In medicine, it is generally difficult to define patient outcomes quantitatively (related to measurable information: exact metrics and observables captured by sensors) and qualitatively (related to nonmeasurable descriptions: categories) because patients are heterogeneous and complex organisms, and an overall assessment of their health status therefore remains a challenge.

Our goal is to identify these variables of interest and place them in a broader understanding of quality assessment, paving the way for mathematical modeling in this field. To this end, we aim to introduce a description of patient outcomes as a fundamental concept for evaluating quality indicators (QIs). These indicators are relevant for decision-making because they provide an estimate for which option holds the greatest desirability or value [12], that is, improvement of the quality of care (eg, improvement of patient well-being, avoidance of patient death, optimal use of resources, etc).

From a medical perspective, a patient outcome is defined in terms of what is meaningful and valuable to the individual patient [13]. However, this definition is very general and difficult to relate to a quantifiable metric. There is still a conceptual problem with the general definition of results in the medical context, such as patient-relevant outcomes (morbidity, mortality, and quality of life) and surrogate or biomarker outcomes [14]. These outcomes can differ significantly depending on the underlying disease. Liu et al [15] summarized the definition of patient outcomes related to clinical practice “as any change [within the patient’s health status] that results from health care [for which each profession] has developed outcome measures that focus on the standards, activities, and impact of its discipline.”

Thus, patient outcomes are manifold and range from PCQC [14] to institutional performance [16] and to patient-reported outcomes [17], which can be assessed from different data sources (for a comprehensive overview, see information provided by Busse et al [18]).

In this perspective article, we focus only on the description of PCQC (Table 1), for which a comprehensive review was published by Kersting et al [14]. We aim to provide a reference for the general concept, which is why we generally refer to medical procedures as follows: A medical procedure is the overarching phrasing for any type of observation, measurement, diagnosis, or treatment performed on the patient.

However, not all patient outcomes can be measured objectively, and only a limited number can be translated into clear, quantifiable metrics. In other words, a mathematical model can reflect only a fraction of the clinical reality.

In the context of clinical trials, great efforts have been made to establish quantifiable patient outcomes. Clinical trials use predefined endpoints to measure patient outcomes. Typical endpoints are patient survival at a given time, incidence of clinical events (such as stroke), clinical performance measures, or patient-reported outcomes [20].

Considering the impact that EHRs have on the quality of care, we provide a definition of patient outcomes on the basis of changes registered in the health record and defined as events [21] (assuming that the data relevant for the definition of patient outcomes are machine-readable). We refer to patient outcomes as events registered in the health records. Such events can include changes in a patient’s health state or changes in medical procedures.

These outcomes enable health care providers and organizations to measure the impact of their services on patient well-being, identify key areas for improvement, and enhance the overall quality of care.

Medical guidelines are systematically developed statements that reflect the current state of medical knowledge. Medical experts prepare these guidelines in cooperation with professional associations. These evidence-based recommendations are based on clinical studies published in scientific literature and/or other existing evidence informed by expert experience and consensus.

The extracted information is then assessed and published by organizations, such as the World Health Organization, for international guidance or by country-specific professional societies (in Germany, the Arbeitsgemeinschaft der Wissenschaftlichen Medizinischen Fachgesellschaften e. V. [22]). Guidelines are not a static corpus and are continuously adapted and further developed on the basis of current clinical studies and scientific findings. They intend to support the decision-making of HCPs in order to enable appropriate care for certain health problems. Medical guidelines consist of complex workflows in mostly narrative form and can also contain multiple, equivalent treatment alternatives for the same disease unit. In general, comparing performed procedures with guidelines is challenging, especially considering that guideline recommendations should not be understood as rigid limits. In medical guidelines, a distinction is made between several grades of recommendations (“can” or “should”) on the basis of available evidence (see, for example, the AWMF register for prostate carcinoma [in German] [23]). Additionally, medical guidelines are country-specific and may differ across national states and regions.

To assess in part how guidelines are applied within an institution and to improve the quality of care, several QIs have been developed by different institutions. These QIs serve as internal quality management tools for medical institutions, enable benchmarking with other institutions, and aim to improve the quality of care for patients [24]. In Germany and Switzerland, routine billing data from hospitals are used for this purpose, if a health care facility decides to participate voluntarily in such initiatives [25]. In Germany, QIs for cancer treatment are defined annually by the German Cancer Society (Deutsche Krebsgesellschaft [DKG]). Hospitals that meet the criteria defined by the DKG are eligible to receive DKG certification. The outcomes of cancer patients treated in these DKG-certified centers are superior in terms of patient survival [26]. However, these QIs are assessed retrospectively on an institutional basis and cover aspects of the application of medical guidelines.

Mathematical modeling is used to represent a real-world phenomenon by capturing and analyzing the most significant relationships between different observables through formalized relationships between mathematical parameters. For the quantitative description of patient outcomes, they need to be described by numerically or categorically measurable parameters. These parameters capture various relevant aspects of the overall performance of patient-centered care quality. These numerical and categorical parameters are referred to as quality metrics (QMs) because they enable an evaluation of patient care. There is a hierarchical ordering that assigns each QI to several QMs, which might be defined in a disease- or context-specific way, to place them into the perspective of mathematical modeling. We refer to QM as a quantitative measure of QIs that relates to adherence to medical guidelines (procedural accuracy); QM also refers to patient-centered care (patient safety and effectiveness of the procedure).

We present an overview of the broad concept of PCQC. According to the classification provided in Table 1, we define in detail the 3 basic categories used for the further discussion of health care quality assessment: patient safety, procedure accuracy, and procedure efficacy.

For PCQC, there is a strong reference to medical guidelines and appropriate clinical practice. Owing to the increasing complexity of medical knowledge reflected within medical guidelines, it is becoming increasingly difficult for health care professionals to keep track of changes. The application of mathematical modeling is intended to help health care professionals comply with the guidelines (ie, a sequence of recommended procedures and support for medical decision-making). Medical practitioners select their decisions on the basis of the probability of initiating a change in the patient’s health state (assuming that physicians are either using heuristics or making rational decisions on the basis of available information [27]). Use cases for the application of mathematical modeling for PCQC are presented in Multimedia Appendix 1.

Mathematical modeling can be included for predicting QMs to enhance health care quality. In this context, the patient’s journey can be understood as a complex chain of procedures that can be logically interconnected. Therefore, it is imperative to look at the whole patient journey as a process in which each procedure is just a point in time (t_i−1, t_i+1). As illustrated in Figure 1, the general way in which health care quality is assessed can be described as follows: collecting data from a patient (element 1), which a clinician uses to make a medical decision regarding the patient’s further procedure (element 2) and whose impact is measured after the procedure is carried out by looking at the patient outcome (element 3), which allows assessment of the quality of the health care provided (element 4). One can distinguish between two methods of implementation: (1) use the prediction in a prospective way (dashed arrow; meaning that before the medical decision is made, the clinician receives feedback from the model and can adjust the decision, leading to potentially different outcomes and improved health care quality) and (2) use a retrospective inclusion (dotted arrow), allowing an evaluation of the observed patient outcome and an assessment of the quality of care.

Another distinguishing feature of mathematical modeling is whether the model works for individual patients or for patient cohorts. Hence, for modeling the QM, we can consider the following representation levels:

The patient journey can be represented as a chain of observed patient outcomes (for example, the main disease and corresponding codiseases) and possible medical decisions that lead again to distinct patient outcomes over time that are patient journey dependent. Medical guidelines specify possible medical decisions following patient outcomes. The quality of the health care received during the patient journey can be assessed by comparing different elements within this decision-tree structure. Three categories of PCQC can be distinguished: patient safety, procedure accuracy, and procedure efficacy. In the following sections, the 3 different categories of modeled patient outcomes are defined by describing their context to the patient journey in a qualitative and quantitative way, as well as identifying suitable QMs.

Providing an exact definition of patient safety is challenging, as this term remains under debate in the literature [29]. Patient safety (related to the taxonomy introduced in Table 1) can be measured by patient safety indicators, which refer mostly to the avoidance of accidental and unwanted events in health care. Patient safety is partially measured within health care facilities via predefined inpatient QIs. Patient safety indicators and inpatient QIs were developed by the Agency for Healthcare Research and Quality (AHRQ [30]), and together, they provide a comprehensive picture of QIs for patient safety. These QIs are mostly determined from the billing data of the hospitals and the medical documentation data contained therein (ICD codes for diagnoses and International Classification of Health Interventions [ICHI] codes [31,32] for medical procedures, although each country can have its own classification schema). Possible examples of QIs that reflect patient safety can be extracted from the AHRQ [30], such as complication-related indicators, procedure-related errors, medical-related errors, and hospital system indicators.

There are a variety of possible patient outcomes, some of which are desirable and some of which are adverse (an adverse outcome is evaluated depending on the specific clinical context). Assuming that physicians are rational decision makers, the medical decision is an evaluation of the risk of previous and future outcomes (according to the classic Theory of Rational Decision Making; see, for example, information provided previously [33]). A mathematical model for patient safety can provide support through checking (retrospective) or predicting (prospective) how likely the outcome of a medical decision will be an element of the set of adverse outcomes.

QMs can be related to the respective QIs for the quantitative representation of patient safety. Examples can be defined as follows: (1) Mortality rate and failure to rescue, which have emerged as important QMs [34]; (2) Emergency readmission to the hospital after a procedure among outpatients, which can happen, for instance, if prophylaxis has been forgotten or if safety guidelines have not been properly followed; (3) Number of reoperations or complications shortly after the operation, which might indicate a need for investigation within the facility; and (4) Admission to intensive care units (ICUs) and length of stay (LOS) in the ICU among inpatients (different from the intermediate medical care unit).

Table 2 depicts the relationships of the example QMs to the respective QIs, the patient status, the data type, and the QM type. The “Combination event with a time period” column indicates that the observed metric is provided in conjunction with a time parameter.

The purpose of medical guidelines is to reduce errors and disparities in medical care while supporting best practices and responsibilities in medicine on the basis of the current state of scientific knowledge. Following established medical guidelines increases the probability of achieving better clinical outcomes [35-37]. Therefore, an appropriate measure of PCQC (related to the taxonomy introduced in Table 1) is the adherence of HCPs to medical guidelines, including the correct and accurate application of a medical procedure, which we refer to as procedure accuracy. To date, there is no structured or automated assessment of whether recommended medical guidelines have been followed by HCPs for an individual patient.

For the inclusion of mathematical modeling for health care quality assessment, it is necessary to identify metrics that can be defined precisely (if possible, with binary or clear multilevel indicators). In most cases, these metrics cannot be defined universally and require disease-specific development. Thus, with respect to procedure accuracy, mathematical modeling can provide useful insights as follows:

Clear results include checking whether the recommendations of a guideline for performing a standardized procedure or a set of procedures have been followed in an individual patient. This information can be retrieved from EHRs, laboratory information management systems, or even reimbursement data. In addition to compliance with a guideline, medical errors can also be reduced. Examples of QMs for procedure accuracy are as follows (Table 3):

While these metrics are easy to extract at the site level, they risk missing relevant contextual factors. It is desirable to integrate additional qualitative or clinical data points for a more holistic assessment. However, in clinical practice, simpler modeling techniques are used.

A general definition of efficacy has been provided by Lynch [42] (“efficacy is the capacity [of a medical intervention] to produce an effect”). For quality assessment, procedure efficacy (related to the taxonomy introduced in Table 1) refers to the optimal selection of procedures aiming at achieving the best possible patient outcome.

Retrospective models could compare the medical decisions chosen by HCPs with other possible choices, which may lead to better outcomes. Hence, procedure efficacy can also be considered for monitoring a decision (made under imperfect information) by observing the patient’s response or for making a prognosis and a subsequent decision under imperfect information. Owing to its nature, procedure efficacy requires the implementation of mechanisms and causal relationships where target/trial emulation is needed. This type of modeling contrasts with conventional predictive risk modeling tasks (which are applied, for instance, to evaluate patient safety and procedure accuracy).

The following are relevant QMs for procedure efficacy (Table 4):

In legal or patient-communication contexts, terms such as “survivability” are typically used, which refer to an inherent capacity or probability of survival but are not considered valid clinical endpoints. In this study, we refer to well-defined outcome measures such as tumor-free survival.

A mathematical representation of the patient’s condition to predict the outcome is essential for a precise evaluation of quality management. In summary, patient safety (Table 2) and procedure efficacy (Table 4) are related to clinical endpoints, whereas procedure accuracy (Table 3) is closely related to precision in the implementation of clinical guidelines and in reference to medical evidence. To avoid conceptual conflation in QMs, procedure accuracy is always defined with respect to documents (scientific literature or guidelines) containing medical or scientific evidence. Importantly, some events, such as patient outliers (Table 3), are estimated inside a time horizon and are mapped into binary states. Despite their convenience for mathematical modeling, there are relevant trade-offs, such as information loss or threshold choices, which can be problematic for the final validity of the models.

In the context of the 3 quality management categories, mathematical models need to perform distinct modeling tasks. First, the information coming from the patient needs to be stored and ordered in a data storage system (graph-based structures, such as knowledge graphs [KGs], can be used to encode patient data and information in a systematic manner). On the basis of the available data, single mathematical models (we focus on ML models) are trained to accomplish tasks that can either be classification tasks of the patient state or prediction tasks of the evolution of a parameter describing the patient response. The model predictions can be either prospective or retrospective.

We focus on KGs to integrate information from single events within the patient journey. By analyzing patterns from these KGs extracted via methods, such as graph neural networks (GNNs), the PCQC can be evaluated. For some QMs, it is necessary to consider the medical process in its specific context within the patient’s journey, which can be understood as a chain of events. This is an important aspect when assessing procedure efficacy, as it strongly depends on the previous and subsequent procedures. We present the concept of HDTs as a model integration framework, enabling a patient path representation.

The growing interest and relevance of mathematical modeling of patient outcomes are indicated in Table 5, which includes the number of published articles from the last 7 years (see Multimedia Appendix 2 for the methodology used to extract these data). However, the number of publications including digital twins (DTs) or GNNs is only a fraction of the total number of publications addressing the application of ML to predict patient outcomes (for a comprehensive review, see the article by Kline et al [43]).

With EHRs, extended research integrating different patient data for health care quality assessment is now possible. Regarding this development, Si et al [44] published a comprehensive survey about the use of current advances in patient representation based on data stored in EHRs.

In recent years, the application of linked knowledge in graphs has become relevant, in part because it allows holistic individualized representation, either by considering individual characteristics in an entire population [38] or by providing a patient-centered representation of the disease and the corresponding outcomes. Data representation as a graph is a method to leverage transparency in model construction. Graph models are also a way to map unstructured data, such as clinical narratives, into machine-readable data via large language models (LLMs). Additionally, graph models, including GNNs, have been considered to represent a way to close the gap between ML and symbolic reasoning [45], introducing the desired interpretability for ML methods.

The sample health care KG in Figure 2 can serve to represent the concept of graph models in health care: the health status of a patient can be modeled as the interlinking of various factors, such as the disease’s attributes (including type, symptoms, and anatomy), the patient’s genetics (including pathways and biological processes), the kind of drug compounds, and the characterization of side effects. Hence, graph models can function as holistic representations of a patient’s health status through the integration of different data sources into a single model.

In graph models, representation learning can be applied to convert multivariate time series and static features into nodes of the graph [47], for example, when multiple steps, such as various procedures in different time steps or disease stages, are converted into nodes in the entire graph.

Hence, starting from the disease stage at time point t_i, procedures, such as diagnosis, can be performed, which introduces a new node into the KG. As clinical work is ideally performed within the frame of the guidelines, suitable medication will be proposed and applied. Any action taken through the performance of a medical procedure can lead to an evolution of the patient’s disease state, which, at a time point, can be monitored by the next procedure, again inducing new nodes in the KG. Relating medical procedures to specific patient characteristics (such as the genetic profile of the patient, side effects, etc) enables a holistic view, allowing guidelines to be tailored to the individual patient and to differ from institutional guidelines. Graphs are not limited to representing a single event and can represent a whole process (ie, the patient’s journey). This is useful for DT implementation, which will be discussed in the following section.

In addition to patient-centered representations, graph models for disease-centered representations can also be constructed, as shown in Figure 3. These can be created from electronic patient records because hospital records contain rich information about the interrelationship of diagnoses (diagnoses and corresponding codiagnoses), the interrelationship of medical procedures (one medical intervention leads to another), and the interrelationship of diagnoses and procedures (due to the causal relationship between the diagnoses and the associated procedures). If a malignant neoplasm of the lungs (C34) is diagnosed, it is already known from earlier observations that it is often accompanied by chronic obstructive pulmonary disease (J44) and secondary malignant neoplasms of the brain (C79.3). Against the background of these relationships, 4 possible methods can be identified: computed tomography, bronchoscopy, magnetic resonance imaging, and positron emission tomography.

In Multimedia Appendix 3, we present an overview of the literature concerning the application of graph models for patient safety, procedure accuracy, and procedure efficacy.

As the graph grows in size, querying and updating can become increasingly resource-intensive, potentially impacting the model’s performance. In addition, algorithms are often based on the message-passing paradigm, which consists of node-by-node iteration (where each node aggregates information from neighboring nodes), which has significant limitations for graphs consisting of highly connected nodes [49]. For this reason, graphs with explicit heterogeneous edge distributions pose a problem and are difficult to evaluate because of the excessive compression of exponentially growing information in fixed-size vectors [50]. Recently, methods based on a different approach than the node-centered approach, which is based on differential geometry, have shown promising results for overcoming graph bottlenecks [51].

KGs often have incomplete data. If the graph does not have all the relevant entities and relationships, the model may lack important information. Incorrect or outdated information can lead to erroneous insights and conclusions.

To model procedure efficacy, it is necessary to analyze the system’s (ie, patient’s) response to an external perturbation (eg, administration of a drug). It is crucial to gain deep insights into a patient’s condition to assess the efficacy of procedures. Here, we want to present how these modeling tasks can be accomplished by HDTs [52] and by health digital shadows (HDSs), specifically when using graph-based methods.

For modeling procedure efficacy, the targets are not only binary values indicating patient outcomes or medical procedures but also biomarkers recorded as time series. For example, it is currently not possible to predict in advance whether a patient will face an allergic reaction or intolerance to medication. However, the ability to predict such events is relevant, as these reactions can be life-threatening and lead to treatment delays or morbidity.

Another problem is that physicians lack the time to review all available medical records of a patient. In the case of complex diseases, such as cancer, with various genetic mutations that could change during the procedure, prior or regular modeling is needed. Procedure adaptations or medication adjustments can be made in cases of side effects or low procedure efficacy.

For example, genetic polymorphisms in drug metabolism are already known for some substances, which must be determined before the substance is administered (eg, dihydropyrimidine dehydrogenase testing before 5-fluorouracil administration). Possible drug interactions must also be considered if several diseases are present. Many patients take several medications, which can sometimes lead to unforeseeable interactions. For patients, these interactions are increasingly difficult to oversee and evaluate, and individual factors also play a role in tolerability. The combination of different target values, along with different time spans, makes the problem definition much more complex than for patient safety and procedure accuracy.

Hence, an algorithmic architecture capable of considering a wide range of patient information and potential interactions would help model patient responses to medical procedures, predict patient outcomes, and test different counterfactual scenarios (ie, estimating a patient’s outcome on the basis of a treatment regime that is distinct from the one that was actually followed and the observed outcome) [53,54]. Therefore, we model the integration of different procedures within a process. Technically speaking, this task represents the need for the integration of various models along a process (ie, patient’s journey).

For this modeling task, the concept of a DT, which originated from engineering sciences [55], can be adapted into a clinical context. The aim of a DT is to generate a digital representation of an existing object within a process or time evolution in a digital space. Its concept is designed to combine time-evolving information from real objects, such as data generated from the Internet of Things, and mathematical modeling in the digital space [56]. Therefore, this digital representation allows the adaptation or optimization of processes in a digital space before transformations or modifications of real processes are made. For example, by integrating different and disparate information sources, it is possible to create different models not only for industrial production but also for the solution of critical and strategic problems, such as pollution, by generating universal avatars [57]. In summary, the design of a DT has the following main characteristics [56]:

In the industry, objects can be fully digitalized (with a bidirectional flow between the object and the digital object), which is limited only by the current evidence and understanding of the underlying complexity of the physical or biochemical processes or materials that should be mimicked. However, in the medical context, patients are not only too biologically complex but also cannot be objectivized in a straightforward way [59]. For this reason, we prefer to refer to HDSs, even when the literature currently mainly uses the term HDTs [60].

Furthermore, the connection between KGs and HDSs represents a logical step in the way different and disparate information sources are connected to create descriptive and predictive models. Relevant patient information as well as medical guidelines (in the form of KGs) can be integrated into a full graph containing patient data (eg, genetic or socioeconomic data) and additional relevant information (eg, drug structure and drug interactions).

In its role as an information and model integrating method, an HDS is well-suited for the following [61]:

The HDS concept is interesting for prospective analysis, that is, in the simulation and prediction of a patient’s health status and response to procedures. For prognostic purposes, it includes the ability to integrate all available patient information. Doctors treating patients can verify in advance the planned treatment regimen, such as medication or radiotherapy, on the patient’s digital representation.

The HDS model runs through all the interactions between the procedure and the patient. In the case of pharmacological treatment, medication interactions are simulated (eg, by integrating ML models for drug-drug interactions [62]). Then, an individual recommendation for the substance and the individual dosage and medication can be obtained by applying appropriate mathematical models (ie, ML). Additionally, if the treatment response is not sufficient, the HDS could propose additional procedures or further treatment strategies. Current databases for potential drug targets could also be considered.

Figure 4 shows how all these steps are included in a closed information loop. Bidirectional communication between the real world and the model is enabled. Therefore, the general model obtains patient-specific information at time point t_i (element 1). This enables the human HDS to make a prognosis (element 2), which can prospectively be used by the clinician to adapt the medical decision (element 3). Alternatively, model prediction can be used retrospectively to evaluate medical decisions by examining patient outcomes at time points t_i+1 (element 4). As time evolves, the patient’s journey produces increasing patient-specific information, leading to better possible adaptation of the general model to a personalized human HDS. As the HDS cohort grows, model readjustments for the general model are performed after certain time periods (element 5).

Thus, HDSs could increase the safety of drug therapy and its efficacy by providing personalized recommendations based on all available information [63]. For example, a similar concept, which is based on a whole-body representation via PBPK models and the integration of multiagent systems representing hepatocyte metabolism, can be used for pharmacology, particularly for the representation of the avatar’s response to a medication [64,65]. This example demonstrates that through the analysis of the avatar’s response, it is possible to plan procedures and even consider individual patient characteristics (eg, personalized metabolite levels that depend on the patient’s genetics) that lead to personalized patient care.

This example implies that procedures can be tested on the DHS to forecast procedure success; therefore, the recommendation level of a procedure can be assessed before being applied to a real patient. Such an application is equivalent to the so-called dynamic HDT, where forecasts based on time series are performed [65].

Consequently, HDSs can be useful for modeling procedure efficacy, as in the following examples [66]:

HDSs are increasingly being used for the estimation of the quality of patient care [55]. To this end, the prediction of patient outcomes is relevant, for example, in a personalized patient assessment to decide whether a procedure with high risk (death, side effects, patient burden, etc) is needed. This prevents the deterioration of a patient’s quality of life. Figure 5 shows an exemplary representation of a human HDS that integrates different data sources (such as omics, EHR, and documentation), which are measured in different time periods along the patient journey, into different models to predict patient outcomes.

As shown in Figure 5, after choosing the best possible medical procedure according to clinical guidelines, different information sources are integrated to predict QMs in different time periods depending on the patient’s evolution. For example, a whole-body (PBPK) model can be used to integrate patient-specific data within the HDT to predict a patient’s pharmacological response. Using pharmacological response prediction and patient data, another model can estimate patient outcomes (eg, through cluster analysis) for survival time prediction.

Chang et al presented a remarkable example of possible DHS realization through the creation of a pipeline for predicting patient outcomes, particularly survival heterogeneity, for patients with breast cancer [68,69]. In this implementation, 3 different data sources were integrated, where low-dimensional embeddings of clinical and molecular features enriched by an external annotation database were the main data sources. In this algorithm, a dimensional reduction method, Uniform Manifold Approximation and Projection, is implemented. This is a general-purpose manifold learning and dimension reduction algorithm to reduce the dimension of the data space. A dynamic HDS requires a pipeline that is accordingly adapted (ie, new data and additional models must be integrated and appropriately adapted according to other outcomes) [68].

Nitschke et al [68] presented a general and unspecialized design that is applicable for HDSs and is predictive, modular, evolving, informed, interpretable, and explainable. This design combines KGs and ensemble learning to represent the patient’s entire clinical journey and assist clinicians in their decision-making. Broad clinical application was shown by presenting 2 explicit cases: prostate cancer biopsy and glioma treatment decisions [68].

HDSs often oversimplify real systems. These approximations can lead to discrepancies between the model and the actual system. The accuracy of the HDS depends on the quality and granularity of the data used. Since more than one model is coupled to generate a full HDS, different models trained on different data and granularities can lead to unbalanced representations. Finally, poor or sparse data can lead to unreliable models [70]. Notably, all these problems are exacerbated by issues related to the aggregation and integration of different types of data that occur in health care, along with the strict privacy policies that must be followed [71], a problem that may be alleviated with the introduction of EHRs [72].

Creating and maintaining an HDS can be computationally intensive, requiring significant resources for simulation and real-time data processing. Integrating an HDS with other systems, sensors, and data sources can be complex and may require sophisticated interfaces and protocols. Ensuring that the HDS reflects real-time changes in the physical system can be challenging, especially in systems with high dynamics [73].

An interesting option is the relation and validation of HDSs using in vitro methods, such as organ-on-a-chip. This option opens the possibility to compensate for potential defects in model predictions and identify their potential problems by validating and comparing results with advanced in vitro patient representations [74]. This, however, is still a challenge in the field of oncology and requires more research to obtain a correct representation of patient outcomes.

The clinical scenario in Textbox 1 illustrates the integration of KGs and HDSs and demonstrates how our proposed framework evaluates all 3 QM categories while simultaneously assessing treatment safety in patients with comorbidities, ensuring guideline accuracy while allowing for personalized adaptation, and optimizing efficacy through predictive modeling of multiple therapeutic pathways for patients with non–small cell lung cancer. This case highlights how KG/HDS systems add value beyond static guidelines by dynamically integrating genomic data, resistance patterns, and real-world evidence to support complex clinical decision-making and establish prospective monitoring protocols.