System for dynamic clinical term normalization and entity linking using a modular natural language processing pipeline

Abstract

A system for the dynamic normalization of clinical terms and entity linking using a modular natural language processing (NLP) pipeline, consisting of: a preprocessing module configured to receive unstructured clinical text and perform operations such as de-identification, spelling correction, abbreviation expansion, and sentence segmentation; a tokenization and embedding processing unit configured to segment the preprocessed text into sub-word units and generate context-related embeddings using transformer-based biomedical language models; a processing unit for dynamic vocabulary expansion configured to update a specialized dictionary in real time by monitoring new abbreviations, acronyms, or institution-specific terms from incoming clinical data streams and feedback from clinical staff; a candidate concept generator configured to retrieve potential ontology mappings by performing an approximate nearest neighbor search via an embedding index linked to medical knowledge bases; a semantic disambiguation controller configured to evaluate candidate concepts using a graph neural network model that takes into account the local linguistic context, patient-specific metadata, and temporal relationships between entities; an ontology linking interface configured to map disambiguated entities to standardized concepts in one or more controlled vocabularies; and a feedback learning module configured to incorporate clinician corrections into the model parameters through reinforcement-based updates, thus enabling continuous self-improvement of entity normalization.

Description

Field of invention

[0001] The invention relates generally to natural language processing (NLP) in the biomedical and clinical field and in particular to a system and device for the dynamic normalization of clinical terms and entity linking using a modular NLP pipeline that can be integrated into electronic health record (EHR) systems, clinical databases and platforms to support medical decisions. Background of the invention

[0002] Clinical documentation is inherently heterogeneous. Physicians, nurses, and other healthcare professionals use diverse terminologies, abbreviations, synonyms, and local codes to describe diagnoses, symptoms, treatments, and laboratory findings. Existing clinical terminology normalization systems typically rely on static dictionaries or pre-trained models that cannot adapt to evolving terminologies, institution-specific vocabularies, or context-dependent usage. This leads to inaccuracies in downstream applications such as clinical decision support, pharmacovigilance, predictive analytics, and regulatory reporting. Furthermore, current solutions rarely offer a modular architecture that allows for domain-specific customization and dynamic integration with multiple controlled vocabularies such as SNOMED-CT, ICD-10, LOINC, RxNorm, and UMLS.

[0003] A persistent technical challenge lies in ensuring both high accuracy and high computing power for real-time term normalization, particularly in large-scale deployments in hospitals and distributed healthcare networks. Furthermore, there is a lack of hardware-integrated devices that enable clinical NLP processing on-premises and at the edge, without relying solely on cloud infrastructure, which limits responsiveness in critical care environments.

[0004] Processing clinical terminology is one of the most complex challenges in biomedical informatics, primarily due to the heterogeneous and dynamic nature of the terminology used across different healthcare systems. Clinicians frequently resort to idiosyncratic expressions, abbreviations, acronyms, institution-specific vocabulary, and even personal habits when documenting patient encounters. A patient presenting with a myocardial infarction may be documented as "MI," "acute coronary syndrome," "STEMI," "heart attack," or with other variations, depending on the clinician, specialty, or hospital convention. This linguistic variability makes it extremely difficult for downstream systems—such as clinical decision support, predictive modeling, pharmacovigilance systems, and population health analytics—to reliably interpret and integrate the meaning of clinical reports.Consequently, over the past two decades, research and commercial solutions have attempted to address the problem of normalizing clinical terms and linking entities. However, every approach has suffered from significant limitations that restrict scalability, adaptability, and accuracy in real-world applications.

[0005] Rule-based and heuristic-driven approaches are also widely used. In these approaches, subject matter experts manually code syntactic and semantic rules to normalize clinical terms and make entities unambiguous. For example, a rule might specify that the word "seropositive" precedes the token "RA," which more likely means rheumatoid arthritis, while the token "echocardiogram" indicates the right atrium. While such systems can achieve greater precision in narrow fields, they suffer from poor generalizability. Writing, testing, and maintaining thousands of rules is labor-intensive and prone to instability. Rules that work in one clinical context may fail in another, and extending such systems to new medical specialties or multilingual clinical texts is prohibitively expensive.Furthermore, rule-based systems often fail when faced with new abbreviations or unexpected writing styles, resulting in reduced robustness in production environments.

[0006] With the advent of machine learning and statistical natural language processing (NLP), more data-driven approaches were introduced, in which clinical terms were normalized using classifiers trained on annotated corpora. Models such as conditional random number generators (CRFs), support vector machines (SVMs), and later, recurrent neural networks (RNNs) were used to learn patterns in clinical texts and map them to standardized terms. These models were superior to purely rule-based systems in handling variability and ambiguity, but their performance was limited by the availability of high-quality, labeled training data. Annotating large clinical corpora requires expertise, is time-consuming, and is further constrained by concerns regarding patient privacy.Therefore, training datasets are often focused on specific institutions or medical specialties, resulting in poor cross-domain generalization. Furthermore, traditional machine learning models require fixed vocabularies and must be retrained whenever the lexicon changes, limiting their ability to dynamically adapt to new terminologies emerging in medical practice.

[0007] The rise of deep learning, and in particular transformer-based architectures like BERT and its biomedical applications, has significantly advanced the field of clinical NLP. These models utilize self-attention mechanisms and contextual embedding to capture nuanced relationships in clinical texts, enabling more precise disambiguation of ambiguous entities. Fine-tuning large, pre-trained models on domain-specific corpora has led to substantial progress in the recognition and association of clinical entities. However, despite their power, these models have several drawbacks when applied to the real-world problem of normalizing clinical terms. First, transformer-based models are computationally intensive, requiring significant hardware resources for both training and inference.Hospitals and clinics with limited infrastructure may find the on-site use of these models impractical, leading to reliance on cloud services, which in turn raises concerns about patient data privacy and regulatory compliance. Second, these models are not inherently modular; updating one part of the system, such as the vocabulary or the embedding mechanism, often requires retraining the entire model, resulting in inefficiency. Third, even the most advanced transformers suffer from context sensitivity issues when faced with domain shifts, such as encountering abbreviations or linguistic patterns specific to a new hospital, department, or geographic region. This lack of adaptability limits their robustness and longevity in clinical use.

[0008] Another research focus is on ontology-based approaches, in which clinical terms are mapped to structured medical knowledge graphs. These approaches use graph traversal techniques and semantic similarity measures to find the most suitable concept for a given entity. While ontology-based methods offer interpretability and leverage structured knowledge, they are often limited by the completeness and quality of the underlying ontology. Many clinical ontologies have gaps, inconsistencies, or overlapping definitions that can lead to errors in entity linking. Furthermore, integrating multiple ontologies is challenging due to differences in scope, granularity, and coding standards. Therefore, ontology-based systems often require significant human effort and cannot dynamically adapt to newly introduced concepts without manual intervention.

[0009] In recent years, hybrid systems have emerged that combine elements of machine learning, deep learning, and ontology-based approaches. For example, candidate terms can be generated through embedding similarity and then evaluated or disambiguated using ontology-based constraints. While hybrid approaches show promise in overcoming some limitations of individual methods, they still face significant challenges regarding scalability, modularity, and real-time adaptation. Many hybrid systems are designed as monolithic architectures where different components are tightly coupled. This makes it difficult to update or replace individual modules without disrupting the entire pipeline. This rigidity contrasts sharply with the rapidly evolving landscape of clinical language, where modular adaptability is critical.

[0010] Another major drawback of existing solutions lies in the lack of mechanisms for incorporating feedback from clinicians and end users. In practice, clinical NLP systems frequently make errors that are noticed and corrected by human experts. However, many current systems lack the ability to capture these corrections and adjust their models accordingly. This leads to repeated errors, reduced clinician confidence, and resistance to adoption. Reinforcement learning and online adaptation techniques have been proposed in research contexts, but few production systems have successfully integrated them in a way that is reliable, efficient, and compliant with medical data management regulations.

[0011] From a deployment perspective, many existing solutions rely heavily on cloud-based infrastructures for computational scalability. While cloud platforms provide the resources needed to train and infer large models, they introduce latency and raise serious privacy concerns, particularly regarding the handling of protected health information (PHI). Regulations such as HIPAA in the US and GDPR in Europe impose stringent requirements on the storage, transfer, and processing of clinical data. Cloud-based solutions risk violating these regulations without comprehensive safeguards, which in turn increases complexity and costs. Conversely, on-premises hospital infrastructure often lacks the specialized hardware accelerators required for the efficient execution of modern NLP models, leading to performance bottlenecks.Therefore, there is an urgent need for dedicated hardware specifically designed for clinical NLP workloads.

[0012] Furthermore, interoperability remains a persistent challenge. Healthcare facilities often use different EHR systems with varying degrees of support for standards such as HL7 and FHIR. Clinical NLP solutions must integrate seamlessly with these heterogeneous systems while ensuring consistent normalization across facilities. Existing solutions frequently suffer from interoperability issues, resulting in fragmented data and inconsistent mappings. This fragmentation hinders cross-facility research, data sharing, and large-scale analytical initiatives.

[0013] In addition to these technical limitations, many existing systems fail to account for the dynamic nature of medical knowledge itself. Medicine evolves rapidly, with new diseases, treatments, and diagnostic criteria being introduced regularly. During the COVID-19 pandemic, for example, new terms such as "long COVID," "cytokine storm," and novel drug regimens quickly entered clinical discourse. Static systems were unable to process these terms in real time, highlighting the importance of adaptability. Similarly, rare diseases and experimental therapies often fall outside the scope of standard ontologies and require systems capable of dynamically integrating new knowledge without necessitating complete retraining or manual curation.

[0014] Dictionary-based, rule-based, statistical, deep learning, ontology-driven, and hybrid approaches each offer partial solutions to the problem of normalizing clinical terms and linking entities. However, none of them fully addresses the combination of adaptability, scalability, modularity, efficiency, and data privacy required for use in real-world healthcare settings. Dictionary- and rule-based systems are static and fragile, machine learning models are limited by data availability and retraining costs, deep learning models are resource-intensive and insufficiently modular, ontology-driven systems rely on incomplete knowledge bases, and hybrid systems often lack flexibility and feedback integration. The reliance on cloud infrastructure further complicates their use in privacy-sensitive environments.Taken together, these disadvantages underline the need for a new system that combines modular design, dynamic vocabulary expansion, real-time adaptability, reinforcement-based feedback learning, and dedicated hardware acceleration, thus overcoming the limitations of existing solutions. Summary of the invention

[0015] The invention provides a system for the dynamic normalization of clinical terms and entity linking using a modular NLP pipeline. This system comprises several interconnected modules in a pipeline configuration, including a preprocessing module, a tokenization and embedding processing unit, a dynamic vocabulary expansion module, a candidate concept generator, a semantic disambiguation engine, and an ontology linking interface. The system utilizes reinforcement-based feedback learning to adapt entity mappings in real time based on user corrections and evolving datasets.

[0016] Another aspect of the invention is a device for modular clinical NLP processing, implemented as a standalone or rack-mounted hardware structure, comprising a high-throughput text capture unit, GPU / TPU-accelerated processing cores, FPGA-based embedding generators, a knowledge graph cache, and secure communication interfaces for integration with EHR systems. The device includes dedicated circuitry for accelerating similarity search and vector disambiguation tasks in clinical NLP pipelines.

[0017] The main objective of the present invention is to provide a system and associated device that can dynamically normalize clinical terms and link them to standardized medical ontologies with high accuracy, adaptability, and efficiency. The invention aims to solve the long-standing problem of the variability of clinical language by developing a modular NLP pipeline capable of processing abbreviations, synonyms, acronyms, institution-specific terminology, and rapidly evolving medical vocabularies.Another objective of the invention is to overcome the limitations of static, dictionary-based or monolithic deep learning models by creating an architecture in which individual modules for preprocessing, tokenization, embedding generation, concept candidate retrieval, semantic disambiguation and ontology linking can be updated, replaced or optimized independently without disrupting the overall pipeline.

[0018] Another objective of the invention is the integration of mechanisms for dynamic vocabulary expansion. This allows the system to automatically adapt to new terms, abbreviations, and concepts introduced in clinical practice, research, or emerging diseases. By integrating reinforcement-based feedback learning, the invention aims to capture corrections from clinicians and user comments, thereby creating a self-improving system that evolves and adapts to practice over time. In this way, the system reduces repetitive errors and increases the confidence of clinicians and healthcare administrators.

[0019] Another objective of the invention is the real-time normalization of clinical units in environments where low-latency processing is critical, such as emergency rooms, intensive care units, and telemedicine platforms. To achieve this, the invention also provides a dedicated hardware device that accelerates NLP workloads using GPU, TPU, and FPGA cores, coupled with specialized knowledge graph cache storage and secure processing enclaves. This device design ensures that sensitive medical text data can be processed locally in hospitals and clinics, thereby reducing reliance on cloud infrastructures that may violate data privacy regulations such as HIPAA and GDPR.

[0020] The invention also aims to ensure interoperability between different medical institutions, standards, and jurisdictions by enabling flexible integration with electronic health records, clinical databases, and health information exchange systems using HL7, FHIR, and other interoperability standards. A further objective is to ensure that the system and device can harmonize data across various ontologies, including SNOMED-CT, ICD-10, LOINC, RxNorm, and UMLS, thereby supporting consistent normalization for comprehensive analyses, pharmacovigilance, and regulatory reporting.

[0021] A further objective of the invention is to improve robustness and reliability in production environments. To this end, a modular housing for the hardware device is being developed, featuring redundant power supplies, cooling, and failover mechanisms, making it suitable for use in hospitals. This ensures uninterrupted performance and scalability in distributed healthcare networks. Furthermore, a secure processing framework is being provided, incorporating hardware-based encryption and controlled access mechanisms to guarantee data integrity and compliance with strict medical governance guidelines. BRIEF DESCRIPTION OF THE FIGURE

[0022] These and other features, aspects, and advantages of the present invention will be better understood if the following detailed description is read with reference to the accompanying drawing, in which the same symbols consistently represent the same parts. The following applies: Fig. Figure 1 shows a block diagram of a system for the dynamic normalization of clinical terms and entity linking using a modular NLP pipeline.

[0023] Experts will also recognize that the elements in the drawing are shown for the sake of simplicity and are not necessarily to scale. For example, the flowcharts illustrate the process by highlighting the main steps to enhance understanding of the aspects of this disclosure. Furthermore, with regard to the design of the device, one or more components of the device may be represented in the drawing by conventional symbols, and the drawing may show only the specific details relevant to understanding the embodiments of this disclosure, so as not to clutter the drawing with details that are readily apparent to those skilled in the art after reading this description. Detailed description of the invention

[0024] For a better understanding of the inventive principles, reference is made below to the embodiment shown in the drawing, which is described in specific terminology. However, this does not limit the scope of the invention. Changes and further modifications of the illustrated system, as well as further applications of the inventive principles, are possible, as would normally occur to a person skilled in the art in the field of invention.

[0025] It is clear to the person skilled in the art that the preceding general description and the following detailed description are exemplary and explanatory of the invention and are not intended as a limitation of it.

[0026] References in this specification to “an aspect”, “another aspect”, or similar expressions mean that a particular feature, structure, or property described in connection with the embodiment is included in at least one embodiment of the present disclosure. Therefore, occurrences of the expressions “in one embodiment”, “in another embodiment”, and similar expressions in this specification may all refer to the same embodiment, but need not.

[0027] The terms "includes," "include," or other variations thereof are intended to cover non-exclusive inclusion, so that a process or method that includes a list of steps may not only contain those steps but may also include other steps not expressly listed or inherent in such process or method. Likewise, the statement "includes..." in the case of one or more devices, subsystems, elements, structures, or components does not, without further limitations, preclude the existence of other devices, subsystems, elements, structures, components, or additional devices, subsystems, elements, structures, or components.

[0028] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as understood by a person skilled in the art in the field of the invention. The system, methods, and examples provided here serve only for illustration and are not to be construed as a limitation.

[0029] Embodiments of the present disclosure are described in detail below with reference to the attached drawing.

[0030] In Fig.Figure 1 is a block diagram of a system for the dynamic normalization of clinical terms and entity linking using a modular NLP pipeline. The system 100 comprises: a preprocessing module (102) configured to receive unstructured clinical text and perform operations such as de-identification, spell correction, abbreviation expansion, and sentence segmentation; a tokenization and embedding processing unit (104) configured to segment the preprocessed text into subword units and generate contextual embeddings using transformer-based biomedical language models; and a dynamic vocabulary expansion module (106) configured to update a specialized dictionary in real time by monitoring new abbreviations, acronyms, or institution-specific terms from incoming clinical data streams and clinician feedback.a candidate concept generator (108) configured to retrieve potential ontology mappings by performing an approximate nearest neighbor search in an embedding index linked to medical knowledge databases; a semantic disambiguation controller (110) configured to evaluate candidate concepts using a graph neural network model that considers the local linguistic context, patient-specific metadata, and temporal relationships between entities; an ontology linking interface (112) configured to map disambiguated entities to standardized concepts in one or more controlled vocabularies, including SNOMED-CT, ICD-10, LOINC, RxNorm, or UMLS;and a feedback learning module (114) configured to incorporate corrections from clinicians into the model parameters through reinforcement-based updates, thus enabling continuous self-improvement of entity normalization.

[0031] In one embodiment, the preprocessing module (102) also includes a de-identification engine configured to replace protected health information with anonymized tokens while preserving the syntactic coherence of the clinical narrative, enabling subsequent NLP modules to process the text without violating the patient's privacy.

[0032] In one embodiment, the tokenization and embedding processing unit (104) uses a hybrid embedding scheme that includes both static, PubMed-trained embeddings and dynamic, context-aware embeddings derived from a fine-tuned transformer, so that embeddings can be adaptively selected depending on the computational limitations of the deployment environment.

[0033] In one embodiment, the processing unit (106) for dynamic vocabulary expansion uses active learning strategies to suggest possible term extensions, with the validation provided by the clinician being logged via a human-in-the-loop interface and subsequently passed on to update both the dictionary and the embedding model without requiring complete retraining.

[0034] In one embodiment, the concept candidate generator (108) performs an approximate nearest neighbor query using a vector database optimized for biomedical embeddings, wherein the database is indexed using locality-sensitive hashing (LSH) or hierarchically navigable small-world graphs (HNSW) to reduce latency in large-scale similarity searches across millions of medical concepts.

[0035] In one embodiment, the controller for semantic disambiguation (110) includes temporal inferences, so that the entity linking is influenced by longitudinal patient records. This allows for differentiation between chronic and acute diseases by modeling their temporal coincidence with laboratory values ​​and medication regimens.

[0036] In one embodiment, the ontology linking interface (112) also includes a bidirectional mapping mechanism that supports both the forward normalization of clinical free text terms into standardized ontology concepts and the backward extension of standardized concepts into institution-specific variants for localized reporting and documentation purposes.

[0037] In one embodiment, the feedback learning module (114) uses reinforcement signals generated from clinician-approved corrections, with the model parameters of the disambiguation engine being updated by policy gradient optimization to maximize alignment with the preferences of human experts across successive normalization tasks.

[0038] In one embodiment, the modular clinical NLP processing comprises: a text ingestion unit configured to interface with hospital information systems, EHR databases, and HL7 / FHIR-compliant data streams; a compute core consisting of one or more GPU, TPU, or FPGA accelerators optimized for transformer inference, graph-based disambiguation, and vector similarity search; a knowledge graph cache memory implemented as fast non-volatile memory arrays configured to store subgraphs of standardized medical ontologies and associated embedding vectors; and a secure processing enclave configured to encrypt and isolate sensitive clinical data during NLP computation to ensure HIPAA and GDPR compliance.and a modular housing configured with redundant cooling, power supply and failover controllers to support uninterrupted operation in hospital environments.

[0039] In one embodiment, the text capture unit also includes a hardware-level HL7 / FHIR parser that converts clinical messages into structured intermediate representations, thereby reducing preprocessing latency compared to purely software-based parsers.

[0040] The invention describes a system and device for the dynamic normalization of clinical terms and entity linking, based on a modular natural language processing (NLP) pipeline and a specialized hardware architecture. The aim is to process unstructured clinical texts from diverse sources such as physician notes, discharge summaries, laboratory reports, and diagnostic imaging annotations, as well as to normalize heterogeneous terminology into standardized medical concepts. The design combines advanced techniques of transformer-based embedding, graph-based disambiguation, and reinforcement learning with hardware accelerators and privacy-friendly mechanisms.

[0041] The system begins with a preprocessing module that receives the raw clinical text. This preprocessing technique employs de-identification techniques that utilize named entity recognition models in combination with rule-based templates to identify and mask personally identifiable information such as patient names, medical record numbers, and addresses. Crucially, the de-identification module maintains syntactic coherence by replacing sensitive tokens with pseudo-identifiers. This ensures that downstream modules can analyze the text without structural disruptions.The preprocessing also applies spell checking using character-level transformer models trained on clinical corpora, abbreviation expansion based on local dictionaries supplemented by contextual embedding, and sentence segmentation optimized for irregular punctuation styles commonly found in medical notes. This ensures that the text input for subsequent modules is standardized, privacy-compliant, and structurally consistent.

[0042] After preprocessing, the text enters the tokenization and embedding processing unit. Here, a hybrid embedding architecture is used, combining sub-word-level tokenization with biomedical transformers. The system uses fragment word or sentence fragment tokenization to break clinical terms down into atomic subunits, thus ensuring robustness against spelling errors and neologisms. Embeddings are generated using a dual-channel framework: static embeddings pre-trained on PubMed and MIMIC III corpora, and dynamic contextual embeddings generated by finely tuned transformer models such as ClinicalBERT. The technique dynamically selects between static and contextual embeddings based on computational constraints and latency requirements.For example, in high-throughput batch normalization, static embeddings may be prioritized for efficiency reasons, while in real-time ICU settings, contextual embeddings offer higher accuracy.

[0043] The dynamic vocabulary expansion unit works continuously to adapt the system to evolving clinical terminology. The technique monitors incoming text streams and identifies out-of-vocabulary (OOV) terms by comparing them to the system's current dictionary. OOV terms are then processed by an active learning framework that uses clustering and similarity metrics to suggest possible meanings. Clinician feedback is incorporated through an interactive interface, where suggested expansions are validated or corrected.

[0044] After confirmation, these updates are incorporated into both the dictionary and the embedding model. To avoid retraining from scratch, the system uses embedding retrofitting, interpolating new term vectors based on existing ontological neighbors. This ensures that the embedding space remains consistent while simultaneously expanding the coverage.

[0045] The concept candidate generator serves as the next stage in the pipeline. This module performs an ANN (Approximate Nearest Neighbor) search across a vector database that indexes embeddings of standardized medical concepts from ontologies such as SNOMED-CT and RxNorm. To ensure low latency, the ANN search technique uses hierarchically navigable small-world graphs (HNSW) or location-sensitive hashing (LSH), depending on hardware capacity. For each detected entity, the search returns a ranked list of ontology concept candidates optimized for a high hit rate. The use of an ANN-based approach ensures that even incomplete embeddings or noisy input terms can yield meaningful candidate matches, thus reducing the probability of missed assignments.

[0046] The disambiguation of candidates is performed by the semantic disambiguation engine. This technique utilizes graph neural networks (GNNs) that consider both the local and global context. Locally, the system applies attention-based sequence modeling to evaluate the syntactic and semantic dependencies between tokens in a sentence. Globally, the system retrieves relevant patient metadata, such as age, comorbidities, lab results, and the temporal sequence of previous diagnoses, encoded as feature vectors. The GNN integrates these multimodal signals and performs message passing between nodes representing candidate concepts to evaluate them for contextual match.For example, if the abbreviation "RA" is found, the engine evaluates not only linguistic clues but also the patient's history; in a patient with elevated rheumatoid factor, "RA" would likely correspond to rheumatoid arthritis, whereas in a cardiology note with echocardiographic parameters, the assignment to "right atrium" is more probable. Temporal reasoning is embedded in the GNN through recurring message-transmission steps that model the evolution of clinical conditions over time and thus differentiate between chronic and acute entities.

[0047] Once the correct entity is selected, the ontology mapping interface performs a bidirectional mapping. In forward mode, normalized entities are mapped to standardized identifiers to ensure interoperability, enabling downstream analysis and harmonization across institutions. In reverse mode, standardized ontology terms are expanded back into institution-specific variants to ensure compatibility with local reporting workflows. This bidirectional design ensures that the system not only contributes to harmonized data aggregation but also integrates seamlessly with the individual documentation practices of clinicians and institutions.

[0048] A key innovation lies in the feedback learning module. This technology captures clinician corrections directly during use, for example, when a physician overrides a suggested match. These corrections are logged as amplification signals and used to update the parameters of the disambiguation engine. A policy gradient amplification learning algorithm adjusts the model weights to maximize the probability of clinician-approved matches in future cases. To prevent catastrophic forgetting, the system employs elastic weight consolidation. This preserves important parameters from previous data while incorporating new corrections. This amplification-based adaptation enables the pipeline to continuously self-improve over time, adapting its behavior to evolving clinical usage without requiring regular, complete retraining.

[0049] The device design enables hardware-based acceleration of the aforementioned pipeline. The text capture unit includes a hardware HL7 / FHIR parser that directly converts structured messages into intermediate representations. By embedding the parsing logic in the firmware, the device reduces preprocessing latency and ensures compatibility with heterogeneous hospital data streams. The computing core integrates GPUs, TPUs, and FPGAs, each optimized for different workloads: GPUs accelerate transformer inference, TPUs handle the generation of large embedding batches, and FPGAs implement quantized transformer models and ANN retrieval for vector similarity. This heterogeneous computing architecture ensures both throughput and efficiency.

[0050] The knowledge graph's cache enables fast access to ontology subgraphs. Concepts and embeddings are stored in compressed sparse-row (CSR) format, allowing simultaneous traversal and similarity calculations. For example, during disambiguation, the GNN queries both the embedding space and the ontology hierarchy; the cache enables these searches without network calls to external servers, significantly reducing latency.

[0051] Security is ensured by the secure processing enclave, which runs the NLP pipeline in a trusted execution environment. All data is encrypted in memory, so clinical texts are never exposed in plaintext outside the enclave. This feature is essential for compliance with HIPAA, GDPR, and other data privacy regulations, allowing hospitals to use the device without fear of data leaks.

[0052] The modular housing is designed for flexible deployment scenarios. In large hospitals, the device can be rack-mounted to support integration with central EHR systems. In field hospitals or emergency vehicles, a portable form factor can be deployed at the perimeter, enabling immediate NLP processing without reliance on cloud infrastructure. Redundant cooling systems, dual power supplies, and automatic failover controllers ensure hospital-grade reliability and enable continuous operation in critical care environments.

[0053] In distributed deployment scenarios, multiple devices can synchronize updates in federated learning mode. In this setup, each device performs local training with clinical corrections without transmitting raw patient data. Instead, model updates are encrypted and aggregated across devices to improve global performance. This federated approach ensures system-wide adaptability while respecting local data privacy restrictions.

[0054] In summary, the invention combines advanced technical modules and specialized hardware into a coherent architecture that addresses the weaknesses of existing solutions. The modular NLP pipeline enables plug-and-play integration of components, dynamic vocabulary expansion ensures adaptability to new terms, the disambiguation engine integrates contextual and temporal reasoning, and the reinforcement-based feedback loop enables continuous learning. The device design accelerates computation, ensures compliance with data protection regulations, and offers flexibility for deployment in hospitals and in the field. Together, these elements provide a scalable, efficient, and privacy-friendly solution for the dynamic normalization of clinical terms and the linking of entities, supporting real-world healthcare environments at scale.

[0055] The system comprises a modular NLP pipeline implemented on a computer framework. Each module can be independently replaced, updated, or customized. The preprocessing module performs de-identification, sentence segmentation, spell correction, and abbreviation expansion using rule-based and statistical approaches. The tokenization and embedding processing unit uses word-level tokenization and transformer-based embeddings trained on biomedical corpora. These embeddings are stored in a memory-optimized vector database and are readily accessible.

[0056] The dynamic vocabulary expansion unit continuously monitors clinical text streams and integrates new abbreviations, acronyms, or rare technical terms into the vocabulary through active learning and clinically validated updates. The candidate concept generator uses nearest neighbor approximation in embeddings to suggest an ordered group of ontology candidates for each recognized entity. The semantic disambiguation controller uses graph neural networks (GNNs) and contextual attention mechanisms to select the most appropriate mapping from among the candidate entities, taking into account the local sentence context, patient-specific data, and the temporal evolution of the terms.

[0057] The drawing and the preceding description show examples of embodiments. Those skilled in the art will recognize that one or more of the described elements can be combined to form a single functional element. Alternatively, certain elements can be divided into several functional elements. Elements of one embodiment can be added to another embodiment. For example, the sequence of the processes described here can be changed and is not limited to the manner described here. Furthermore, the actions of a flowchart need not be implemented in the sequence shown; nor does it necessarily have to be performed by all actions. Actions that are not dependent on other actions can also be performed in parallel with the other actions. The scope of the embodiments is in no way limited by these specific examples.Numerous variations are possible, whether explicitly stated in the specification or not, such as differences in structure, dimensions, and material usage. The range of embodiments is at least as broad as specified in the following claims.

[0058] Advantages, further benefits, and problem solutions have been described above with reference to specific embodiments. However, the advantages, benefits, problem solutions, and all components that can lead to an advantage, benefit, or solution occurring or becoming more apparent are not to be construed as critical, necessary, or essential features or components of individual or all claims. REFERENCES 100 A system for the dynamic normalization of clinical terms and the linking of entities using a modular NLP pipeline. 102 Preprocessing module 104 Tokenization and Embedding Processing Unit 106 Module for dynamic vocabulary expansion 108 Candidate Concept Generator 110 Semantic Disambiguation Controller 112 Ontology Linking Interface 114 Feedback Learning Module

Claims

[1] A system for the dynamic normalization of clinical terms and entity linking using a modular natural language processing (NLP) pipeline, consisting of: a preprocessing module configured to receive unstructured clinical text and perform operations such as de-identification, spelling correction, abbreviation expansion, and sentence segmentation; a tokenization and embedding processing unit configured to segment the preprocessed text into sub-word units and generate context-related embeddings using transformer-based biomedical language models; a processing unit for dynamic vocabulary expansion configured to update a specialized dictionary in real time by monitoring new abbreviations, acronyms, or institution-specific terms from incoming clinical data streams and feedback from clinical staff; a candidate concept generator configured to retrieve potential ontology mappings by performing an approximate nearest neighbor search via an embedding index linked to medical knowledge bases; a semantic disambiguation controller configured to evaluate candidate concepts using a graph neural network model that takes into account the local linguistic context, patient-specific metadata, and temporal relationships between entities; an ontology linking interface configured to map disambiguated entities to standardized concepts in one or more controlled vocabularies; and a feedback learning module configured to incorporate clinician corrections into the model parameters through reinforcement-based updates, thus enabling continuous self-improvement of entity normalization. [2] System according to claim 1, wherein the preprocessing module further comprises a de-identification engine configured to replace protected health information with anonymized tokens while preserving the syntactic coherence of the clinical narrative, thereby enabling subsequent NLP modules to process the text without violating the patient's privacy. [3] System according to claim 1, wherein the tokenization and embedding processing unit uses a hybrid embedding scheme that includes both static embeddings and dynamic context-related embeddings derived from a fine-tuned transformer, so that embeddings can be adaptively selected depending on the computational limitations of the deployment environment. [4] System according to claim 1, wherein the processing unit for dynamic vocabulary expansion uses active learning strategies to suggest possible term expansions, wherein the validation provided by the clinician is logged via a human-in-the-loop interface and subsequently passed on to update both the dictionary and the embedding model without requiring complete retraining. [5] System according to claim 1, wherein the controller for semantic disambiguation comprises temporal inferences such that the entity linking is influenced by longitudinal patient records, thereby differentiating between chronic and acute diseases by modeling their temporal coincidence with laboratory values ​​and medication courses. [6] System according to claim 1, wherein the ontology linking interface further comprises a bidirectional mapping mechanism that supports both the forward normalization of clinical free text terms into standardized ontology concepts and the backward extension of standardized concepts into institution-specific variants for localized reporting and documentation purposes. [7] System according to claim 1, wherein the feedback learning module uses reinforcement signals generated from clinician-approved corrections, wherein the model parameters of the disambiguation engine are updated by policy gradient optimization to maximize alignment with the preferences of human experts across successive normalization tasks.

Images