Anshika Jain
For decades, the electronic health record has primarily served as a digital version of the medical chart. It stores diagnoses, laboratory results, prescriptions, imaging reports, procedures, clinical notes and other information generated during a patient’s interaction with the healthcare system. The transition from paper records to electronic health records transformed how medical information could be stored and accessed, but the basic purpose remained largely descriptive: the record documented what had happened to the patient.
The next stage of healthcare computing is considerably more ambitious. Instead of simply recording a patient’s medical history, computational systems are increasingly being designed to interpret that history, identify patterns across time and estimate what might happen next. This creates the possibility of transforming the electronic health record from a passive archive into the foundation of a dynamic computational patient.
A computational patient is not simply a digital copy of a medical record. It is a continuously interpreted representation of an individual’s health state, constructed from longitudinal clinical information and potentially enriched with laboratory measurements, medical images, genomic information, wearable-device signals, medications, environmental exposures and other sources of health data. Machine learning and other computational methods can use these data to model disease trajectories, estimate risks and support clinical decisions.
This transformation is closely connected to the emerging idea of human digital twins. Recent research describes human digital twins as virtual representations that integrate real-world data with computational models to simulate physiological behaviour, treatment responses and disease trajectories. However, a predictive patient model does not automatically constitute a full digital twin. The computational patient exists along a spectrum, ranging from predictive models built from EHR data to increasingly dynamic systems that continuously update themselves as new information becomes available.
The Electronic Health Record as the Foundation
The modern EHR contains a remarkable amount of information about an individual. A single longitudinal record can contain years of diagnoses, medications, laboratory measurements, hospitalisations, procedures and clinical observations. When viewed collectively, these records can reveal patterns that are difficult to identify during an individual clinical encounter.
Traditional healthcare often examines data encounter by encounter. A physician may review the patient’s current symptoms, recent laboratory values and relevant medical history before making a decision. Computational models can approach the same information differently. Instead of focusing primarily on the present encounter, they can examine relationships across months or years.
This longitudinal perspective is important because many diseases develop gradually. Changes in kidney function, blood pressure, glucose regulation, cardiovascular risk or medication response may become meaningful only when viewed as trajectories rather than isolated measurements.
The computational patient therefore begins with a fundamental shift in perspective: the patient is not represented as a collection of disconnected medical events but as an evolving system.
Turning Records Into Computable Information
One of the biggest obstacles to building predictive patient models is that EHR information is not naturally organised for machine learning.
Some information is structured. Diagnoses may be represented by medical codes, medications may have standard identifiers and laboratory tests may be stored as numerical measurements. Other information exists in free-text clinical notes written by physicians and other healthcare professionals.
Those notes can contain some of the most valuable information in the record. They may describe symptoms, treatment decisions, adverse effects, adherence problems, family history, clinical reasoning and observations that are not completely captured by structured fields.
A September 2026 Nature Medicine study demonstrated the importance of this problem by developing methods to extract computable clinical information from both structured and unstructured EHR data. The researchers noted that much real-world evidence research traditionally relies on structured fields even though substantial clinical information exists in narrative text. Large pretrained language models were used to transform this unstructured information into data that could be analysed computationally.
This represents an important development because the computational patient cannot be more informative than the data from which it is constructed. If important clinical information remains trapped inside narrative notes, predictive models that ignore that information may produce incomplete representations of the patient.
From Medical History to Patient Trajectory
The most important change in computational healthcare may be the movement from describing a patient’s past to modelling the patient’s trajectory.
A conventional EHR can tell a clinician that a patient previously had elevated blood pressure, received a particular medication and returned for follow-up. A predictive model can attempt to determine whether the patient’s pattern of measurements and treatments indicates increasing cardiovascular risk.
The distinction is subtle but powerful. Historical records answer questions about what has happened. Predictive models attempt to answer questions about what may happen next.
This does not mean that a prediction is a certainty. Medical outcomes are influenced by biology, behaviour, treatment, environment and chance. Instead, predictive modelling produces estimates of probability that can help clinicians identify patients who may need closer attention.
The computational patient therefore becomes a dynamic representation of risk rather than merely a chronological archive.
The Role of Machine Learning
Machine learning provides many of the computational techniques required to transform longitudinal health data into predictive representations.
Traditional statistical models can identify relationships between selected variables and clinical outcomes. Modern machine-learning systems can analyse larger and more complex datasets, potentially identifying nonlinear relationships and interactions that are difficult to specify manually.
Deep learning can process medical images, time-series measurements and other high-dimensional data. Natural-language processing can interpret clinical notes. Transformer-based models can analyse sequences of medical events and identify relationships across long periods of a patient’s history.
Different approaches can therefore contribute different parts of the computational patient.
An imaging model may represent structural information. A language model may extract information from clinical narratives. A time-series model may analyse physiological measurements. A risk model may estimate the probability of a future outcome.
The challenge is to combine these representations into a coherent patient-level model.
The Computational Patient Is Multimodal
Human health does not exist inside separate data categories. A patient’s physiology, behaviour, environment and medical history interact continuously.
This is why future computational patients are likely to become increasingly multimodal.
An individual’s model could potentially combine EHR information with medical imaging, laboratory results, genomic or molecular measurements, wearable-device signals and other sources of longitudinal information. Research on human digital twins increasingly identifies this multimodal integration as a central component of personalised and predictive healthcare.
Consider a patient with cardiovascular disease. Their computational representation could incorporate years of blood-pressure measurements, medication history, cardiac imaging, laboratory results and physiological signals from wearable devices. Instead of examining each dataset independently, an integrated model could attempt to understand how the different signals relate to one another.
The result would not simply be a larger medical record. It would be a computational representation capable of generating predictions from relationships across different types of evidence.
Predicting Disease Before It Becomes Obvious
One of the most important potential applications of computational patients is earlier identification of disease risk.
Traditional healthcare frequently becomes more intensive after symptoms appear or a diagnosis has already been established. Predictive healthcare aims to identify meaningful changes earlier.
For example, a computational model could recognise that a combination of laboratory trends, medication changes and clinical observations is associated with increasing risk of a particular complication. A clinician could then investigate the patient before the condition becomes more severe.
This approach is particularly relevant to chronic diseases because they often evolve over long periods.
Research into digital twins and predictive healthcare has explored applications across cardiovascular disease, oncology, chronic disease management and other areas. A 2026 scoping review of cardiovascular digital twins identified potential applications including personalised management, precision treatment, risk prediction and clinical-trial optimisation, while also highlighting challenges involving data limitations, model validation and clinical adoption.
The promise is therefore not simply earlier diagnosis. It is the possibility of creating healthcare systems that recognise changing risk before conventional clinical pathways would necessarily trigger action.
Predictive Models Versus Digital Twins
The phrase “digital twin” is increasingly used in healthcare, but it is important not to treat every AI prediction system as a digital twin.
A predictive model may take a set of patient variables and calculate the probability of a future event. A digital twin aims to go further by maintaining a dynamic representation of a specific patient or physiological system and updating that representation as new information arrives.
The distinction can be understood as a difference between prediction and simulation.
A predictive model might estimate a patient’s risk of hospitalisation. A more advanced computational patient could potentially simulate how that patient’s risk might change under different treatment strategies.
Research published in 2026 describes human digital twins as systems that can combine multimodal data with physiological modelling and machine learning to represent individual anatomy, physiology and clinical characteristics. Such systems are being investigated for applications including treatment response prediction and disease-trajectory modelling.
However, fully functional clinical digital twins remain an emerging research area. Many current systems are better understood as specialised computational models rather than complete virtual replicas of individual patients.
Simulating Treatment Decisions
The most ambitious version of the computational patient could eventually allow clinicians to explore potential interventions before applying them to the real patient.
Imagine a model that incorporates a patient’s disease state, previous treatment responses, physiology and relevant biological information. A clinician could potentially use such a model to compare different treatment scenarios and estimate how the patient’s condition might respond.
This is particularly attractive in fields where treatment decisions are complex and patient responses vary substantially.
Oncology is one example. Cancer is not a single disease, and tumour biology can differ considerably between individuals. Recent reviews of digital twins in oncology have explored applications ranging from treatment selection and radiotherapy to drug development and surgery, while emphasising that data integration, model validation and scalability remain significant barriers.
The computational patient could therefore become a platform for asking not only “What is happening?” but also “What could happen if we choose this intervention?”
Learning From Continuous Health Data
The EHR traditionally captures healthcare interactions at specific points in time. Wearable devices and connected sensors can potentially provide a much more continuous stream of information.
Heart rate, activity, sleep patterns, glucose measurements and other physiological signals can add temporal detail to the computational patient. Instead of receiving a small number of measurements during clinical appointments, a model may eventually incorporate information describing how the patient’s physiology behaves between appointments.
This could make the patient model more dynamic.
However, more data does not automatically mean better prediction. Continuous measurements can contain noise, missing values, device errors and behavioural changes. The computational system must therefore distinguish meaningful signals from irrelevant fluctuations.
The challenge is not simply collecting data. It is determining which data genuinely improves the representation of the patient’s health.
The Problem of Data Quality
A computational patient can create an illusion of precision if the underlying information is incomplete or inaccurate.
EHRs may contain duplicated information, outdated medication lists, inconsistent terminology or missing clinical observations. Different hospitals may store similar information using different systems. Data collected by wearable devices can vary according to device quality and user adherence.
These limitations can directly influence predictive performance.
A 2026 methodological review of healthcare digital twins identified interoperability and heterogeneous data integration as foundational barriers to widespread adoption. The review emphasised the need for standardised data models, privacy-preserving learning, clinical validation and workflow-aware design.
This means that building the computational patient is partly an infrastructure problem. Advanced AI cannot compensate indefinitely for fragmented or unreliable data.
Interoperability Becomes Essential
A patient rarely receives all healthcare from one institution. Medical information may be distributed across primary-care clinics, hospitals, diagnostic laboratories, pharmacies and specialist practices.
If these systems cannot communicate effectively, the computational representation of the patient becomes fragmented.
Interoperability is therefore a critical requirement for predictive healthcare. The model needs access to relevant information regardless of where that information was generated.
This is particularly important for longitudinal modelling. A missing hospitalisation or medication change can alter the interpretation of subsequent events. A model that sees only part of the patient’s history may generate a prediction that appears mathematically sophisticated but is clinically incomplete.
The future computational patient consequently depends as much on connected health-data infrastructure as on AI algorithms.
Privacy and the Computational Identity of a Patient
The more comprehensive a computational patient becomes, the more sensitive it becomes.
A conventional medical record is already highly private. A computational representation could contain not only historical diagnoses but also predictions about future disease risk, treatment response, behaviour and physiological changes.
This raises important questions about who controls the computational representation and how it can be used.
Patients may reasonably want to know which data contributes to their model, who can access it and whether predictive information can be shared beyond direct healthcare.
Privacy-preserving technologies such as federated learning may become increasingly important because they can support collaborative model development while reducing the need to centralise sensitive patient data. Recent digital-twin research has identified federated approaches as one possible direction for addressing data-sharing constraints.
Bias Can Become Part of the Patient Model
Predictive models learn from historical data. If the data reflects unequal access to healthcare, differences in documentation, demographic bias or other structural inequalities, the resulting model may reproduce those patterns.
This creates a particularly important problem when computational systems are used to determine who receives additional monitoring or intervention.
A prediction that appears objective may actually reflect patterns in the healthcare system rather than pure biological risk.
For this reason, computational patient development requires evaluation across diverse populations and healthcare settings. Models must be tested not only for overall performance but also for whether their predictions remain reliable across different patient groups.
Personalisation should not mean creating highly sophisticated models that work exceptionally well for some populations while performing poorly for others.
The Human Clinician Remains Essential
The computational patient should not be understood as an autonomous replacement for clinical judgment.
Healthcare decisions involve values, preferences, uncertainty and communication. A model can estimate risk, but a patient may choose differently depending on their priorities, circumstances and tolerance for risk.
Clinicians also understand contextual information that may not be fully represented in datasets. A physician may recognise that a patient’s recent measurement is unusual because of a temporary circumstance rather than a meaningful deterioration.
The strongest model therefore remains part of a human decision-making system.
The purpose of computational healthcare should be to provide clinicians with richer evidence, not to remove the human relationship from medicine.
From Retrospective Records to Living Models
The long-term transformation of the EHR may ultimately involve a change from retrospective documentation to continuously evolving computational representation.
The traditional EHR asks what happened during healthcare encounters. A predictive computational system asks what the accumulated information suggests about the patient’s current state and future trajectory.
An advanced digital twin asks an even more ambitious question: how might the patient’s state change under different circumstances?
These are different levels of computational medicine, and healthcare is currently moving gradually between them.
The transition will not happen simply because more powerful AI models become available. It will require interoperable data infrastructure, high-quality longitudinal information, reliable validation, privacy protection, clinical integration and clear governance.
The Future of the Computational Patient
The computational patient could become one of the most important conceptual developments in data-driven medicine. Instead of treating every clinical encounter as a separate event, healthcare could increasingly view each encounter as another update to a continuously evolving patient model.
New laboratory results could modify risk estimates. A medication change could alter predicted treatment response. A new scan could update disease-state modelling. Wearable data could reveal changes occurring between clinical visits.
Over time, the model could become increasingly personalised.
Yet the ultimate goal should not be to create the most complicated virtual representation possible. The goal should be to create computational representations that are clinically useful, scientifically defensible and understandable enough to support real decisions.
A sophisticated model that cannot be validated or integrated into clinical practice has limited value. A simpler model that reliably identifies meaningful risk at the right moment may have a much greater impact.
Conclusion
The electronic health record began as a way to digitise medical documentation. Its next evolution may be to become the foundation for computational representations of individual patients.
By combining longitudinal EHR information with artificial intelligence, clinical notes, medical imaging, laboratory measurements, wearable signals and other sources of health data, researchers are beginning to explore systems capable of modelling disease trajectories rather than simply recording past events. Recent research into computable patient journeys and human digital twins demonstrates how quickly this field is developing.
The computational patient does not mean that every person will soon have a perfect digital replica. Significant challenges remain in data quality, interoperability, validation, privacy, bias, regulation and clinical adoption. Current digital-twin research continues to identify these limitations as major barriers to translation into routine healthcare.
Nevertheless, the underlying direction is clear. Healthcare is gradually moving from static records toward dynamic models that can interpret history, recognise patterns and anticipate possible futures.
The ultimate transformation may be from an EHR that tells clinicians what happened to a computational patient model that helps them understand what is happening, what may happen next and how different decisions could change the trajectory.
That shift could redefine the role of medical data itself. Instead of remaining a record of healthcare, data could become an active component of how healthcare is predicted, personalised and delivered.