Dynamic mirroring: unveiling the role of digital twins, artificial intelligence and synthetic data for personalized medicine in laboratory medicine

Introduction

The digital twins (DTs) paradigm, characterized by the creation of virtual replicas mirroring physical entities or processes, has recently gained substantial attention in several fields of informatics, technology, and healthcare. Especially in this last context, the advent of integrated digitalization and tailored artificial intelligence (AI) technologies, including advancements like natural language processing, genome sequencing, and imaging, has greatly facilitated the process of creating cornucopias of patient’s health data [1]. In addition, generative AI has been proven to generate synthetic healthcare data efficiently, with clear advantages over the experimental collected data, such as the possibility of virtually completely expanding a few datasets from a single patient to a large population [2].

The concept of DT starts from the data gathered from different sources for a single person (physical entity). This requires not only results from diagnostics (including laboratory and pathology test results), but also real-time updates of personal data by wearables, and self-tracking tools capturing physical activities, heart rate, etc. [3]. Thereafter, a specific virtual replica, referred to as a digital twin, is created [4, 5]. Additionally, the creation of DTs could be facilitated by using synthetically generated data [6], particularly in cases where population data are unavailable or limited. The concept of DT in healthcare can be applied to patients, healthcare communities, and entire hospital wards; in addition, DT can be set up for specific medical devices of IVD analytical platforms. Thus, DTs are emerging as transformative tools, offering unprecedented opportunities not only for diagnosis, prognosis, and real-time monitoring of patients but also for assessing the precision and accuracy of medical devices, entire wards, or clinical laboratories.

Currently, DTs in the field of medical disciplines are transitioning beyond their initial stages. It is interesting to note that while DTs are gaining attention in various fields, including healthcare, there is a lack of comprehensive documentation for certain disciplines, such as laboratory medicine [7]. DTs putative uses are rapidly evolving in “-omics” sciences [8], in oncology [9, 10], in radiology [11], cardiology [12, 13], and potentially in all other fields of medicine.

Interestingly, the digital replica that stands behind the digital twins concept is not a static representation, such as a digital shadow or a simulation model. DTs can be considered a dynamic, real-time simulation that mirrors its physical counterparts regarding physiological and operational aspects. Thus, simulation should also include forecasting some clinical measurement trajectories that can be predicted by varying parameters within the AI models, exploring the effect within a range of outcomes [14]. Furthermore, wearable sensors or new electronic monitoring devices have the potential to be fully integrated with DT, contributing to a more holistic approach to health monitoring [15]. Thus, digital twins might offer a window of opportunities to fix or solve the intricacies of healthcare processes, providing actionable insights for improved decision-making and patient outcomes. However, the successful application of this digital revolution requires addressing several technological issues still present, mainly due to the lack of data integration across several platforms and the management of interoperability in medical information generated during the patient’s journey within the healthcare system [4, 5].

The potential role of AI and DTs in moving from population-based to patient-based laboratory medicine

The fundamental principle of the scientific method involves employing structured datasets to validate hypotheses reliably. However, it is undebatable that the recent availability of large volumes of data in medical sciences (big data), the use of AI tools, in addition to the increasing availability of genetic and molecular information have suggested the possibility of stratifying data or generating equations and algorithms following unveiled tailored schemes, for personalized treatments [16]. Recent reviews described DTs as associated with using AI for personalized medicine [7]. AI for personalized medicine and digital twins are two related but distinct concepts. AI uses algorithms to perform tasks that typically require human intelligence, such as assisted diagnosis, prognosis, and treatment recommendation. DTs are high-resolution models of individual patients updated in real-time with available data [17]. Despite DT technology being often powered by AI, DTs allow simulation and modeling of patients’ results for diagnostic tests and treatments and planning predictive analyses [9]. Hence, DTs are technologies used to dynamically simulate the variation in one or several patients’ characteristics (e.g. disease response), based on the real-time effect of changing of one or multiple parameters (Figure 1).

Figure 1:

Possible roles of digital twins (DT) in laboratory medicine for creating clinically actionable information. (A) In contrast to AI alone, DTs might facilitate the clinical monitoring of patients and might be used for real-time prediction of worsening of clinical parameters; (B) DTs could be of utility for defining personalized RI, capturing inherent biological diversity among individuals and by using patient’s data; (C) by leveraging the pre-test probability, DTs can improve diagnostic test significance.

Physiological rhythms and biological variation for personalized medicine

In biological systems, rhythmic variations in the concentrations/activities of molecules can be observed over a wide range of time scales from fractions of a second to years. These variations are classified into three subgroups: ultradian, circadian, and infradian [18]. Ultradian variations represent ‘short-term rhythms’, with a frequency of fewer than 24 h, but typically with periods in the range of 20 min to 6 h. Circadian rhythms are daily oscillations corresponding to the earth’s 24 h rotation cycle, and infradian rhythms are periods that range from days to years [18, 19]. First of all, ultradian and circadian rhythms can greatly influence the diagnosis, screening and monitoring of disease. Diurnal significant variation of albumin and TSH was demonstrated by Andersen and colleagues. For albumin, variations could be attributable to varying hydration and metabolism during the day; for TSH, the pulsatile secretions from the thyroid, the control exerted by hypothalamus-releasing hormone, as well as dopaminergic mechanisms were speculated to explain diurnal variation [20]. Andersen et al. concluded that for both albumin and TSH, the use of floating RI or 2 h intervals RI could be important for the correct clinical classification of patients [20]. Moreover, for a large series of laboratory tests, including hormones, electrolytes and leukocytes it has been demonstrated that light exposure, sleeping conditions and chronophenotype (brevidiurnal, longidiurnal) factors affect their levels during day [18]. In the context of biological variations, AI and DTs applications can be helpful for predicting the influence of circadian and ultradian on laboratory parameters, for ameliorating diagnosis, screening and patient monitoring. In particular, DTs can serve to deliver realistic simulations to physicians, predicting customized, patient-centric perspectives on the physiological fluctuations of clinical values. Furthermore, they will prove beneficial in forecasting the anticipated changes resulting from alterations in patients’ conditions, taking into account chronotypes and utilizing insights into the rhythmicity of physiological parameters.

DTs in cancer therapy monitoring: advancements for tailored treatment strategies

In several cancer types, the elevated mortality is due to the recurrence and metastasis of the tumor after surgery. The evaluation of cancer during the therapy is crucial for the effectiveness of the treatment; however, tumor heterogeneity can cause a variation in therapeutic responses, posing challenges and inefficiencies in patients’ monitoring [21]. For instance, colorectal cancer (CRC) is known for its significant variability observable not only between patients but also within a single individual [22]. Patients’ lifestyle and exogenous factors, together with the remarkable genomic instability of this neoplasia, might lead to an elevated spatial heterogeneity of tumor cells. This latter term refers to clones genetically distinguished between groups of tumor cells with different metastatic and invasiveness that can be detected within the same individuals and that are not often easy to be determined [22]. It has been suggested that to capture the originality of the tumor, in addition to molecular pathology, liquid biopsy techniques could represent a promising diagnostic tool, allowing the evaluation of circulating tumor cells, circulating tumor DNA, or exosomes of microRNA. Pathology and liquid biopsy could serve for determining tumor heterogeneity and evolution, particularly in light of the intra-patient variability observed in the invasiveness of different metastases within the same patient [22, 23]. By using highly accurate patients’ data from imaging, molecular diagnostics, clinical biomarkers, clinical parameters, and biopsies, DTs may delineate a virtual copy of the patient’s tumor status, metastasis, and treatments. These data could be used by clinicians as well as by AI tools to simulate tumor response to therapies that could be delineated with extreme precision. Hence, DTs might provide reliable predictive representations of tumor growth and differentiation (also of different metastasis), finally offering the opportunity for improvements in the surgical treatment of target metastasis.

Simulation of personalized reference intervals (pRI)

For personalized reference data, individual-specific information is needed, and this can be obtained using repeated measurements in single individuals at specified conditions. For every analyte, each individual exhibits unique fluctuations, known as within-person variation (CV_P), which highlight the inherent biological diversity among individuals [16]. Recently, it has been reported that personalized reference intervals (pRI) might be derived by biological and analytical variation and using previous test results [24]. The idea behind the pRI is that instead of using a reference population to establish reference values, steady state results of an individual (homeostatic set point) could be used, together with data from biological variation (BV), to calculate the width of possible results for that analyte [25]. pRIs have also been suggested as important in evaluating individual metabolomics/proteomic results [26]. The estimation of pRI calls for a complex mathematical approach, including evaluating data quality (analytical accuracy). From this perspective, AI and DTs might potentially forecast (among the multiple predictions of patients’ information) pRI using individual and population data, particularly for the parameters with sufficient patient data to compute homeostatic set points and robust reference values(Figure 1). This presents an exceptional prospect for physicians to gain a comprehensive understanding of anticipated fluctuations in patients’ parameters, applicable not just in clinical environments but also for monitoring vital function impairments in outpatient settings by general practitioners.

DTs and synthetic data for improving test results interpretation

The correct interpretation of some diagnostic test results is based on the cut-off (threshold). The probabilistic approach at the basis of the dichotomization of patient’s results (namely positive or negative) can be either the optimization of sensibility/sensitivity (diagnostic test operating characteristics) or of positive/negative predictive values (PPV/NPV) [27]. However, the last two parameters depend not only on the operating characteristics of the test, but also on disease prevalence (pre-test probability). However, there is a concrete difficulty in correctly estimating the pre-test probability in most clinical contexts, and this might lead to overrating the probability of having the disease in the case of positive results (PPV) or not having the disease for negative results (NPV) [28]. Different methods can be used to estimate the pre-test probability [29]. Direct estimation of disease prevalence from population studies, or studies based on large data cohorts are mostly used. Also, the clinical reasoning process should yield a final probability of having a disease, despite it could be cumbersome [30]. Some specific tools have been speculated to be useful for achieving prevalence estimation, such as clinical prediction rules (that quantify the relative importance of clinical data points) or analysis models [31, 32]. DTs may have the potential to improve the accuracy of disease prevalence estimation by leveraging real-time data integration, finally leading to a more accurate post-test probability of disease [33] (Figure 1). In this context, synthetic data could also be used. Indeed, it has been demonstrated that synthetic patient and population health data for the state of Massachusetts could be useful for generating ML algorithms for predicting cancer risk [6]. Furthermore, the correct interpretation of diagnostic test results, also facilitated by reporting interval-specific probabilistic estimates of disease presence in laboratory reports, is expected to add actionable information to laboratory test results, enhancing interpretation of many tests, including autoimmuno and serological test result [28]. Finally, AI can finally facilitate the interpretation of clinical reports reporting interval-specific probabilistic estimates of disease.

DTs for improving automation

Total laboratory automation (TLA) could be described as systems able to prepare, sort, manage (centrifugation) and transport specimens to the laboratory analyzers to finally provide analytical results. These systems can perform several other tasks, such as measuring HIL indexes, automatically prioritizing emergency samples for reducing TAT and adequately storing samples [34]. The functions performed by TLA systems are similarly a series of rules devised by humans to manage what has been planned in preceding viewpoints. On the contrary, AI applied to TLA could autonomously adapt the flow of operations it carries out to minimize delays and recover from downtimes [35]. However, enforcing AI rules might lead to unforeseen behaviors, possibly resulting in catastrophic outcomes for laboratories. In automation, DTs serve as a “digital replica” of the program’s rules and algorithms. Thus, DTs will not only enable precise determination of the behavior of advanced AI-based platforms in various scenarios but also automatically generate tests to ascertain the outcome of each scenario. Finally, every variation in predefined rules will be automatically tested for possible errors, thus avoiding the risk of human errors and unpredicted outcomes.

Final remarks

Currently, there is an incredible interest in AI applications in healthcare. A huge limitation of ML is represented by data availability, and this is primarily true for healthcare information, which includes sensitive data and requires anonymization procedures. Several strategies are being studied to solve this problem; firstly, aggregation procedures could be used to gather data from different studies. Aggregative approaches present the limitation of being preferentially applicable only to datasets collected by specific studies. In contrast, this approach could be successful for some tools (diagnostic algorithms), but it could not be feasible for real-time application when patients are continuously evaluated over time. To overcome this problem, synthetic patient data technologies have already been proven to be an alternative, efficient, and reliable source of information for testing new processes involving healthcare; recently, a tool has been developed for synthetic electronic health record generation [6]. Although population-based ML and AI algorithms offer physicians support in predicting patients’ outcomes, they are generated from general data and do not consider the unique inherent complexity of each individual. The emerging technologies of DTs seem to be able to fill this gap. Enhancing the capabilities of AI to individualized data, DTs would present some peculiar features, such as the feasibility of real-time updates. In the field of clinical laboratory, there are many applications that could benefit from DTs. Some examples of potential improvements could be personalized RI, tailored BV rhythmicity and more accurate monitoring of cancer responses through biochemical markers. Further, DTs of analytical platforms (including the automation workflow) and devices, will make advantages for optimizing procedures, including stat tests execution in consolidated labs, and teaching to new personnel. However, AI and DTs’ widespread adoption in healthcare in the future can be limited by data silos and lack of data integration, issues that should be solved for upscaling laboratory technological advancements [36]. All in all, while DTs widespread adoption in healthcare is not imminent, it is not far off. Shortly, this technology will likely offer innovative and definitive solutions for dynamically evaluating treatments and more precise diagnosis, thereby achieving personalized medicine.