnosed, treatments are generally less effective, more complex, and more expensive. As example, treatment for cancer patients diagnosed early is 2–4 times less expensive than treating people diagnosed with cancer at more advanced stages [11].

Identifying at-risk patients before disease onset using current state-of-the-art protocols is difficult. Uncertainty persists when translating research data into evidence-based precision medicine because the average outcome of carefully chosen study participants may not represent a specific patient—especially in complex and ambiguous cases. Even with current NCD behavior modification consulting and screening programs, doctors commonly cannot diagnose or treat NCDs until symptoms appear. Sometimes, disease diagnosis is further delayed if symptoms are common to less severe conditions or the case is complex due to multimorbidity or other compounding factors.

Precision medicine strives to reduce inaction or waste by applying more granular population stratification to disease control programs. But without causal knowledge, filling the gaps between epidemiology research and clinical practice is challenging. The results of epidemiological studies designed to determine patient risks for common chronic ailments such as cardiovascular, cancer, diabetes, and neurodegenerative diseases tend to produce marginal, contradictory, irreproducible, or hard-to-interpret results [12].

Evidence suggests that more screening does not always translate into fewer deaths [13]. For example, over 1300 women aged 50–59 need a mammogram to save one life [13]. Some screenings lead to relatively small reductions in mortality. One study found that lung cancer screenings yield a 0.4% reduction in patient mortality rates [13]. Further screenings can lead to unnecessary follow-up screening tests, anxiety caused by false-positive results, and strains on resources due to the overdiagnosis of diseases [13]. Studies show that 19% of screen-detected breast cancers [14] and 20–50% of screen-detected prostate cancers are overdiagnosed [15].

Population-based statistics can also lead to inaction due to personal beliefs, attitudes, or fear of a diagnosis. For example, research findings suggest only 10–20% of smokers will develop lung cancer. Patients may choose to continue smoking and avoid lung cancer screening programs—believing they will be in the 80–90% of the population who smoke but do not develop cancer.

With innovations, such as genomics, slowly getting integrated into care, the complexity of NCD knowledge is only increasing. At the same time, the acumen of doctors is decreasing because their longitudinal experience with patients is decreasing. Which means doctors do not see their mistakes play out.

Over the last decade, many within the medical community have recognized the need to complement traditional research with system-based approaches to support the goals of personalized preventive medicine [16, 17, 18, 19, 20, 21, 22]. Many theoretical frameworks have been proposed, and yet, most practical progress is still achieved through reductionist processes that heavily rely on statistical data to build models that explain and discover functions at a single system level (e.g., gene, protein, cellular, tissue, or organ level) [16]. Using the resulting circumstantial data as inputs for machine learning (ML) and large language models (LLM), only introduces more, not less, uncertainty in preventive medicine due to standard errors, data integrity issues, model bias, and unexplainable results [23].

Human digital twins (HDTs) support a paradigm shift to personalized right-time preventive care by adding the missing context of complex disease processes and promoting further understanding of how dynamic interactions between thousands of determinants impact a patient’s susceptibility to disease. Such efforts aim to build an iterative process that combines epidemiological research with digital twin technologies to deepen the understanding of disease progression and cognitively discover promising new areas for scientific research.

The precision health knowledge gained from HDTs can be used to extrapolate predictions about disease behaviors not covered by conventional data-based deductions. Scientists can use HDTs to accelerate the translation of research into effective primordial and primary precision preventive protocols. Representative virtual models of patients can help doctors make clinical care decisions based on a patient’s unique circumstances and predicted health risks. Explainable, personalized health metrics can empower people to take control of their health. For public health authorities, HDTs can prove the benefits of preventive healthcare programs, inform policy choices, and support highly deterministic clinical pathways characterized by high reproducibility of care.

2. Causality is key to personalizing prevention

NCDs are dynamic processes that evolve through environmental factors, genetic predisposition, disease agents, and lifestyle choices. Interrelations among the components of health often include dynamic challenges, such as feedback loops and changes over time that cannot be understood using linear or reductionist paradigms. Often, two or more patients with the same disease are very different regarding causes, evolution, and temporal conditions. These dynamic differences are illustrated in COVID-19 patients and other autoimmune diseases. Further, epidemiological research has demonstrated the strong connections between seemingly disparate diseases, as evidenced by the number of patients who share diagnoses with two or more “unrelated” disease processes [24].

Complexity science asserts that to develop a more comprehensive and complete understanding of the whole, it is necessary to reinforce the importance of understanding how each part interacts with all the other parts and emerges into something new. A key concept of systems theory is that the whole is greater than the sum of its parts—meaning that when holistically examining how smaller systems come together to affect the entire complex system, specific characteristics of the whole cannot be easily explained or rationalized when looking singularly at any one of its parts.

Therefore, the goal of precision preventive medicine cannot be fully obtained without a method to model a human’s health dynamics, constraints, and conditions and then elucidate principles (such as purpose, measure, methods, and tools) in ways that can be applied to understand and predict NCD behaviors for other humans. This requires the formulation of predictive mathematical and computational models that cover the functional, graph connectivity, and regulatory networks and are as complex as necessary to account for specific details while being computationally tractable [16]. Figure 1 summarizes the differences between the current consequential versus the proposed circumstantial approach to epidemiology.

Figure 1.

Difference between consequential vs. circumstantial approach to NCD prevention.

To reproduce system-wide behavior and uncover the origins of known as well as unknown behaviors, three factors must be considered: (1) context, which covers the values that represent all components related to the process being studied; (2) time, which considers the changing characteristics of each component; and (3) space, which accounts for the topographic relationships between and among components [19]. This approach naturally leads to the development of mechanistic models that support the analysis of one patient at a time instead of averages across large patient populations. If a model combines a large group of people, many characteristics that make an individual unique must be ignored; otherwise, the solution becomes computationally intractable.

2.1 Hierarchical model

The mechanisms of human health are hierarchical in that lower-level dependencies influence and decompose higher-level behaviors. Therefore, to adequately draw quantitative predictions, the model needs to capture these hierarchical relationships horizontally within and vertically between multiple levels. Further, health is not a closed system. At any point, a patient’s health is influenced by external factors such as the environment, individual behaviors, access to healthcare, and socioeconomics (See Figure 2).

Figure 2.

Simplified diagram of human health determinants and biological causal relationships.

2.2 Nonlinear dynamics

A nonlinear system is defined as a system in which the change of the output is not proportional to the change of the input. Natural and living systems’ flexibility and ability to adapt to and cope with different conditions originate from this nonlinearity [25]. As such, nonlinearity plays a fundamental role in the performance, functionality, and evolution of health. Consequently, understanding nonlinearity is vital to accurately predicting the pre-onset of diseases and pre-pathogenesis conditions for an individual.

Since the 1800s, Claude Bernard and then Walter B. Cannon popularized homeostasis as a guiding principle of medicine—emphasizing the body’s remarkable ability to maintain stability and constancy in the face of stress [26]. From this point of view, illness is defined as a failed homeostatic mechanism, and treatment aims to correct health failures by reestablishing parameters within a normal range. As such, the homeostasis paradigm has placed a significant emphasis on static stability (i.e., normal ranges) and not on dynamic stable states, such as oscillatory [27] or chaotic behavior [28, 29, 30, 31, 32].

Further, homeostasis implies that human health operates as a closed-loop system when in fact, human health is better defined as an evolutionary system, which perpetually changes over time in response to interactions with the environment and surrounding circumstances. The evolution of health (or disease) takes multiple forms, and the predictability of system behaviors is difficult to explain by linking a cause to an observed effect when using historical statistical data, which was only valid when it was captured. In many medical models, data extraction, e.g., obtaining serum glucose level or blood pressure, creates a loss of time, space, or context information. This leads to the loss of rich information that, if captured, would help improve understanding of the systemic and dynamic behavior of the human body [19]. Disease progression and mutations provide examples of chaotic behaviors that are difficult to prevent and/or treat without the concept of nonlinearity in medicine.

Failure to include dynamic states in models can often lead to treatments that are either ineffective or detrimental [19]. Many scientists and clinicians have largely ignored that less intuitive treatments may yield more effective outcomes or that the correction itself may invoke harmful system-wide effects [33, 34, 35]. To address this shortcoming, it must be recognized that the behavior of a nonlinear system can be sensitive to the system’s parameters and initial conditions. More specifically, by changing a nonlinear system’s parameters, the system’s behavior can change qualitatively, as well as quantitatively—and often dramatically, especially in the case of chronic diseases.

Studies have shown that chaos is a widespread phenomenon throughout the biological hierarchy, ranging from simple enzyme reactions to entire ecosystems [36]. Mutation demonstrates the nonlinearity of health. For example, variants of SARS-CoV-2 are caused by a mutation at the protein level [37], and mutations in the DNA sequence contribute to cancer progression [38]. The dynamics associated with a mutation can cause a patient’s health to become unstable. Therefore, the goal of preventive medicine should be to sufficiently understand chaotic, evolutionary, and transformative processes so that it becomes possible to predict their occurrence and control the associated risks.

2.3 Identifying the right mathematical solution

Auffray et al. assert that most mathematical theories are unsuitable to deal with the nonlinearity of health because they do not fully address the space and time scales characteristic of biological systems and cannot make experimentally testable predictions relevant to fundamental biological questions [16]. However, even though some aspects of a nonlinear system’s dynamic behaviors can appear counterintuitive, unpredictable, or chaotic, the behavior is not random. Therefore, the right mathematical expression, paired with adequate solvers and supporting digitalization technologies, should support the prediction of system behaviors with a high degree of accuracy [30, 31].

A mathematically robust solution is necessary to build a representative model because the variations of parameter constituents of NCDs are considerable. To support a move towards systems medicine, many proposed models use network principles [39] that describe relations and behaviors of elements that reside in a common spatial level (or layer) of the system being studied [40]. Further, the reliance on massive amounts of data, Bayesian logic, machine learning, and numerical solutions introduce a priori knowledge, uncertainty, spatiotemporal dependence, and/or computational challenges [23].

Mechanistic mathematical models provide a viable way to overcome the limitations of statistical and numerical solutions. An added advantage of a mechanistic approach is that the resulting model can be used repeatedly across multiple medical applications, whereas statistical and numerical models must be continually reconfigured to cover any change in time or space. Typically, mechanistic models within medical sciences have been perceived as too simplistic to cover the full scope of complex systems and sub-systems that represent health [41]. However, other sciences, such as elementary particle physics and gravitational mechanics [42, 43, 44], have shown it is possible to use mechanistic mathematical models to cover a full hierarchy of perturbed graphs as required to discover the underlying dynamics of systems—even from sparse or noisy experimental data [45, 46].

A systems-based framework and perturbation mathematical foundations can add the missing context of complex disease processes and promote further understanding of how dynamic interactions between thousands of determinants, such as DNA [47], pathogens [18], autoimmune responses [48], metabolism [49], and aging [50] influence an individual’s health.

2.4 Why use a mechanistic model?

A mechanistic mathematical model can be defined as the mathematical description of the elements forming a system, their mutual interactions, and their interaction with the environment [51]. Mathematical modeling, emulation, and optimization are already used in systems medicine, and mechanistic modeling is currently the main focus [51, 52, 53].

The purpose of a mechanistic model is to mimic real-life events by making assumptions about the underlying mechanisms and refining the model until the desired accuracy is achieved. Typically, this involves constructing mathematical formulations of causal mechanisms and using analytical tools to determine whether the range of possible input-output behaviors predicted by the model is consistent with experimental observations.

Mechanistic modeling relies upon a two-stage process: (1) a subset of the available data is used to construct and calibrate the model, and (2) in a validation phase, further data is used to confirm and/or refine the model, thereby increasing its accuracy. The advantage of using a mechanistic model instead of a data-driven or machine learning/black box model is that the mechanistic model can address the complexity of human biology and the lack of detailed information about biological system elements by supporting what-if scenarios as well as the ability to confirm the representativeness of the model through targeted emulation exercises and randomized experiments.

When modeling a system as complex as human health, the modeler lacks empirical data to explain critical phenomena through which independent variables interact to produce complex and synergetic nonlinear effects. Statistical methods attempt—without success in most cases—to develop a fundamental understanding of the root causes that impact the most crucial factors within a system using probabilistic treatment of production data or experimental results.

Finding mathematical solutions that enable modelers to improve the certainty of predictions and cover the unknowns is critical to advancing the goals of many areas of scientific study, as well as artificial intelligence and machine learning. This is especially true for decisions that may have a profound impact on human life and well-being. It is, therefore, essential to have mathematical solutions that can expose and define the appropriate mitigating actions for scenarios that may not be represented by experience or historical data.

2.5 Combined use of graph and perturbation theory

Graphs are mathematical structures used to model pairwise relations between objects. Graph theory provides a mathematical nonlinear data structure capable of representing various kinds of physical structure—consisting of a group of vertices (or nodes) and a set of edges that connect the two vertices. In practical applications, vertices and edges of graphs often contain specific information, such as labels or weights.

Many real-life scenarios are better modeled by time-dependent graphs when sequences of time-dependent elements activate the edges [54]. For instance, in bioinformatics networks, graphs reflect the similarity and regulation of biomolecules, such as proteins, genes, and enzymes [55, 56, 57]. The connections in biological functions are not always active, but the status may change over time [58, 59].

The proposed perturbed graph solution’s structure is like those applied in quantum and celestial mechanics [60]. As such, this method can be considered a generalized approach that provides the foundation to find an acceptable approximation for an exact, unperturbed solution. The solution is complex because the problem at hand is complex—both as a product of dynamics/nonlinearity and because the determining factors of health behaviors may be found anywhere in the hierarchal structure of the human being studied. Horizontally, dependencies may drive behaviors; vertically, behaviors may result from direct and indirect causes. Finding these causes is sometimes more important than finding the unperturbed solution itself.

Using perturbation theory and the associated mathematical equations as the basis for building a mechanistic model of patient health provides an analytical framework that allows practitioners to build an approximation method for separable structures and small divisors. The small divisors generally appear late in analytical expansions as inequalities that seem trivial but can produce significant contributions to the final solution. The small divisors often represent chaotic behaviors missing from traditional research but are nevertheless vital in understanding the causality of complex disease processes. Therefore, new insights about their influence on health can be gained by identifying these small divisors and measuring their effect on the perturbed solution.

Compared to alternative modeling methods, perturbation theory provides the framework necessary to capture and reduce the complexity and maintain computational control of the equations. It is, therefore, possible to represent the entire spatiotemporal evolution of disease processes efficiently and accurately, such as multimorbidity, without creating ill-posed problems or non-physically acceptable solutions.

3. Personalizing prevention with HDT

A digital twin, also called a virtual twin, is a software representation of a real-world asset, system, or process designed to detect, prevent, predict, and optimize the system being studied through mathematical or statistical analysis. Many industries are using digital twins to improve the performance of real-world systems through computational analysis of a virtual counterpart [48]. Human digital twins (HDTs) extend the concept of digital twins to cover the mathematical replication of human health process dynamics.

HDTs built from graph-based perturbation theory can encourage a better understanding of how thousands of dynamically complex determinants interact horizontally and vertically across multiple scales to influence an individual’s predicted NCD risk or effectiveness of interventions. When used as an experimentation and decision-validation platform, HDTs may offer researchers, care providers, and public health agencies the toolset they need to predictively prevent NCDs through highly customized interventions.

3.1 HDT requirements

To support accurate prediction for an individual, HDTs must start with a representative model that accurately captures the dynamics of health based on the following criteria:

• Patient-specific: To factor in how genes, autoimmune responses, environment, and lifestyle choices may influence morbidity for an individual patient, HDT models should be computationally efficient and robust enough to analyze many-to-many causal relationships between millions of interdependent and time-sensitive biological and nonbiological health factors.

• Accurate prediction of dynamic behaviors: HDTs should allow users to calculate a future event precisely without the involvement of randomness. Deterministic models can limit uncertainty in clinical applications because they provide a mathematical representation in which every variable alters according to a mathematical formula. This makes it possible to trust predictions about behaviors not covered by conventional data.

• Reproducible: The model should always produce the same output from a given starting condition or initial state. Reproducibility makes it possible to repeatedly use the model across multiple medical applications, whereas probabilistic models must be continually reconfigured to cover any change in time or space.

• Explainable: Finding correlations is not the same as proving causation. ML based on statistical or correlation studies bypasses the need for causality and focuses exclusively on prediction. This leads to unexplainable results that are difficult to trust for critical health decisions. Mechanistic models focus on the causality of input-output relationships. This provides a way to explain the underlying dynamics of health—even when starting from sparse or noisy experimental data [45, 46].

3.2 Iterative HDT methodology

A symbiotic relationship between mechanistic models and machine learning can help accelerate the accrual of medical knowledge [23] and support the use of HDTs in research and clinical applications [61, 62]. Synthetic data, also called algorithmic cognition, derived from HDTs can be used by machine learning algorithms both as transient inputs and as validating frameworks, while machine learning can be harnessed to improve the scalability of mechanistic models (see Figure 3).

Figure 3.

Pair observed knowledge with algorithmic cognition to accelerate the accumulation of epidemiology knowledge.

Algorithmic cognition or synthetic data combined with robotization will provide the flexibility necessary to account for the evolutionary process of morbidity and move precision preventive medicine from concept to value in research and clinical applications. For instance, ML can be used to query a patient database to predict which existing immunotherapy treatment will be most effective in preventing cancer for an individual patient based on past observations, but intrinsically, a learning algorithm cannot suggest new treatment protocols or accurately predict the outcome of new treatments [23]. When machine learning was used to predict the success rate for endoscopic third ventriculostomy as a treatment for hydrocephalus [63], the algorithm could predict the success rate of the actual procedure, but it was not able to consider general physiological variables that may pose a risk for a particular patient and predict a more favorable outcome using a different procedure. In new or complex situations, the deductive capabilities of HDTs are needed to extrapolate behavior predictions that are not present in the original data [64] and consider all the specific dimensions for a particular patient versus the averages of a population of patients.

HDTs provide the framework necessary to holistically compare the patient’s current state, future risks, and available treatment options that support the optimal outcome for a given patient. Figure 4 outlines the 6 steps used to digitally transform the mechanisms of health into an HDT that covers the full hierarchy of graphs necessary to gain accuracy and certainty in behavior predictions that can be used for discovery and experimentation. The outcome of each cycle supports an improvement process whereby HDTs reveal the missing data, which can be vetted through randomized experimentation and data analysis. Then, once the algorithmic cognition is confirmed, ML can cover individual patient scenarios more exhaustively in broader research and clinical applications.

Figure 4.

Iterative analysis methodology for multidimensional medicine.

3.2.1 Step 1: deconstruct components of health

The first step to building an HDT is deconstructing human health using causal deconstruction theory [65] and network mapping methodology. This is necessary to understand the constituent components of health and their dependencies. The goal is to build a directed graph that maps the interdependencies, topology of structures, justification of choices, operational constraints, modes of operations, and data necessary to discover the hidden structures that form over time. Figure 5 illustrates an example of a directed graph covering the complex interdependencies carried by an individual parameter as possible perturbations of multiple interacting networks for patient health.

Figure 5.

Example of generic health directed-graph vertices and edges.

3.2.2 Step 2: build a mechanistic model

From this knowledge, a generic HDT can be built using a perturbed graph solution (see Figure 6) [66]. The HDT should cover previously validated morbidity diagnoses, remediation, mitigation actions, and any influencing biological, public policy, healthcare, habits, or social and environmental factors that improve the HDT’s ability to reliably predict NCD risks and identify possible interventions.

Figure 6.

Example of variable organization and perturbed representation of patient health.

3.2.3 Step 3: perform scenario analysis

Once the HDT is constructed, sensitivity analysis and what-if analysis can be performed to determine how different values of an independent variable impact a dependent variable under a given set of assumptions. This analysis exposes the unknown influences of disease(s) or treatment processes. It delivers the synthetic knowledge necessary to rapidly identify a potential health risk with immediate analysis of root causes and proposed intervention.

By analyzing predictive outcomes produced by the HDT versus known cases under various conditions, the modeler can confirm the HDT’s representativeness, acceptable accuracy, and robust predictability. If done correctly, the HDT should closely match known scenarios. If not, the HDT parameters can be adjusted until the desired representativeness, accuracy, and predictability are achieved.

This provides an opportunity to reverse engineer the cause of the discrepancy and ultimately leads to the discovery of new insights or the identification of unknowns by finding missing definitions or introducing new parameters at various levels of the graphs. Such efforts are particularly important once the modeling process has been established since changes in health can occur at any time and necessitate the ability to quickly identify a new risk and apply therapeutic actions.

3.2.4 Step 4: validate findings

To confirm HDT scenario findings and further enhance subject matter expertise, any new predictions or causal relationship insights should be vetted through traditional randomized experiments designed to mimic digital findings in the physical world. Performing targeted experiments to understand failed predictions is a proven method for systematically discovering new epidemiological knowledge [67]. Since model predictions are based on a system reconstruction that represents the totality of what is known about human health, such predictions are a critical test of the current comprehensive understanding of epidemiology for a target patient. Incorrect predictions can be used to discover determinants by classifying them and understanding their underlying causes.

3.2.5 Step 5: certify models

After validating the findings, dynamic patterns encapsulating behaviors, dependencies, and surrounding rules for health behaviors should be stored in a database to support ML and predictive analysis activities. Rules can include remedial options for risk avoidance or preventive interventions. Every human has some biological characteristics that make him or her unique, but in large part, the essential components of health share many commonalities across patients. Building libraries that contain certified dynamic patterns of generic patient models that cover the full hierarchy of genes and biology and various disease states as well as external health influencers, such as behavioral, environmental, and socioeconomic factors, can help provide base models for many HDT applications as well as accelerate model customization tasks in clinical practice. Encouraging contributions to open libraries from related areas of science provides an opportunity to continuously enhance the domain knowledge, support better collaboration across disciplines, and provide greater accessibility to the latest research.

3.2.6 Step 6: use in clinical applications

All combinations of certified generic HDT models and related cognitive findings can be customized with patient-specific data collected from electronic health records and/or medical evaluations to be used in clinical care settings (see Figure 7). After completing this customization, care providers or artificial (AI) automated technologies can use the HDT to evaluate scenarios under different patterns of initial conditions and dynamic constraints to identify the circumstances under which health risks increase and use the corresponding information to evaluate the case.

Figure 7.

Libraries of generic models and AI rules provide a starting point for custom patient HDTs in clinical practices.

Working to establish universally accepted health metrics would provide clinicians with a framework to quantify morbidity risk and evaluate various intervention options. Table 1 provides an overview of proposed HDT health scoring metrics.

Human health dependability (HhD) score	Represents the impact of time sensitive interdependencies on human health metrics, which provides an early indication of whether the patient is approaching a health risk.
Energy reserve	Measures the health of patient’s immune system and its ability to fight disease.
Mean tolerance score	Measures how quickly the health of the patient may deteriorate over time.

Table 1.

Definition of patient health metrics.

3.3 Example HDT research use case

The following exploratory multimorbidity use case demonstrates how HDTs can provide a framework to unify NCD research across multiple disciplines. In the following scenario, an HDT is created for a hypothetical patient named Bob Smith. The ledger shows changes in health metrics as the patient’s circumstances evolve and new datasets are added, e.g., new test results (see Figure 8).

Figure 8.

HDT ledger for hypothetical patient.

Using the situational data revealed from the HDT, the ultimate objective is to prevent NCDs for individual patients. Scientists can use HDTs to predictively analyze how changing patient parameters may result in a health risk or support a more optimal health outcome (see Figure 9).

Figure 9.

Use of HDTs for preventive medicine research and personalization of clinical care disease prevention decisions.

In clinical settings, HDT capabilities could rapidly identify a potential problem with immediate analysis of root causes and proposed corrective actions. The active monitoring of HDT outcomes would provide a fully vetted platform to support individualized and proactive patient risk avoidance and suggest personalized preventive protocols when necessary (see Figure 10).

Figure 10.

HDT clinical decision tree.

4. Open challenges and opportunities

A new preventive medicine paradigm that provides a path to contain skyrocketing healthcare costs, extend lifespans, improve citizens’ well-being, and increase nations’ productivity is a worthy pursuit. Still, obstacles persist.

4.1 HDT ethical challenges

Achieving timely preventive diagnosis, prognosis, intervention, or treatment optimization involves various tasks to build and operate the HDT, including data collection, analysis of current or future health states, and risk scoring. The quality of data inputs and outputs is always a concern. Reliable results cannot be achieved with low-quality data or the wrong analysis methods. Each stage of HDT development and use presents different ethical challenges.

4.1.1 HDT development

Across all industries, establishing credibility and confidence in AI is key to success. Disease prevention and precision medicine require AI to compute all the parameters influencing human health. All HDT vendors should adhere to responsible AI standards to ensure their use for critical applications is safe, trustworthy, and unbiased. The use of AI and ML in treatment or diagnosis must avoid distributional shifts—or else the target data will not match ongoing patient data and will lead to inaccurate conclusions, misjudgment, or incorrect risk scoring. This can be accomplished with modeling methods that are Robust, Explainable, Ethical, and Efficient. Many efforts are underway by technology and research communities to establish responsible AI standards.

HDT developers must use data that reflects the characteristics of the person served by the application and pay extra attention to calibrating and validating algorithms to avoid biased or discriminatory results. Also, it is essential to consider how HDT developers’ values and conception of health or diseases may influence decisions regarding which information to present and how to score risks. Any proposed capabilities will need research validation and future collaborations with research centers. All results should be published in peer-reviewed journals and presented at conferences to help target audiences understand how AI backed with HDT predicts patient outcomes and informs clinical decisions.

4.1.2 HDT data inputs

HDTs are constructed and configured using nonidentifiable and identifiable data. For instance, generic HDTs can be built using de-identified data sets, but to make predictions on a person’s health trajectory requires specific identifiable data provided by the patient, the patient’s care provider, or imported from the patient’s electronic health records (EHR). Potential patient-identifiable data includes medical history, demographics, habits, family history, genetic profile, exposure and incidents, medical encounters, diagnosis, orders and prescriptions, and test results. To avoid violating a person’s privacy and autonomy rights, informed consent is always required whenever data is collected or used. Measures should be taken to prevent service providers from secretly collecting more data than necessary and exploiting the data for financial gains [68].

4.1.3 HDT data outputs

HDTs compute 3 categories of synthetic data outputs that can be used for personalized health monitoring, diagnosis, prognosis, prevention, and treatment.

• Descriptive information covers what has happened or is happening to a person’s health.

• Predictive information offers foresight into what will likely happen to a person’s health.

• Prescriptive information suggests which action or intervention should be taken to improve or restore a person’s health.

One of the biggest challenges in the transition to rules-based medicine is the impact on the role of physicians. Any change perceived as decreasing a doctor’s control over a patient’s fate may not be well received. If AI predicts a certain outcome, doctors and patients want to understand the logic behind the prediction. Therefore, any technology vendor must work with target audiences to define the right product features/benefits and plan a phased rollout of product capabilities that support desired use cases.

Contrary to the goal of empowerment, HDTs might burden patients with a sense of powerlessness, guilt, and anxiety, especially in the case of early lifestyle interventions such as diabetes, hypertension, and obesity [69]. The goal of earlier diagnosis and intervention could lead to overdiagnosis and bodily harm [70]. For example, many bioethicists and clinicians believe genetic testing for BRCA1 and BRCA2 mutations might cause overtreatment [71, 72].

Treating predicted NCD cases and patients with confirmed diseases similarly is morally problematic. Predictive diagnosis labels may increase concern about potential diseases and the desire for more invasive treatments. Evidence suggests that disease labels affect people’s psychological responses and healthcare decisions [73]. For example, Nickel et al. [74] suggest removing cancer labels for low-risk conditions may help reduce overdiagnosis and overtreatment.

4.2 Healthcare stakeholder opportunities

The innovative use of HDTs has the potential to align healthcare stakeholders towards a common goal of improving the value of healthcare for all citizens. Innovation and collaboration across healthcare stakeholders can help take HDT from concept to value.

• Scientists: HDT will speed precision medicine research by enabling new causal analysis capabilities that were previously impossible. The computational capacity to evaluate the effect of millions of dynamic multiscale variables exists, but readying HDT for clinical use requires that researchers work with mathematicians to develop an accurate representation of the parts that define health and the relationships between these parts. The validation of HDT-derived findings through in vivo studies and traditional data analysis will help translate research into effective precision medicine protocols and build confidence in the approach.

• Technology providers: Technology companies are building clinical decision support systems (CDSS) to help clinicians keep up with expanding medical knowledge and incorporate precision medicine concepts into practice. CDSS matches an individual patient’s characteristics to a computerized clinical knowledge base and then offers clinicians patient-specific assessments or recommendations for a decision [75]. The objective is admirable, but doctors hesitate to use AI-based CDSS recommendations without transparency. By replacing population averages with the exact mechanisms that define how genes, autoimmune responses, environment, and lifestyle choices influence the health of individual patients, HDT removes variability and randomness. Pairing the determinism of HDT with ML can help technology providers build confidence in HDT-based CDSS solutions.

• Care providers: One of the biggest obstacles to HDT rules-based medicine is the impact on the role of physicians. Trust and acceptance of new AI technologies must be established to improve healthcare efficiency. Through scoring metrics that are easy to interpret and explain, HDT can allow care providers to predictively identify at-risk patients and diagnose asymptomatic cases. With HDT, doctors can easily customize and approve recommended treatment plans based on the latest evidence-based medical guidelines and the patient’s unique medical characteristics—biomarkers, genes, and other clinical indicators. As new risks form, automated algorithms will help care providers proactively recommend the right actions to patients at the right time.

• Public health authorities: Most experts agree that a significant decrease in mortality, leading to improved healthcare economics and survival rates, can only be obtained through earlier diagnosis. The perception that disease control programs are too expensive or complicated significantly reduces the willingness of public health agencies to invest in such services. It is difficult to prove a return on investment for population-based preventive services or justify the anticipated strain on limited healthcare resources. HDT can help eliminate these barriers by providing a more granular way to identify at-risk patients, plan interventions that have the best chance of success, and enforce highly deterministic clinical pathways characterized by a high level of reproducibility of care. By investing in HDT research and explainable AI-enabled CDSS that support non-invasive asymptomatic disease detection years before conventional diagnosis, countries can measurably reduce national healthcare costs, improve citizen wellbeing, judiciously manage healthcare resources, and ensure equal access to quality care.

• Regulators: In many cases, medical device certification requirements, which apply to CDSS as software medical devices, hinder innovation. A collaborative solution between technology providers and regulators will be necessary to modernize regulatory frameworks and accelerate the time to market for AI-based CDSS. Certifying AI rules using outdated, static medical control frameworks is impossible. To translate state-of-the-art precision medicine research and HDT-derived intelligence into clinical value, AI rules must be as dynamic as the health problems they are meant to solve. In response, regulators must expedite the adoption of new risk-based controls or other flexible frameworks that meet market requirements for robust, explainable, ethical, and efficient AI-enabled CDSS.

5. Conclusion

Operationalizing HDT in clinical practice will take strong commitment and collaboration between scientists, doctors, technologists, policymakers, regulators, and payors, but the promised payoffs are worth the pursuit. Collectively, HDT, AI, and CDSS technologies can provide the impetus citizens, governments, payors, and care providers need to finally shift the focus of medicine from sickcare to healthcare.

Healthcare faces an NCD crisis that has not improved despite many technological and biological breakthroughs of the twenty-first century. While some people may push for collecting more health record data, more rigorous diagnosis training programs, or better coordination between healthcare specialists as the solution, none of these actions will improve the standard of patient care or judiciously manage NCD-associated costs if uncertainty persists.

Gaining confidence in any given analysis is the backbone of human and artificial decision-making. The uncertainty inherent in probability-based NCD risk stratification must be solved, or any attempts to meet the goals of personalized precision medicine will remain prone to error, especially in multimorbidity cases, because complex and ambiguous presentations cannot be covered by experience and research data alone. No matter how much data is collected or how well-trained a doctor may be, the dynamics of human health cannot always be inferred from knowledge, experience, or intuition.

On the surface, the call to upend hundreds of years of preventive practices built upon analogies and statistics may seem radical, but the disruption is necessary. Research has proven that complex interactions between genetics, biochemistry, physiology, microbiology, and biomechanics affect health and indirectly impact less obvious factors, like extrinsic factors such as public policy, environment, or social interactions with others. To improve NCD control, new methods are needed to relate these complex causes quickly and reliably into effects for an individual patient.

Now is the time to advance preventive care decisions from an art form to a scientific discipline that takes advantage of the deterministic modeling methods used in other branches of science. Pairing mechanistic HDTs with robotics and AI capabilities will allow clinicians to autonomously predict patient outcomes without a modeling expert’s assistance. To this end, scientists, researchers, and practitioners should collaborate to prove the robustness of the solution in prospective research cases and build the interfaces necessary to support the use of HDT in critical preventive medicine use cases.

Conflict of interest

While the author did not receive, at any time, any payment or services from a third party for any aspect of the submitted work, he wishes to declare the use of a proprietary algorithm and methodologies, which are subject to a pending patent.