Multimorbidity, the co-occurrence of 2 or more long-term conditions, has been associated with numerous adverse outcomes and overwhelming financial burdens on health care systems [1]. Its prevalence in the United Kingdom is projected to almost double by 2035, and at least two-thirds of the gain in life expectancy for people aged >65 years will be spent with 4 or more chronic conditions [2]. Multimorbidity is most common in older people, but it is seen in patients of all ages [3]. It is more prevalent, and at younger ages, among people who live in socioeconomically deprived areas [4]. Health, behavioral, and environmental factors significantly influence individuals’ risk and many processes that lead to multimorbidity begin with chronic illness at much earlier ages. Multimorbidity can occur in clusters with predictable disease pathways [5]; however, the specialization of health care has created a challenge for identifying and treating co-occurring health conditions. An interdisciplinary approach combining clinical knowledge of disease pathways, clusters, and risk factors with artificial intelligence (AI) technology could improve understanding of the biosocial factors that are associated with developing multimorbidity. Understanding key risk factors for multimorbidity could enable doctors to monitor and treat patients more effectively, potentially mitigating or preventing multimorbidity.

A holistic, patient-centered system will be essential to address the challenge of preventing and managing multimorbidity and will require a combination of prediction, monitoring, and intervention. Digital twins—an engineering concept that uses real-world data and AI to provide a virtual representation of a physical counterpart—are starting to be explored in health care. The concept of the digital twin can be interpreted in relation to patients as a means of improving diagnostics and treatments by processing vast amounts of data to develop predictive health trajectories for individuals [6]. Our aim is to adapt this engineering solution to draw on continuously updated individual data on factors and health outcomes to simulate an individual’s future health status [7].

There is little established evidence on effective means of identifying risk factors or preventing and managing multimorbidity throughout the life course [8,9]. The vast majority of multimorbidity studies aim to identify disease clusters at a single point in time, providing little information about how multimorbidity develops over time within individuals [10]. This study will develop a system where innovative AI approaches are used to analyze complex longitudinal data and predict risk levels of multimorbidity. Using simulation, digital twins could identify key risk factors, consider weaknesses in source data through quantified underreporting, model potential adverse health outcomes, and recommend the most effective intervention when given adequate information. The use of digital twins is aimed at making significant enhancements within health care and advancing knowledge of disease [6]. When used in conjunction with clinician experience and knowledge, it will aid decision-making and highlight risk factors early on, so that steps can be taken to mitigate them and avoid later health consequences. The proposed digital twin solution has the potential to have a significant positive impact on individuals and health care systems.

The aim of this study is to identify key risk factors that predict multimorbidity throughout the life course by developing an intelligent agent using digital twins. The hypothesis is that this intelligent agent can use a wide range of biosocial variables (including physical and mental health and behavioral, socioeconomic, relational, and environmental conditions throughout an individual’s life) to predict risk of multimorbidity.

The three main objectives of the Life Course Digital Twins–Intelligent Monitoring for Early and Continuous Intervention and Prevention (LifeTIME) study are the following:

Using an implementation science theoretical framework and a retrospective cohort design, this will be a feasibility study aiming to address the following research question: Can historical data and data captured via a dynamic remote monitoring system be used to develop a digital twin that can predict individual risk of developing multimorbidity over a lifetime? An implementation science framework provides guidance regarding design and conduct to translate innovative ideas into practice [11]. It also gives ideas and suggestions to achieve various intended research pursuits.

For this study, 4 longitudinal databases will be used to determine how the presence or absence of particular factors (health, behavioral, and environmental) during childhood is associated with the development of health conditions and multimorbidity over the life course. The longitudinal data are needed to inform the model about childhood factors and health outcomes over the life course to identify potential early predictors of later multimorbidity. The study will consist of two phases: (1) prototype development, where the model will be trained to identify associations using subsets of the data sets; and (2) validation, where the model will be tested on the remaining subsets. Textbox 1 and Figure 1 provide an overview of the study design and framework.

The health care data that will be used to train and validate the model will be collected from several pre-established databases. To satisfy the data requirements of model training and validation, a few longitudinal data sets will be used and a variety of factors—including mental and physical health, and behavioral, environmental, and socioeconomic indicators—will be examined. The cohort used to train and validate the model will be built using data from previously established databases and split into random subsets. It will build upon previous work conducted on data linkage, the harmonization of multiple sources of patient-related e-records [12], and use of AI for health data analytics [13,14]. Using multiple databases will allow for a broader range of variables to be included in model development. In total, 4 longitudinal databases (the National Child Development Study [NCDS] [15], the Clinical Practice Research Datalink [CPRD] [16,17], the North West London Integrated Care Record [NWL ICR] [18], and Cerner’s Real-World Data from UK and Irish population [19]) were selected because they cover a range of mental and physical health outcomes, health behaviors, and other characteristics over decades in a diverse UK population (see Table 1).

Vast amounts of data will be used to develop the data sets, which will then generate a computational model to establish the relationship between biosocial factors and later health outcomes. Machine learning will be used to train a model using a variety of factors. The model will need to be trained on longitudinal data that includes both biosocial characteristics and later health conditions to establish links between factors and outcomes and account for historical differences in risk factor and multimorbidity prevalence.

Once the predictive model has been developed based on the 4 longitudinal databases detailed in Table 1, two additional data sets will be incorporated into the development of the digital twins. Data from the Early-Life Data Cross-linkage in Research (eLIXIR) study [20] and the Children and Young People’s Health Partnership (CYPHP) randomized controlled trial [21] will be included in the development of the digital twins because they track early life variables and use of health care services. Using the preliminary model developed from the longitudinal data sets in combination with machine learning to identify clusters of factors that best predict the later risk of developing multimorbidity, a series of digital twin personas will be created to provide simulations of different types of people (see Figure 2).

Although the 2 additional data sets do not contain longitudinal health outcomes, incorporating them at this stage will enable the digital twin personas to be developed based on the clustering of a more comprehensive set of variables. For instance, if a particular variable was identified in the model as a predictor of multimorbidity, these additional data sets can provide more detail about variables that commonly co-occur with that predictor. This is beneficial because these data sets will provide more current data about early life factors than the longitudinal databases. The number of digital twin personas developed will depend on the clustering of the predictive factors. For this study, these personas will be developed to represent children, but in the long term, they will be developed to simulate different stages throughout life.

The validity of the developed model and the digital twin prototypes will be assessed using the k-fold cross-validation procedure [22], with k being determined when the total cohort size is known. The machine learning algorithms will be tested on the validation data set to determine which model has the best fit. Model validity will be examined using a variety of indicators, including receiver operating characteristic and precision-recall curves (see Table 2). The Prediction Model Risk of Bias Assessment Tool (PROBAST) will be used to assess the model’s risk of bias [23].

Ethical approval for this study will be sought from the University of Plymouth’s Faculty of Health Research Ethics and Integrity Committee. All the data being used in this proof-of-concept study will be procured from pre-existing databases. Therefore, all data access will be dependent on the approval of the study protocol by the data controllers for those databases; for instance, the CPRD database requires that study protocols be approved by an Independent Scientific Advisory Committee before data are shared. All storage and use of the data provided will comply with database-specific and General Data Protection Regulation (GDPR) requirements. It will be confirmed before access that the data have been sufficiently anonymized to comply with GDPR requirements and that the databases’ patient consent processes extend to cover the sharing of anonymized data for other research purposes (ie, that their personal health data is being used in a way that could be reasonably expected by the participants). Processes for the removal of patient data—at their request to the specific database—will be established before data use with each of the database managers. Potential ethical issues relating to the development and implementation of digital twins in health care will be identified from the literature and discussions with a patient and public steering group.

By associating a patient with a digital twin persona and then personalizing it with their evolving data, health care providers and researchers could receive a personalized risk score representing the probability that the patient will develop further health complications. With the capability to start at a young age, a digital twin system has the potential to provide significant clinical value by identifying risk factors early, so that preventive interventions can be made to reduce an individual’s risk of developing long-term health conditions. This would save time and resources for health care providers, hospitals, and health care systems and reduce health problems and related financial costs for patients.

If the study results are positive, future research could develop and incorporate a remote monitoring system that can collect and collate individual data with the population data in a pilot trial. This would enable the digital twins to be further developed and iteratively improved based on feedback from the participants. As the system will use sensitive personal data and have ethical implications, safety and efficacy data will be an essential next step on the path to achieving wide-scale patient benefit.

Based on a user’s personal data, health care providers and researchers would be able to associate an individual patient with the digital twin persona that is the best fit for them. This will then provide a risk score representing the probability that the patient will develop further complications or health conditions. In the future, the digital twin will also be trained with data on interventions so that it can provide recommendations to health care providers about what interventions are most likely to be effective for that individual. In the long term, data collected from this remote monitoring system will be combined with childhood digital twin personas to increase the variables for prediction and to personalize the best-fitting persona based on individual health data.

Potential limitations of the study have been identified, with plans for their mitigation, to maximize the potential usefulness of the study.

Incorporating a remote monitoring system would enable tracking of ongoing lifestyle data, which could be combined with the individual’s personal health data and population health and lifestyle data to potentially increase the potential impact. Personalized patient simulations could be used to optimize prevention or treatment selection, saving patients unnecessary side effects, reducing the number of treatments attempted, and reducing the need for repeat hospital visits. Prevention and early intervention to manage multimorbidity will support healthy aging and reduce its negative impacts on quality of life and the workforce.