The widespread adoption of electronic health records (EHRs) has fundamentally transformed how patient information is generated, stored, and accessed in health care systems worldwide [1,2]. These digital repositories enable comprehensive documentation of medical histories, clinical notes, diagnostic results, and other sensitive data, frequently captured in unstructured textual formats [3]. Safeguarding this information is critical not only to preserve patient privacy and confidentiality but also to ensure the integrity and trustworthiness of modern health care systems [4,5].

Core principles of medical data governance, including privacy, confidentiality, integrity, and nonrepudiation, guide how patient information is accessed, shared, and protected [6]. Anonymization has thus become a vital requirement, enabling secondary uses of health data for research and quality improvement while minimizing the risk of patient reidentification [7,8]. The challenge of effective anonymization is particularly acute for unstructured clinical text, which contains complex linguistic patterns, domain-specific terminology, and contextually embedded sensitive information [9,10].

To address these concerns, comprehensive data protection regulations have been established worldwide. In the United States, the Health Insurance Portability and Accountability Act (HIPAA) mandates the deidentification of Protected Health Information (PHI) before secondary use, specifying either Safe Harbor or Expert Determination methods for compliance [11,12]. The European General Data Protection Regulation (GDPR) sets rigorous standards for processing health data, emphasizing techniques such as anonymization and pseudonymization with stringent consent requirements [13,14]. Similarly, Brazil’s Lei Geral de Proteção de Dados Pessoais (LGPD) aligns with international norms, underscoring the need for robust technical solutions to protect sensitive information in the Brazilian health care context [15].

Over time, various anonymization methods have been developed to address these regulatory requirements [16]. Conventional rule-based and pattern-matching techniques, though reliable for structured identifiers such as social security numbers and standardized date formats, often struggle with the linguistic complexity, variability, and context dependence of natural language in medical texts [17,18]. These approaches typically require extensive manual rule creation and maintenance, limiting their scalability across different health care institutions and documentation styles.

Machine learning and deep learning models, particularly those developed for Named Entity Recognition (NER), offer greater adaptability and accuracy in identifying sensitive entities within clinical narratives [19-21]. Transformer-based architectures, including BERT (Bidirectional Encoder Representations from Transformers) and specialized biomedical variants such as BioBERT and ClinicalBERT, have demonstrated strong performance across multiple deidentification benchmarks by effectively capturing contextual and domain-specific linguistic patterns [22,23]. However, these approaches require large annotated corpora, extensive hyperparameter tuning, and considerable computational resources for both training and inference [24].

Most recently, large language models (LLMs) such as GPT-4, Llama, and their instruction-tuned derivatives have emerged as promising alternatives for anonymizing medical texts [25,26]. These models can capture nuanced context, understand complex semantic relationships, and identify entities that may elude traditional pattern-matching systems [27]. Evidence suggests that locally deployable LLMs can achieve competitive performance with state-of-the-art NER systems while offering flexibility in entity categories through prompt engineering [28,29]. Nevertheless, LLM-based approaches remain at a nascent stage in the clinical deidentification literature, with persistent concerns regarding model interpretability, computational overhead, hallucination phenomena, and potential risks when processing sensitive data through external services [30].

Even more exploratory are efforts in quantum computing paradigms, which have begun to investigate whether quantum algorithms can be applied to tasks such as pattern recognition, optimization, and privacy-preserving computation [31,32]. Quantum Natural Language Processing (QNLP) represents an emerging field that leverages quantum mechanical principles for linguistic tasks, with theoretical foundations suggesting potential advantages in expressiveness and computational efficiency [33,34]. The application of quantum machine learning to text processing has been explored through various frameworks, including the translation of classical operations into Quadratic Unconstrained Binary Optimization (QUBO) problems suitable for quantum annealing [35,36]. Neuromorphic quantum computing platforms have emerged that emulate quantum behavior through memristor-based circuits, enabling practical deployment of quantum-inspired algorithms on classical hardware [37]. Despite growing interest, the practical application of quantum-enhanced approaches to clinical machine learning remains largely unexplored, and no systematic evaluations have compared quantum-assisted architectures with contemporary LLM-based or transformer-based methods for clinical text anonymization. The aim of this study is to evaluate the effectiveness and computational efficiency of LLM-based and quantum-enhanced hybrid anonymization architectures on Portuguese clinical notes from a single Brazilian tertiary cardiology hospital, across 5 protected health information entity categories. We hypothesized that integrating a global QUBO-based optimization layer over the final attention layer of an instruction-tuned LLM would yield higher anonymization quality, particularly in precision, than the same LLM operating in stand-alone prompt-based mode, while remaining computationally tractable for routine de-identification workflows.

Recent studies have explored the use of LLMs for clinical text deidentification. Wiest et al [28] developed an open-source anonymization framework based on locally deployable LLMs, demonstrating that models such as Llama-3-70B can achieve sensitivity exceeding 0.992 on German clinical letters. Their work emphasized 2 critical aspects: the importance of local deployment for privacy preservation and the flexibility of prompt-based approaches for customizing entity definitions. Similar findings have been reported for English clinical texts, where instruction-tuned LLMs have achieved competitive performance against fine-tuned NER systems [26].

Systematic reviews of deidentification approaches have documented the evolution from rule-based systems to hybrid machine learning methods, with transformer-based models representing the current state of the art [19,38]. The 2014 i2b2-UTHealth corpus remains the most frequently used benchmark, where top-performing systems achieve a micro-averaged F₁-score of 0.940 [39]. However, most benchmarks focus on English texts, with limited evaluation of methods for other languages, including Portuguese.

In the quantum computing domain, research in QNLP has demonstrated theoretical advantages for text classification and semantic analysis tasks [33,40]. Hybrid quantum-classical approaches have been proposed that integrate quantum circuit optimization with classical neural network components [41]. However, to our knowledge, no prior work has specifically evaluated quantum-enhanced architectures for medical text anonymization.

In this context, this study makes several contributions. First, we present a comparative evaluation of LLM-based and quantum-enhanced hybrid architectures for medical text anonymization using a purpose-built corpus of Portuguese clinical notes. Second, we demonstrate the application of neuromorphic quantum optimization algorithms to improve anonymization precision while maintaining high recall. Third, we provide a detailed explanation of how transformer attention mechanisms can be reformulated as QUBO problems and solved via quantum annealing. Fourth, we provide a comprehensive assessment of both quality metrics and computational efficiency. Finally, we discuss the implications of our findings for regulatory compliance and practical deployment in health care settings.

The primary aim of this study was to perform a comparative evaluation of stand-alone LLM anonymization approaches and quantum-enhanced hybrid architectures for the deidentification of Portuguese clinical notes. Specifically, we aimed to determine whether the proposed Dynex-QML framework could improve anonymization performance compared with conventional prompt-based LLM approaches, particularly regarding precision, false-positive (FP) reduction, and overall F₁-score, while preserving high sensitivity and practical computational efficiency.

Quantum annealing is an optimization technique that exploits quantum mechanical phenomena to find the global minimum of an objective function [44]. The method is particularly effective for solving combinatorial optimization problems that can be formulated as finding the ground state (lowest energy configuration) of a quantum system.

The mathematical foundation of quantum annealing rests on the Ising model [45,46], which describes a system of interacting binary spins. For a system with N binary variables si∈{-1,+1}, the energy function (Hamiltonian) is expressed as:

where hi represents the local field acting on spin i, and Jij represents the interaction strength between spins i and j. The goal is to find the spin configuration that minimizes this energy function.

For computational convenience, the Ising model can be reformulated as a QUBO problem by transforming the binary variables from {-1,+1} to {0,1} domain using the substitutionsi=2xi-1. The resulting QUBO formulation is:

where xi∈{0,1} and Q are real-valued matrices encoding both the linear and quadratic terms of the objective function. This formulation is particularly suitable for representing constraint satisfaction problems and optimization tasks that arise in NLP [35,47].

Modern LLMs are built on the transformer architecture, which relies fundamentally on the attention mechanism to process sequential data [48]. The multihead attention mechanism computes weighted representations of input tokens by learning which parts of the input sequence are most relevant for predicting each output token.

For a given input sequence, the attention mechanism computes 3 matrices: Query (Q), Key (K), and Value (V), derived from the input embeddings through learned linear transformations. The attention weights are computed as:

where dk is the dimensionality of the key vectors. The softmax operation normalizes the attention scores to create a probability distribution over the input tokens.

In a typical transformer-based LLM, multiple layers of multihead attention are stacked, with each layer refining the representations. The final layer produces output logits that are passed through a softmax function to generate probability distributions over the vocabulary for each token position. In text generation tasks, including anonymization, these probabilities guide the selection of output tokens.

The Dynex platform implements neuromorphic quantum computing through an architecture that emulates quantum behavior using memristor-based circuits [37]. Unlike gate-based quantum computers that operate on physical qubits in quantum superposition, Dynex uses algorithmic qubits, that is, mathematical representations of quantum states that are simulated within GPU video random access memory (VRAM).

The core innovation of the Dynex platform lies in its ability to construct problem-specific quantum circuits composed of linear and quadratic quantum gates. These gates are not physical quantum components but rather mathematical models of memristive elements whose behavior is governed by systems of ODEs. The platform solves these ODEs through numerical integration to simulate the quantum annealing process.

For linear quantum gates representing individual qubits, the voltage dynamics are described by:

where vn is the qubit voltage (representing the qubit state), QGm is the quantum gate state computed as QGm=(1.0-pivi)/2.0, Wm is the quantum gate weight (encoding the QUBO coefficients), and pi is the polarity indicator (+1 for i§amp;gt;0, -1 for i≤0).

For quadratic quantum gates representing interactions between qubits, more complex dynamics govern the system evolution, incorporating both gradient terms and rigidity constraints that ensure the quantum circuit converges to a stable solution. The platform automatically optimizes the annealing parameters (α,β,γ,δ,ϵ,ε,ζ) for each problem class to maximize solution quality.

The use of memristor-based simulation allows the Dynex platform to leverage the massive parallel processing capabilities of modern GPUs while maintaining the mathematical properties of quantum annealing. This approach provides computational efficiency that scales with the number of GPU cores available, enabling practical deployment of quantum-inspired algorithms on existing hardware infrastructure.

To assess the effectiveness and efficiency of the anonymization methods, we evaluated both quality-related and computational performance metrics on 500 of the 1000 clinical texts in the dataset, selected at random to constitute the test set.

The quality of anonymization for each method was quantitatively determined based on three main criteria:

These counts served as the basis for calculating the primary quality metrics for each entity type considered (<NAME>, <DATE>, <LOCAL>, <ORGANIZATION>, and <ID>):

The macro–F₁-score was calculated as the unweighted mean of F₁-scores across all entity types to provide an overall performance measure. To estimate statistical uncertainty, 95% CIs for precision, sensitivity (recall), and F₁-score were computed using bootstrap resampling based on aggregate multinomial TP, FP, and FN counts for each model and entity category. For each bootstrap iteration, TP, FP, and FN counts were resampled from the observed aggregate count distribution, and the corresponding metrics were recomputed. We used 10,000 bootstrap iterations. Macro-level CIs were calculated by averaging the bootstrapped entity-level metrics across the 5 entity categories.

Pairwise model comparisons were performed using empirical 2-sided bootstrap tests on the difference in F₁-score. Prespecified within-size comparisons were Dynex-QML (Llama 8B) versus stand-alone Llama 3.1 8B and Dynex-QML (Llama 70B) versus stand-alone Llama 3.3 70B. For each comparison, the empirical P value was calculated as twice the smaller tail probability of the bootstrap distribution of the F₁-score difference. Statistical significance was defined as P<.05.

For computational performance, uncertainty in mean internal processing time (TI) and mean total elapsed time (TT) was estimated using note-level bootstrap resampling of the 500 observed per-note timing measurements for each model. Pairwise time comparisons used paired bootstrap resampling by clinical note ID, preserving the matched structure of the timing data across models. Empirical 2-sided P values were calculated from the bootstrap distribution of paired mean time differences, with statistical significance defined as P<.05.

Interannotator agreement was evaluated using the 3 independent annotation sets produced by the 3 annotators for the same 100 randomly selected clinical notes. Pairwise comparisons were performed for the 3 possible annotator pairs.

Agreement was assessed at the entity level for NAME, DATE, ID, LOCAL, and ORGANIZATION. An entity was considered concordant only when both annotators identified the same complete text span and assigned the same protected-entity category.

Entity-level F₁-scores were calculated separately for each protected-entity category and annotator pair. For each category, the mean pairwise F₁-score was calculated as the arithmetic mean of the 3 pairwise comparisons. The macro–F₁-score for each annotator pair was calculated as the unweighted mean of the 5 entity-level F₁-scores. The overall mean pairwise macro–F₁-score was calculated as the arithmetic mean of the 3 pairwise macro–F₁-scores. Percentile-based 95% CIs were estimated using 1000 note-level bootstrap samples with replacement.

This study was conducted in accordance with the Institutional Review Board (CAPPesq – IRB Number 68, Brazilian National Research Ethics System), under registration number CAAE 89528325.8.0000.0068. The clinical notes were retrospectively obtained from routine clinical care. Original notes were accessed only by authorized researchers within a secure institutional environment for the manual annotation procedure. Before model development and evaluation, all annotated identifiers were replaced with synthetic values to generate pseudonymized clinical texts. The ethics committee waived the requirement for individual informed consent because of the retrospective study design and the data-protection measures adopted. All data-handling procedures complied with the Brazilian General Data Protection Law, LGPD.

The Heart Institute research team executed all anonymization algorithms on remote high-performance servers to ensure reproducible computational servers provided by Dynex within a Sandbox environment. For each model evaluation, 2 complementary time metrics were recorded to characterize computation performance:

Both processing time metrics were recorded for each text and aggregated to provide per-text averages (mean and SD) and totals for the complete test set of 500 clinical notes. This dual-metric approach enables a clearer distinction between intrinsic model efficiency and external latency effects introduced by remote execution. TI reflects the intrinsic computational efficiency of the anonymization model itself, whereas TT better approximates the latency experienced in real-world deployment scenarios where remote execution and network communication are required.

The evaluation framework was assessed using the following standard classification metrics: TP, FP, FN, precision, recall (sensitivity), and F₁-score, along with processing time (TI and TT) to capture computational efficiency.

Table 2 summarizes the computing environment used for all model evaluations, including hardware specifications and software configurations that supported the execution of both LLM-based and quantum-enhanced models.

The same 100 randomly selected clinical notes were independently reannotated by 3 annotators. Pairwise entity-level agreement results and the numbers of entities identified by each annotator are presented in Table 3. The overall mean pairwise macro–F₁-score across the 5 protected-entity categories was 0.786 (SD 0.038; 95% CI 0.700‐0.869). Pairwise macro–F₁-scores ranged from 0.744 to 0.816.

Agreement was highest for NAME, with a mean pairwise F₁-score of 0.966 (SD 0.015; 95% CI 0.929‐0.990), followed by DATE at 0.947 (SD 0.018; 95% CI 0.928‐0.963), ORGANIZATION at 0.849 (SD 0.045; 95% CI 0.793‐0.897), and LOCAL at 0.782 (SD 0.131; 95% CI 0.649‐0.895). ID showed the lowest mean pairwise F₁-score at 0.388 (SD 0.138; 95% CI 0.000‐0.778).

The full set of anonymization quality metrics is presented in Table 4. Performance varied substantially across entity categories and model architectures. Dynex-QML (Llama 70B) achieved the highest macro–F₁-score (0.855, 95% CI 0.823‐0.880), followed by Dynex-QML (Llama 8B; 0.733, 95% CI 0.709‐0.756), stand-alone Llama 3.3 70B (0.726, 95% CI 0.704‐0.747), and stand-alone Llama 3.1 8B (0.602, 95% CI 0.588‐0.615). At the macro level, Dynex-QML (Llama 70B) improved F₁-score by 0.128 compared with stand-alone Llama 3.3 70B (95% CI 0.091‐0.163; empirical 2-sided bootstrap P<.001). Dynex-QML (Llama 8B) improved macro–F₁-score by 0.131 compared with stand-alone Llama 3.1 8B (95% CI 0.104‐0.158; P<.001).

Figure 1 displays the entity-level F₁-score comparison across all protected entity categories for the 4 evaluated models. To further characterize model behavior beyond F₁-score, Figure 2 shows the macro-sensitivity and macro-precision trade-off across architectures. Dynex-QML (Llama 70B) achieved the most favorable overall balance, combining higher macro-precision with preserved macro-sensitivity compared with the other evaluated models.

Figure 3 provides an entity-level error analysis by separating FP and FN counts across protected entity categories. In the context of clinical text anonymization, FNs are particularly important because they represent missed protected health information and therefore potential residual privacy risk, whereas FPs mainly represent overanonymization and potential loss of clinical information utility. Across entity categories, Dynex-QML (Llama 70B) reduced FP counts substantially compared with stand-alone Llama 3.3 70B, especially for NAME, ID, LOCAL, and ORGANIZATION entities. Importantly, this reduction in overanonymization was not accompanied by an increase in FN. The model maintained similarly low FN counts for DATE, NAME, LOCAL, and ORGANIZATION entities, and only one FN for ID entities. These results suggest that the main advantage of Dynex-QML (Llama 70B) was not improved precision alone, but a more favorable privacy-utility trade-off, combining high sensitivity with fewer unnecessary anonymization errors.

This study presents a comparative assessment of anonymization performance using both conventional LLMs and a quantum-enhanced hybrid architecture (Dynex-QML). Based on the primary macro–F₁-score metric, Dynex-QML-70B achieved the best overall performance (0.855, 95% CI 0.823‐0.880), followed by Dynex-QML-8B, Llama-3.3-70B, and Llama-3.1-8B. The observed macro–F₁-score improvement of Dynex-QML (Llama 70B) over stand-alone Llama-3.3-70B was supported by bootstrap analysis (mean difference +0.128, 95% CI 0.091‐0.163; P<.001). Entity-level analyses showed statistically significant improvements for most categories, with the strongest gains for ID and NAME entities. Computationally, Dynex-QML (Llama 70B) also showed a statistically significant reduction in total elapsed time compared with stand-alone Llama 3.3 70B (mean paired difference −0.55, SD 2.36; 95% CI −0.75 to −0.34 seconds per note; P<.001). These findings demonstrate that quantum-enhanced optimization provides substantial improvements in anonymization quality, particularly reducing FP rates and improving entity-level predictions.

One of the important observations is the precision improvement achieved by quantum-enhanced models. For the NAME entity, Dynex-QML-70B achieved a precision of 0.971 compared to 0.624 for the stand-alone Llama-70B, driven primarily by reduction in FP (10 vs 202). Similar trends were observed for DATE (precision 0.960 vs 0.919) and LOCAL (precision 0.919 vs 0.873). These improvements suggest that the QUBO-based optimization layer acts as an effective corrective mechanism constraining the model’s tendency to overidentify borderline entities, a known limitation of prompt-driven LLM anonymization approaches [28]. By imposing structured optimization criteria, the quantum-enhanced framework appears to reduce overgeneralization while preserving high recall.

A key consideration in clinical anonymization is that sensitivity and precision do not have equivalent practical implications. FN correspond to missed protected health information and therefore represent the primary privacy and regulatory risk. In contrast, FP correspond to overanonymization, which may reduce the clinical interpretability and downstream research utility of the resulting text. For this reason, the performance gain observed with Dynex-QML (Llama 70B) should not be interpreted as a simple precision improvement. Rather, the relevant finding is that the model achieved higher precision while preserving high sensitivity across most entity categories. This pattern is illustrated in Figure 2, where Dynex-QML (Llama 70B) occupies the most favorable region of the macro-sensitivity and macro-precision space, and in Figure 3, where the reduction in FP is not accompanied by a meaningful increase in FN.

Relative performance across entity categories also highlights persistent challenges in clinical text anonymization. All models achieved their strongest results on DATE entities, which is expected given their relatively standardized formats and limited contextual ambiguity in clinical narratives. In contrast, the ID category remained the most difficult to anonymize reliably; even the highest-performing model (Dynex-QML-70B) reached only a 0.548 F₁-score. This limitation likely reflects the substantial variability in identifier formats across institutions, specialties, and documentation contexts, as well as the inherent difficulty in distinguishing clinically relevant numerical values from patient-specific identifiers.

The interannotator agreement analysis provides important context for interpreting the model results. The overall mean pairwise entity-level macro–F₁-score of 0.786 (SD 0.038) indicates generally consistent application of the annotation guidelines, although agreement varied across entity categories. This variability suggests that some model errors may reflect both model limitations and uncertainty in the reference annotations. However, all models were evaluated against the same operational gold standard, and the improvements achieved by Dynex-QML 70B were observed across multiple entity categories. The lower and less precise agreement observed for ID should be interpreted cautiously because this category was represented by very few entities in the reannotated subset.

Our findings align with extended recent work on LLM-based deidentification. Previous work [28] reported that Llama-3-70B achieved over 0.992 sensitivity on German clinical letters, consistent with our observation of high sensitivity rates (0.951‐0.990) for the larger models. However, their work did not address FP problems that we found to be substantially mitigated by quantum-enhanced optimization.

The performance levels achieved in our study are somewhat lower than those reported on the English i2b2 benchmarks, where top systems achieve a micro-averaged F₁-score of 0.940 [39]. This difference reflects several factors: the Portuguese language context introduces different linguistic challenges; our entity definitions may be broader than i2b2 categories; and prompt-based approaches generally show lower performance than fine-tuned models [26]. Nevertheless, relative improvements from quantum enhancement should generalize across contexts.

In the quantum computing domain, our work represents one of the first practical applications of quantum-enhanced methods to clinical NLP tasks. Prior work on QNLP has focused primarily on theoretical analyses and simple classification tasks [33,40]. The success of our QUBO-based approach suggests that quantum optimization can provide practical benefits even with current hybrid quantum-classical architectures.

From an operational standpoint, Dynex-QML-70B showed a lower total elapsed time than Llama-70B (7.97, SD 2.87 seconds per note vs 8.52, SD 3.32 seconds per note). This finding is notable because a quantum-enhanced approach involves additional optimization steps that might be expected to increase computational overhead. The efficiency gains may result from the global optimization reducing the need for multiple inference passes or from the QUBO solver converging faster than iterative token generation in challenging cases.

For smaller models, computational trade-offs are less favorable for the quantum-enhanced approach. Dynex-QML-8B requires substantially more processing time than Llama-8B (8.59, SD 3.62 seconds per note vs 5.22, SD 2.12 seconds per note), suggesting that the overhead of quantum optimization may be more significant relative to the faster base model inference. This makes the 70B configuration more suitable for production deployment where quality is prioritized over raw speed.

Improvements in precision achieved by quantum-enhanced approaches have important implications for regulatory compliance. Under HIPAA Safe Harbor provisions, all 18 categories of identifiers must be removed with no residual information that could enable reidentification [11]. High FN rates (missed entities) create direct compliance risks, while high FP rates (overanonymization) can degrade data utility for research purposes.

The Dynex-QML-70B model’s combination of high recall (minimizing compliance risk) and high precision (preserving data utility) represents an optimal balance for health care applications. For sensitive entity types such as names, the model’s 0.951 recall ensures that very few identifying names remain in anonymized output, while the 0.971 precision ensures that clinical terms are not erroneously removed.

Several limitations of this study should be acknowledged. First, our evaluation was conducted using Portuguese outpatient clinical notes obtained within a single health care system. InCor is part of HC-FMUSP, the largest academic health care complex in Latin America, serving a large and diverse patient population through multiple clinical pathways, different documentation styles, clinical specialties, and languages. Institutional differences in writing patterns, abbreviation usage, entity distributions, and documentation workflows may substantially affect anonymization performance in external settings. Second, the dataset size of 1000 texts, while sufficient for comparative evaluation, may not capture the full variability of clinical documentation. Third, although the same 100 randomly selected notes were independently reannotated by 3 annotators, the agreement analysis covered only 20% of the evaluation set and 10% of the complete corpus. It also included 3 of the 5 original annotators rather than the complete annotation team. Fourth, our implementation of quantum-enhanced optimization uses simulated quantum annealing on GPU hardware rather than actual quantum processors, though this reflects the current state of accessible quantum computing resources and the practical deployment model of the Dynex platform.

Additionally, we did not conduct systematic hyperparameter optimization for either the prompt design or the QUBO formulation parameters, suggesting that further improvements may be achievable. Energy consumption measurements were not included in this study due to methodological challenges in accurately comparing energy usage between classical LLM inference and hybrid quantum-classical systems. The Dynex platform’s use of algorithmic qubits emulated in GPU VRAM introduces complexity in energy accounting, as the quantum annealing process involves intricate patterns of memory access and computation that are not straightforwardly comparable to standard neural network inference. Specifically, the algorithmic qubits are not isolated physical devices whose draw can be metered, but mathematical state variables coresident in GPU video memory alongside the classical LLM inference workload, and the Dynex platform operates on a distributed GPU network where node-level power readings are confounded by other tenant workloads sharing the same hardware. We therefore took the methodologically conservative position of removing the energy comparison from this paper rather than reporting a comparison whose denominator is not well defined. Future work should develop standardized methodologies for energy measurement in neuromorphic quantum computing systems to enable fair comparisons with classical approaches.

Future research should extend this evaluation to additional languages and clinical domains, including radiology reports, pathology notes, and nursing documentation. Systematic benchmarking against fine-tuned NER models and other state-of-the-art systems would provide additional context for the quantum-enhanced approach. As quantum hardware continues to evolve, evaluation on alternative quantum computing platforms would help characterize the generalizability of the QUBO-based anonymization approach.

The QUBO formulation approach may also be applicable to other clinical NLP tasks where global optimization could improve upon sequential processing, such as relation extraction, temporal reasoning, and clinical summarization. Exploring these applications represents a promising direction for quantum-enhanced health care AI. Additionally, investigating the interpretability of the quantum optimization process could provide insights into which linguistic features most strongly influence anonymization decisions.

This study demonstrates that quantum-enhanced hybrid architectures can achieve substantial improvements in medical text anonymization compared to stand-alone LLMs. The Dynex-QML-70B model achieved the highest overall performance (macro–F₁-score 0.855, macro-precision 0.794, and macro-sensitivity 0.971) while slightly reducing processing time compared to Llama-3.3-70B (7.97, SD 2.87 seconds per note vs 8.52 SD 3.32 seconds per note). The primary advantage of quantum enhancement lies in substantially improved precision by 25% (0.794 for Dynex-QML-70B vs 0.636 for Llama-3.3-70B) while maintaining high sensitivity with a 2% increase (0.971 for Dynex-QML-70B vs 0.949 for Llama-3.3-70B).

The detailed exposition of the quantum-enhanced methodology demonstrates how transformer attention mechanisms can be reformulated as QUBO problems and solved via neuromorphic quantum annealing. This approach leverages the global optimization capabilities of quantum computing to address the precision challenges inherent in prompt-based LLM anonymization.

These findings highlight the growing maturity of hybrid quantum-classical architectures and their promise for accurate and scalable anonymization in real-world health care applications. As regulatory requirements for health data protection continue to evolve and the volume of clinical text data grows, quantum-enhanced approaches may represent a viable path toward sustainable solutions for clinical text anonymization that balance quality, efficiency, and compliance.