Paper deep dive
Human-Centered Explainable AI for TinyML Edge Devices: A Pareto-Based Selection Framework with LLM-Guided Design
Zeinab Dehghani, Dhavalkumar Thakker, Koorosh Aslansefat, Kuniko Paxton, Bhupesh Kumar Mishra, Baseer Ahmad, Rameez Raja Kureshi
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 88%
Last extracted: 8/10/2026, 4:14:31 AM
Summary
The paper proposes a human-centered framework for selecting Explainable AI (XAI) methods for TinyML edge devices, specifically in clinical applications. The framework integrates Large Language Model (LLM)-guided design to map stakeholder preferences to candidate XAI methods, followed by deterministic feasibility filtering based on hardware constraints (SRAM, Flash) and Pareto-based optimization to balance explanation fidelity, stability, and deployment cost. A proof-of-concept evaluation using skin lesion classification demonstrates the framework's ability to identify Pareto-efficient trade-offs.
Entities (17)
Relation Signals (13)
Evaluation → uses → HAM10000
confidence 95% · proof-of-concept skin lesion classification study using the HAM10000 dataset
Edge-AI → enables → local edge devices
confidence 90% · Edge Artificial Intelligence (Edge AI) enables the deployment of AI models directly on local edge devices
Framework → evaluates → Fidelity
confidence 90% · The framework exposes trade-offs among explanation fidelity, stability, and proxy-based deployment cost
Framework → evaluates → Stability
confidence 90% · The framework exposes trade-offs among explanation fidelity, stability, and proxy-based deployment cost
LLM → guides → XAI method selection
confidence 90% · integrates a large language model (LLM)-guided design interface that maps qualitative stakeholder preferences to candidate XAI methods
XAI → servesas → human-AI interface
confidence 90% · explainable artificial intelligence (XAI) can serve as a human-AI interface intended to support healthcare professionals’ and patients' understanding
Evaluation → uses → MobileNetV3-Small
confidence 90% · MobileNetV3-Small [23]
Framework → uses → Pareto Optimization
confidence 90% · followed by deterministic feasibility filtering and Pareto-based optimization
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Edge Artificial Intelligence (Edge AI) enables the deployment of AI models directly on local edge devices, while such deployments are subject to strict resource constraints, particularly in clinical applications requiring local and timely inference. In such contexts, explainable artificial intelligence (XAI) can serve as a human-AI interface intended to support healthcare professionals' and patients' understanding of model predictions and informed decision-making. To fulfill this role, XAI method selection for TinyML deployments can be formulated as a human-centered multi-objective design problem that jointly considers qualitative stakeholder preferences, explanation quality, and proxy-based deployment cost. We propose a framework that integrates a large language model (LLM)-guided design interface that maps qualitative stakeholder preferences to candidate XAI methods, followed by deterministic feasibility filtering and Pareto-based optimization. The framework exposes trade-offs among explanation fidelity, stability, and proxy-based deployment cost while characterizing their implications for explanation quality and estimated deployment feasibility. A proof-of-concept evaluation on a skin lesion classification task illustrates how the framework systematically compares candidate XAI methods and identifies Pareto-efficient trade-offs. The present evaluation covers the computational selection stages, while physical MCU deployment and empirical human-expert validation remain outside the scope of this study.
Tags
Links
- Source: https://arxiv.org/abs/2608.07091v1
- Canonical: https://arxiv.org/abs/2608.07091v1
Trouble viewing inline? Open PDF directly →
Full Text
81,667 characters extracted from source content.
Expand or collapse full text
Human-Centered Explainable AI for TinyML Edge Devices: A Pareto-Based Selection Framework with LLM-Guided Design Zeinab Dehghani Dhavalkumar Thakker Koorosh Aslansefat Kuniko Paxton Bhupesh Kumar Mishra Baseer Ahmad and Rameez Raja Kureshi Zeinab Dehghani, Dhavalkumar Thakker, Koorosh Aslansefat, Kuniko Paxton, Bhupesh Kumar Mishra, Baseer Ahmad, and Rameez Raja Kureshi are with the School of Digital and Physical Sciences, University of Hull, Cottingham Road, Hull, HU6 7RX, United Kingdom (e-mails: zeinab.dehghani068@gmail.com, D.Thakker@hull.ac.uk, e-mail: K.Aslansefat@hull.ac.uk, k.paxton@hull.ac.uk, Bhupesh.Mishra@hull.ac.uk, Baseer.Ahmad@hull.ac.uk, R.Kureshi@hull.ac.uk). Corresponding author: Kuniko Paxton Abstract Edge Artificial Intelligence (Edge AI) enables the deployment of AI models directly on local edge devices, while such deployments are subject to strict resource constraints, particularly in clinical applications requiring local and timely inference. In such contexts, explainable artificial intelligence (XAI) can serve as a human-AI interface intended to support healthcare professionals’ and patients’ understanding of model predictions and informed decision-making. To fulfill this role, XAI method selection for TinyML deployments can be formulated as a human-centered multi-objective design problem that jointly considers qualitative stakeholder preferences, explanation quality, and proxy-based deployment cost. We propose a framework that integrates a large language model (LLM)-guided design interface that maps qualitative stakeholder preferences to candidate XAI methods, followed by deterministic feasibility filtering and Pareto-based optimization. The framework exposes trade-offs among explanation fidelity, stability, and proxy-based deployment cost while characterizing their implications for explanation quality and estimated deployment feasibility. A proof-of-concept evaluation on a skin lesion classification task illustrates how the framework systematically compares candidate XAI methods and identifies Pareto-efficient trade-offs. The present evaluation covers the computational selection stages, while physical MCU deployment and empirical human-expert validation remain outside the scope of this study. IEEEkeywords EdgeAI, TinyML, Explainability, LLM, Resource-restricted, Medical Image 1 Introduction The use of artificial intelligence (AI) has expanded rapidly in healthcare. In particular, driven by the need to protect patient data privacy and the demand for real-time processing, there is growing interest in edge AI, which runs AI models on resource-constrained platforms such as microcontrollers (MCUs) and low-power embedded devices. In these environments, models must operate under strict latency, memory, and energy constraints, motivating active research in TinyML to maintain predictive performance under limited computational resources [1, 2, 3]. On the other hand, medical AI requires not only predictive performance but also safety and reliability. Therefore, explainable artificial intelligence (XAI) is regarded as an important component for examining the basis of diagnostic predictions and identifying potential risks [4]. In fact, since clinicians and patients may use explanations to support their understanding of AI-assisted decisions, explanations can serve as an important interface that supports appropriate trust and informed use [5, 6, 7, 8]. However, many existing XAI methods are designed with computationally rich environments in mind, and challenges remain regarding their application to TinyML environments. For example, activation- and gradient-based methods such as Class Activation Mapping (CAM) [9] and Gradient-weighted Class Activation Mapping (Grad-CAM) [10] require access to internal model information, while perturbation-based methods such as Local Interpretable Model-agnostic Explanations (LIME) [11] and Randomized Input Sampling for Explanation (RISE) [12] involve multiple rounds of inference, resulting in high execution costs on devices [13, 14]. In addition, hardware heterogeneity and environmental variability may affect explanation stability over time, reducing predictability in operational settings [15, 3]. As a result, in real-world edge environments, the choice of explanation method often depends on the developer’s experience and judgment, and alignment with hardware constraints and operational requirements is not sufficiently considered. Furthermore, XAI in healthcare is not merely a model analysis tool but can also function as a decision-support interface for healthcare professionals and patients. It is well known that the quality and presentation of explanations can influence user trust and reliance behavior, and inappropriate explanations can lead to over-reliance or erroneous judgments [5, 6, 16]. Nevertheless, existing research tends to treat the selection of explanation methods as a purely technical problem, and the systematic incorporation of intended-user needs and usage context into the selection process has received limited attention. Moreover, with the diversification of XAI methods in recent years, selecting a method suited to the intended use and operational environment has become a critical challenge in itself. Recent research has highlighted the importance of human-centered XAI, emphasizing the selection and presentation of explanations that align with users’ objectives and decision-making contexts [17]. In other words, explanations should not only be technically correct but also understandable and usable by healthcare professionals and patients [18]. Thus, selecting an XAI method in a TinyML environment requires simultaneously considering diverse requirements, such as hardware constraints, the characteristics of the explanation method, explanation quality, task objectives, and human interpretability requirements. However, comprehensively evaluating these requirements and selecting an appropriate explanation method is not easy and places a significant burden even on experts. Recent advances in large language models (LLMs) have highlighted their potential to function as interfaces between human intent and system design [19, 20]. In particular, LLMs can support the exploration, generation, and preliminary ranking of candidate solutions based on stakeholder requirements, thereby reducing the manual effort traditionally required to navigate complex design spaces while leaving optimization and feasibility enforcement to deterministic procedures [21]. Therefore, in this study, we formulate the selection of XAI methods in TinyML environments as a human-centered, constrained multi-objective optimization problem. The proposed framework evaluates trade-offs among explanation fidelity, stability, and proxy-based deployment cost using Pareto optimization, while leveraging LLMs to map qualitative stakeholder-oriented requirements, represented in this study through predefined goal profiles, to candidate XAI methods. The LLMs support the preliminary ranking of candidate methods according to qualitative stakeholder requirements, while deployment constraints are enforced separately through deterministic feasibility rules. Additionally, feasibility assessment and Pareto-based candidate identification are performed using deterministic procedures, supporting transparency and auditability. The computational selection stages of the framework are demonstrated through a proof-of-concept skin lesion classification study using the HAM10000 dataset [22] and MobileNetV3-Small [23]. 2 Literature Review This section reviews prior work on (i) XAI deployment in TinyML environments, (i) human-centered explainability, and (i) frameworks for XAI method selection, including recent LLM-assisted approaches. 2.1 Explainability Method Applications in TinyML This review also considers related work from the broader TinyML and edge AI domains because studies specifically addressing XAI in medical TinyML remain limited, while these broader studies address similar challenges in deploying explainability under resource constraints. Existing research has explored the use of XAI to support TinyML system design and optimization, including model pruning, hardware design, robustness enhancement, and anomaly detection [24, 25, 26, 27, 28]. These studies primarily used XAI to support the design and optimization of TinyML systems rather than to provide explanations directly to end users. Consequently, the computational implications of user-facing XAI under on-device deployment have received limited attention. Other studies have integrated XAI into TinyML and edge AI applications, with explanations generated either on servers [29, 30] or directly on edge devices [31, 32, 33]. However, studies in this category often provide limited analysis of the computational costs associated with XAI implementation and hardware constraints. Recent work has begun to consider hardware constraints and the computational cost of explanation generation in edge AI environments [34, 13, 14]. Nevertheless, XAI selection in these studies remains largely developer-driven with limited consideration of stakeholder preferences or end-user evaluation of explanation quality. Furthermore, few studies explicitly incorporate stakeholder-oriented requirements into the design and selection of XAI methods for resource-constrained deployments. 2.2 Human-Centered Explainability Explainability plays a critical role in shaping human interaction with AI systems, particularly in safety-critical domains such as healthcare [4]. Human-centered explainability focuses on how explanations are designed, presented, and interpreted by users within specific decision-making contexts [35] and considers how users perceive and respond to explanations [36]. Empirical studies show that explanation quality can influence user trust and reliance, potentially leading to over-reliance, under-reliance, or inappropriate trust depending on the characteristics of the explanation [5, 6]. However, trust is not synonymous with reliance, and this distinction underscores the necessity of employing appropriate and validated evaluation measures to guide the design and evaluation of human-centered AI systems [37]. These findings highlight that explanations should be evaluated not only in terms of technical correctness but also in terms of their potential effects on human judgment and decision-making. In medical applications, explanations should align with domain knowledge and support clinically meaningful reasoning [7]. However, existing work often treats human factors and system constraints separately, limiting the applicability of these approaches to real-world deployments. 2.3 Selection of Explainability Methods Over the years, several frameworks have been proposed to support the selection of appropriate XAI techniques. Prior approaches include a multi-criteria decision-making framework for ranking candidate methods [38], property-oriented selection frameworks targeting data scientists [39], conceptual frameworks for XAI centered on stakeholders and their explanatory needs [40, 41], and taxonomy-based approaches for organizing and comparing XAI techniques [42]. These studies have contributed to a more structured understanding and selection of XAI methods. That said, they typically rely on manual evaluation, focus primarily on developers or data scientists, or remain at the conceptual level with limited empirical involvement of domain experts. A recent study incorporated LLMs into the design of XAI for small MCUs [21]. While this highlighted the potential of generative AI for automated design support, the resulting recommendations remain largely driven by hardware and model-performance considerations, with limited representation of domain-user requirements in the design process. To address this gap, we propose a human-centered TinyML XAI selection framework that combines stakeholder-oriented goal profiles, LLM-guided candidate generation, deterministic feasibility assessment, Pareto-based method selection, and a proposed human-expert review stage. The present proof-of-concept study evaluates the computational selection stages, while the human-expert review stage remains to be empirically implemented and validated. 2.4 Main Contributions In this manuscript, we propose four main contributions that support the selection of appropriate XAI configurations under task-specific and resource constraints. The framework is intended to enable transparent and resource-aware XAI design for resource-constrained edge AI applications. Its computational components are demonstrated through a proof-of-concept medical imaging case study. 1. Human-Centered, Stakeholder-Oriented Formulation: We formulate XAI method selection as a constrained multi-objective optimization problem that jointly considers qualitative stakeholder priorities, explanation quality, and proxy-based deployment cost. 2. LLM-Guided Candidate Generation: We introduce a mechanism that maps qualitative stakeholder preferences to candidate XAI methods, while feasibility assessment and Pareto-based evaluation remain governed by deterministic procedures. 3. Pareto-Based Selection: We develop a Pareto-based selection framework that captures trade-offs among explanation fidelity, stability, and proxy-based deployment cost. 4. Transparent and Auditable Selection: We provide a transparent and auditable comparison of Pareto-efficient XAI configurations to support subsequent human-expert review and deployment-oriented planning. 3 Proposed Method This work proposes a human-centered, hardware-aware framework for selecting XAI methods for TinyML edge deployment scenarios, as illustrated in Fig. 1. Figure 1: Overview of the proposed human-centered, hardware-profile-aware framework for selecting XAI methods for TinyML edge deployment scenarios. The framework integrates stakeholder-oriented goal profiles, LLM-guided candidate generation, deterministic feasibility enforcement, Pareto-based evaluation, and a proposed human-expert review stage for final method validation. 3.1 Problem Definition Stage The proposed framework begins with an explicit problem-definition stage in which the predictive-model specification and stakeholder intent are recorded independently of XAI method and hardware metadata. This separation is important for safety-sensitive XAI design because it prevents qualitative stakeholder requirements from being conflated with device-specific resource constraints [15, 4]. The model specification describes the architecture family, task definition, input representation, and output format. Stakeholder intent is represented through three predefined goal profiles, Clinical, TinyML, and Balanced, which encode the decision context, intended users, primary and explanation goals, operational mode, time sensitivity, explanation-quality expectation, risk tolerance, explanation representation, and safety considerations. Deterministic rule-based mappings subsequently translate the profile fields into policy constraints used during candidate assessment and feasibility enforcement, following a structured approach suitable for human-centered and safety-critical AI systems [43, 44]. 3.2 Knowledge Source Stage The construction of candidate explanation pipelines is supported by two structured knowledge sources: an XAI method knowledge base and a hardware profile library. 3.2.1 XAI Method Knowledge Base The XAI method knowledge base records the properties required for candidate enrichment and deterministic feasibility assessment. These properties include method family, execution scope, required access to model weights, activations, or gradients, the number of forward passes, and available analytical resource proxies. The resource fields are used for relative comparison and feasibility screening rather than as direct measurements of runtime, energy consumption, or physical memory usage. The knowledge base also distinguishes methods intended for on-device execution from host-only methods and records the requirements associated with each method. 3.2.2 Hardware Profile Library The hardware profile library represents the target execution environment, including supported numerical precision, runtime backend, SRAM and Flash capacity, system-memory reservations, stack and heap reservations, and model-memory reservations. Given an instantiated target profile, the remaining SRAM and Flash headroom available for XAI is calculated deterministically. 3.3 LLM-Guided Design Stage The LLM receives the predictive-model specification, the qualitative stakeholder goal profile, and a structured summary of the complete XAI method catalog. The catalog summary includes method family, execution scope, and method requirements, but does not provide measured latency, energy, SRAM, or Flash values. The LLM is instructed to return a ranked shortlist containing no more than five methods and to align the ranking with the qualitative stakeholder profile. The prompt prevents the LLM from generating resource estimates or enforcing hardware feasibility. Method metadata and deployment constraints are attached and evaluated deterministically in the subsequent stage. A separate Pareto-constrained LLM call is used during final configuration selection, as described in Section 3.6. 3.4 Feasibility Enforcement Stage Let M=m1,…,mNM=\m_1,…,m_N\ denote the ranked methods proposed by the LLM. Each candidate is first enriched using the method knowledge base, which supplies its execution scope, model-access requirements, forward-pass requirement, and available resource proxies. Deterministic filtering is then performed in two stages. The first stage checks SRAM and Flash limits whenever both the candidate estimate and the corresponding hardware limit are defined. The second stage applies execution-scope and stakeholder-derived policy constraints, including forward-pass limits and the permitted use of multi-pass methods. A host-only method is never classified as MCU-feasible. When host fallback is permitted, a host-only method may be retained as a hybrid alternative, but it is not included in the MCU-feasible set. When host fallback is not permitted, the method is rejected from the intended deployment. Let MCUC_MCU denote the set of applicable MCU constraints. The MCU-feasibility indicator is defined in Eq. 1. MCU(mi)=1,if mi satisfies every applicable c∈MCU,0,otherwise.I_MCU(m_i)= cases1,&if m_i satisfies every applicable c _MCU,\\ 0,&otherwise. cases (1) The MCU-feasible method set is therefore defined in Eq. 2. MMCU=mi∈M∣MCU(mi)=1.M_MCU= \m_i∈ M _MCU(m_i)=1 \. (2) Fidelity and stability are evaluated across the complete parameterized configuration set. For each stakeholder profile, however, only configurations whose base method belongs to MMCUM_MCU and for which all required objective values are available are included in the profile-specific deployment-cost and Pareto analyses. 3.5 XAI Validation Stage The validation stage evaluates explanation fidelity, explanation stability, and relative deployment cost. These quantities are subsequently used to identify non-dominated trade-offs among the feasible configurations. 3.5.1 Measure Explanation Quality Explanation quality is evaluated from two complementary perspectives: attribution fidelity and explanation stability. Fidelity measures the extent to which highly attributed image regions affect the model score. Stability measures the structural consistency of the resulting heatmaps under mild input variations [45, 46, 47]. Fidelity Score: Because the case study addresses skin-lesion image classification using heatmap-based explanations, attribution fidelity is evaluated using deletion AUC, insertion AUC, and Area Over the Perturbation Curve (AOPC) [12, 48]. For each input image x∈ℝH×W×3x ^H× W× 3, the evaluation class is fixed to the class predicted from the unmodified input, as defined in Eq. 3: c⋆(x)=argmaxkzk(x),c (x)= _kz_k(x), (3) where zk(x)z_k(x) denotes the logit of class k. The fixed-class score used throughout the perturbation process is defined in Eq. 4. Sx(x′)=zc⋆(x)(x′),S_x(x )=z_c (x)(x ), (4) where x′x denotes the original image or one of its perturbed versions. The implementation uses logits rather than softmax probabilities. Let hm(x)h_m(x) denote the heatmap produced by configuration m. The heatmap is resized to the input resolution. Its absolute values are then independently min–max normalized for each input, as shown in Eq. 5: h¯m(x)=|hm(x)|−min|hm(x)|max|hm(x)|−min|hm(x)|+ϵ. h_m(x)= |h_m(x) |- |h_m(x) | |h_m(x) |- |h_m(x) |+ε. (5) The normalized heatmap is partitioned into non-overlapping patches of size p×p× p. In the present implementation, p=16p=16. For the 224×224224× 224 inputs, the number of patches is calculated using Eq. 6: P=22416×22416=196P= 22416× 22416=196 (6) patches. The patches are ranked in descending order according to their mean absolute attribution magnitude. The evaluation uses L=20L=20 perturbation intervals rather than applying one perturbation step for every patch. At interval ℓ∈0,…,L ∈\0,…,L\, the cumulative number and fraction of modified patches are defined in Eq. 7. qℓ=round(ℓPL),uℓ=qℓP.q_ =round ( PL ), u_ = q_ P. (7) Let πqℓ(x,hm) _q_ (x,h_m) denote the set containing the qℓq_ highest-ranked patches. A blurred version of the input is used as the baseline image xbasex_base. The implementation constructs this baseline using a 21×2121× 21 uniform depthwise blur. During deletion, the highest-ranked patches are progressively replaced by the corresponding regions of the baseline image, as shown in Eq. 8: xdel(ℓ)=Replace(x,xbase,πqℓ(x,hm)).x^( )_del=Replace (x,x_base, _q_ (x,h_m) ). (8) Insertion begins with the blurred baseline and progressively restores the highest-ranked regions from the original image, as shown in Eq. 9: xins(ℓ)=Replace(xbase,x,πqℓ(x,hm)).x^( )_ins=Replace (x_base,x, _q_ (x,h_m) ). (9) The deletion and insertion curves are defined in Eq. 10. Dm(uℓ)=Sx(xdel(ℓ)),Im(uℓ)=Sx(xins(ℓ)).D_m(u_ )=S_x (x^( )_del ), I_m(u_ )=S_x (x^( )_ins ). (10) The areas under these curves are numerically calculated using the trapezoidal rule, as shown in Eqs. 11 and 12: AUCdel(m)=∑ℓ=1LDm(uℓ−1)+Dm(uℓ)2(uℓ−uℓ−1),AUC_del(m)= _ =1^L D_m(u_ -1)+D_m(u_ )2 (u_ -u_ -1 ), (11) AUCins(m)=∑ℓ=1LIm(uℓ−1)+Im(uℓ)2(uℓ−uℓ−1).AUC_ins(m)= _ =1^L I_m(u_ -1)+I_m(u_ )2 (u_ -u_ -1 ). (12) Lower deletion AUC and higher insertion AUC indicate better attribution fidelity because removing highly attributed regions should rapidly reduce the fixed-class score, whereas inserting them should rapidly restore it [12, 49]. AOPC measures the average reduction in the original fixed-class logit across the L non-zero deletion intervals, as defined in Eq. 13: AOPC(m)=1L∑ℓ=1L[Sx(x)−Dm(uℓ)].AOPC(m)= 1L _ =1^L [S_x(x)-D_m(u_ ) ]. (13) Higher AOPC indicates that removing highly ranked regions causes a larger reduction in the score of the original predicted class [48]. The image-level deletion AUC, insertion AUC, and AOPC values are averaged over the evaluation dataset for every parameterized configuration. Since the three metrics have different numerical ranges and optimization directions, their configuration-level means are independently min–max normalized across the evaluated configuration set. The normalized quantities are denoted in Eq. 14. AUC~del(m),AUC~ins(m),AOPC~(m). AUC_del(m), AUC_ins(m), AOPC(m). (14) Because deletion AUC is minimized, its normalized value is inverted. The composite fidelity score is then calculated using equal weights, as shown in Eq. 15: F(m)=13[AUC~ins(m)+AOPC~(m)+1−AUC~del(m)].F(m)= 13 [ AUC_ins(m)+ AOPC(m)+1- AUC_del(m) ]. (15) Higher values of F(m)F(m) indicate better composite attribution fidelity across the deletion, insertion, and AOPC criteria. Stability: Explanation stability is evaluated using three mild perturbed versions of each input image. Each perturbation applies reflection padding of six pixels followed by a random crop back to the original input dimensions. This introduces a small spatial translation. Zero-mean Gaussian noise with a standard deviation of 0.020.02 is then added in the model-input value scale. The implementation does not explicitly verify preservation of the predicted class after perturbation. These transformations are therefore treated as mild input variations rather than guaranteed label-preserving perturbations. Brightness and contrast transformations are not used in the stability calculation. For each original and perturbed input, the explanation heatmap is resized to the input resolution and independently min–max normalized to [0,1][0,1]. Let hm(x)h_m(x) denote the resulting heatmap for the original input and let hm(x~j)h_m( x_j) denote the heatmap for the jjth perturbed input. Structural similarity is measured using the Structural Similarity Index Measure (SSIM) [50, 51]. For two heatmaps a and b, SSIM is defined in Eq. 16. SSIM(a,b)=(2μaμb+C1)(2σab+C2)(μa2+μb2+C1)(σa2+σb2+C2),SSIM(a,b)= (2 _a _b+C_1)(2 _ab+C_2)( _a^2+ _b^2+C_1)( _a^2+ _b^2+C_2), (16) where μa _a and μb _b denote local means, σa2 _a^2 and σb2 _b^2 denote local variances, σab _ab denotes local covariance, and C1C_1 and C2C_2 are stabilizing constants. In the present implementation, J=3J=3 perturbed inputs are generated for each image. The image-level stability score is defined in Eq. 17. Stabm(x)=1J∑j=1JSSIM(hm(x),hm(x~j)).Stab_m(x)= 1J _j=1^JSSIM (h_m(x),h_m( x_j) ). (17) The configuration-level stability objective is obtained by averaging over the complete evaluation dataset D, as shown in Eq. 18: S(m)=1||∑x∈Stabm(x).S(m)= 1|D| _x Stab_m(x). (18) Higher values of S(m)S(m) indicate greater structural consistency of the explanations under the applied input variations. Stability is considered jointly with fidelity because strongly compressed or nearly invariant but uninformative heatmaps may obtain high SSIM values. 3.5.2 Identification of Trade-Offs Deployment overhead is represented using runtime and SRAM proxies rather than direct measurements of physical energy consumption. For the CAM-family configurations included in the profile-specific Pareto analysis, the runtime proxy is defined in Eq. 19. T(m)=t¯shared_fwd+t¯CAM+t¯post(m),T(m)= t_shared\_fwd+ t_CAM+ t_post(m), (19) where t¯shared_fwd t_shared\_fwd is the mean time of the shared model forward pass, t¯CAM t_CAM is the mean time required to construct the base CAM from the shared activations and logits, and t¯post(m) t_post(m) is the mean configuration-specific post-processing time. For the unmodified CAM configuration, t¯post(m)=0 t_post(m)=0. An SRAM proxy, denoted by Mtensor(m)M_tensor(m), is obtained from the analytical resource metadata in the XAI method knowledge base. The proxy represents the estimated tensor storage associated with the base method. In the current implementation, parameterized configurations belonging to the same base method may share the same SRAM proxy. It should therefore not be interpreted as a parameter-specific measured memory footprint. Within each stakeholder profile, the runtime proxy is min–max normalized over configurations with available timing values, whereas the SRAM proxy is min–max normalized over configurations with available analytical SRAM metadata in the merged profile table. Pareto filtering is subsequently restricted to configurations with complete fidelity, stability, runtime, and SRAM-derived deployment-cost values. The relative deployment-cost proxy is defined in Eq. 20. Cg(m)=wTmmg(T(m))+wMmmg(Mtensor(m)).C_g(m)=w_T\,m_g (T(m) )+w_M\,m_g (M_tensor(m) ). (20) The weights are fixed as shown in Eq. 21. wT=0.85,wM=0.15,w_T=0.85, w_M=0.15, (21) to prioritize the runtime component while retaining a contribution from the SRAM proxy. Lower values of Cg(m)C_g(m) indicate lower estimated deployment cost. This quantity is a relative proxy and is not a measurement of physical energy consumption on an MCU. The three optimization objectives are therefore to maximize composite fidelity F(m)F(m), maximize stability S(m)S(m), and minimize deployment cost Cg(m)C_g(m). For two feasible configurations mim_i and mjm_j in profile g, configuration mim_i dominates configuration mjm_j according to Eq. 22. F(mi)≥F(mj),S(mi)≥S(mj),Cg(mi)≤Cg(mj),F(m_i)≥ F(m_j), S(m_i)≥ S(m_j), C_g(m_i)≤ C_g(m_j), (22) with at least one strict inequality. The profile-specific Pareto set is defined in Eq. 23. g=m∈g∣∄m′∈g such that m′ dominates m,P_g= \m _g \,m _g such that m dominates m \, (23) where gV_g denotes the valid feasible configuration set for profile g. Every configuration in gP_g represents a non-dominated trade-off among fidelity, stability, and relative deployment cost [52]. 3.6 Pareto-Constrained Final Selection The final decision stage operates exclusively on the profile-specific three-objective Pareto set gP_g. Configurations excluded by feasibility enforcement or dominated during Pareto filtering are not permitted to re-enter the selection process. Within each Pareto set, fidelity, stability, and deployment cost are independently min–max normalized. Let F^g(m) F_g(m), S^g(m) S_g(m), and C^g(m) C_g(m) denote the normalized values for configuration m∈gm _g. The profile-specific goal score is defined as Rg(m) R_g(m) =αgF^g(m)+βgS^g(m) = _g F_g(m)+ _g S_g(m) (24) +γg(1−C^g(m)),m∈g. + _g (1- C_g(m) ), m _g. The profile-specific weights are (αg,βg,γg)=(0.45, 0.40, 0.15),g=Clinical,(0.20, 0.20, 0.60),g=TinyML,(0.34, 0.33, 0.33),g=Balanced.( _g, _g, _g)= cases(0.45,\,0.40,\,0.15),&g=Clinical,\\ (0.20,\,0.20,\,0.60),&g=TinyML,\\ (0.34,\,0.33,\,0.33),&g=Balanced. cases (25) The Clinical profile places the greatest emphasis on fidelity and stability, the TinyML profile prioritizes deployment efficiency, and the Balanced profile assigns approximately equal importance to the three objectives. The ranked Pareto configurations, together with their fidelity, stability, deployment-cost, and goal-score values, are supplied to GPT-4.1 mini with a temperature of 0.10.1. The API request uses the gpt-4.1-mini model identifier, with the executed responses recorded as gpt-4.1-mini-2025-04-14. The structured-output schema restricts every method field to exact configuration identifiers from gP_g. For the Clinical profile, the LLM must select exactly one primary configuration and no fallback. The selected configuration must have the highest fidelity in gP_g; fidelity ties are resolved by lower deployment cost and then higher stability. For the Balanced and TinyML profiles, the LLM must select two distinct primary configurations and one distinct fallback. One primary configuration must be the maximum-fidelity, lowest-cost anchor. The second, complementary primary configuration must satisfy F(mcomp) F(m_comp) ≥0.70, ≥ 70, (26) S(mcomp) S(m_comp) ≥S(manchor)+0.02, ≥ S(m_anchor)+02, C(mcomp) C(m_comp) ≤C(manchor)+0.05. ≤ C(m_anchor)+05. The fallback must differ from both primary configurations, provide higher fidelity than the complementary primary configuration, and remain within the same low-cost interval: F(mfallback) F(m_fallback) >F(mcomp), >F(m_comp), (27) C(mfallback) C(m_fallback) ≤C(manchor)+0.05. ≤ C(m_anchor)+05. The structured LLM response is subjected to deterministic validation. The validator checks the stakeholder profile, exact Pareto membership, configuration cardinality, duplication, anchor selection, and all numerical constraints in Eqs. 26 and 27. A response that fails validation is returned to the LLM together with the detected errors for correction. A response that fails validation is returned to the LLM together with the detected errors for correction. If a valid response is not obtained, the procedure terminates without reporting a final selection. 3.7 Proposed Human Review and Method Finalization The framework includes a proposed human-expert review stage to ensure that final explanation selection is not based solely on quantitative optimization. In a future clinical deployment, the Pareto-valid explanations would be reviewed by medical experts using clinically relevant criteria, including consistency with established dermatological patterns, localization of salient regions to medically relevant areas, and the absence of spurious artifacts or background bias. The review stage is intended to identify explanations that may achieve strong quantitative scores while remaining clinically misleading or difficult to interpret. Human-centred evaluation can influence trust, usability, and decision quality in AI-assisted systems [5, 16, 44, 53]. This human-expert review stage was not empirically implemented in the present proof-of-concept study and remains part of future validation. 4 Experimental Setup The experimental study focuses on skin-lesion classification using three stakeholder profiles: (1) Clinical, which prioritizes explanation fidelity and stability; (2) TinyML, which prioritizes on-device deployability; and (3) Balanced, which represents a compromise among explanation quality and deployment cost. Because the predictive task is image classification, the required explanation representation is a class-specific spatial heatmap. The LLM proposal stage considers the complete structured XAI method catalog. A separate parameterized configuration space is then constructed for the quantitative fidelity and stability evaluation. For the profile-specific timing and Pareto analyses, this configuration space is filtered according to the base methods retained by the deterministic feasibility stage. 4.1 Data, Model and Training Methods The proposed framework is evaluated on the HAM10000 skin-lesion dataset [22, 54]. The dataset contains 10,015 dermoscopic images belonging to seven diagnostic categories. MobileNetV3-Small [23] is used as the predictive backbone because it is designed for computationally constrained environments. The images are divided using stratified sampling with a fixed random seed of 42. The resulting subsets contain 7,010 training images, 1,001 validation images, and 2,004 test images, corresponding approximately to 70%, 10%, and 20% of the dataset, respectively. Table 1: Summary of the experimental setup. Component Description Dataset HAM10000; 10,015 images and seven classes Split Stratified training/validation/test split: 7,010/1,001/2,004 images (70%/10%/20%70\%/10\%/20\%) Random seed 42 Input 224×224224× 224 RGB images; MobileNetV3 preprocessing Batch size 64 for model training and evaluation Backbone MobileNetV3-Small with ImageNet-pretrained weights Training stage 1 Frozen backbone; two epochs; Adam with learning rate 10−310^-3 Training stage 2 Last approximately 40% of backbone layers unfrozen; two epochs; Adam with learning rate 3×10−53× 10^-5 Loss and imbalance handling Categorical cross-entropy with label smoothing of 0.05 and balanced class weights Augmentation Horizontal flip; rotation factor 0.10; zoom factor 0.20; contrast factor 0.20 Evaluated XAI families CAM, Tiny-Saliency, LR-CAM, Micro-CAM, Binary-CAM, TopK-CAM, and TopK×Binary Full XAI configuration set 67 parameterized configurations Execution model One shared model forward pass followed by configuration-specific post-processing LLMs GPT-4.1 mini and Gemini 2.0 Flash; Gemini evaluated in a separate run GPT configuration OpenAI API; temperature 0.1 Hardware profile Generic Cortex-M7 profile with 512 kB SRAM and 2 MB Flash Explanation metrics Deletion AUC, insertion AUC, AOPC, and SSIM Optimization objectives Composite fidelity, SSIM-based stability, and relative deployment-cost proxy All images are resized to 224×224224× 224 pixels and processed using the MobileNetV3 preprocessing function. The training set is augmented using horizontal flipping, a rotation factor of 0.10, a zoom factor of 0.20, and a contrast factor of 0.20. Augmentation is not applied to the validation and test sets. A two-stage transfer-learning procedure is used. During the first stage, the ImageNet-pretrained MobileNetV3-Small backbone is frozen and the classification head is trained for two epochs using Adam with a learning rate of 10−310^-3. During the second stage, approximately the last 40% of the backbone layers are unfrozen and the model is fine-tuned for two additional epochs using a learning rate of 3×10−53× 10^-5. Both stages use class weighting and categorical cross-entropy with label smoothing of 0.05. A ReduceLROnPlateau callback monitors validation loss. The principal experimental settings are summarized in Table 1. 4.2 Explainability Methods All 67 explanation configurations are implemented using a shared-forward execution model. A single model inference returns the final convolutional activation tensor and the class logits. The base CAM is calculated from these shared outputs, after which the parameterized explanation variants are generated using configuration-specific post-processing. Tiny-Saliency is also generated from the shared activation tensor and does not require an additional model inference. The complete set of 67 configurations is evaluated for fidelity and stability under each stakeholder profile. Before the profile-specific deployment-cost and Pareto analyses, the configuration set is filtered according to the base methods retained by the deterministic feasibility stage. The complete parameterized design space is reported in Table 2. Table 2: Parameterized XAI configuration space. The symbols d, s, τ, and k denote the LR-CAM downsampling factor, requested Micro-CAM spatial size, Binary-CAM threshold, and retained TopK-CAM activation ratio, respectively. Requested Micro-CAM sizes greater than the 7×77× 7 feature-map resolution are clipped to seven by the implementation. XAI family Parameter Values No. CAM – Baseline 1 Tiny-Saliency – Baseline 1 LR-CAM Downsampling d 1,…,101,…,10 10 Micro-CAM Requested size s 1,…,101,…,10 10 Binary-CAM Threshold τ 0.1,…,0.90.1,…,0.9 9 TopK-CAM Retained ratio k 0.02,0.04,…,0.500.02,0.04,…,0.50 25 TopK×Binary k and τ 11 parameter pairs 11 Total – – 67 For TopK×Binary, the parameter grid is constructed from k∈0.05,0.15,0.30,0.50,τ∈0.3,0.5,0.7.k∈\0.05,0.15,0.30,0.50\, τ∈\0.3,0.5,0.7\. (28) Eleven of the twelve Cartesian-product combinations are evaluated; the combination (k,τ)=(0.50,0.7)(k,τ)=(0.50,0.7) is not included. For every included combination, TopK-CAM is first applied to the base CAM, followed by the binary-threshold operation. LR-CAM reduces the 7×77× 7 CAM resolution according to Hd=max(1,⌊7d⌋),Wd=max(1,⌊7d⌋).H_d= (1, 7d ), W_d= (1, 7d ). (29) Micro-CAM resizes the base CAM to seff=min(s,7),s_eff= (s,7), (30) which means that requested sizes s=7,…,10s=7,…,10 produce the same effective 7×77× 7 spatial resolution. Binary-CAM thresholds the normalized base CAM at τ. TopK-CAM retains at least one activation and otherwise retains max(1,⌊49k⌋) (1, 49k ) (31) of the 49 CAM locations. 4.3 LLM-Guided Models The explanation-proposal stage uses OpenAI’s GPT-4.1 mini and Gemini 2.0 Flash [55]. GPT-4.1 mini is accessed through the OpenAI API using a temperature of 0.1. Gemini 2.0 Flash is evaluated in a separate run using the same prompt structure. For each stakeholder profile, the prompt contains the predictive-model specification, the qualitative stakeholder goal profile, and a structured summary of the XAI method catalog. The method summary includes method family, execution scope, and method requirements such as the number of forward passes. Explicit SRAM, Flash, latency, and energy values are not provided to the LLM. The prompt instructs the LLM to propose and rank no more than five methods without performing hardware-feasibility filtering or generating resource estimates. 4.4 Hardware and Deployment Constraints The proof-of-concept uses an instantiated generic Cortex-M7 hardware profile with 512 kB of total SRAM and 2,048 kB of total Flash. The profile reserves 64 kB of SRAM for the system, 64 kB for the stack and heap, and predefined model footprints of 200 kB SRAM and 1,500 kB Flash. Under these predefined values, the resulting headroom available for XAI is 512−64−64−200=184kB512-64-64-200=184\,kB (32) of SRAM and 2048−1500=548kB2048-1500=548\,kB (33) of Flash. These footprint values are used for deterministic proof-of-concept screening and are not reported as measurements obtained from a deployed physical MCU. Runtime measurements are obtained from a batch of 16 test images. Two untimed warm-up executions are performed, followed by five repeated measurements. The runtime proxy for each CAM-family configuration contains the shared model forward-pass time, the base-CAM construction time, and the configuration-specific post-processing time. Within each stakeholder profile, runtime is min–max normalized over configurations with available timing measurements, whereas the SRAM proxy is normalized over configurations with available analytical SRAM metadata. The relative deployment-cost proxy assigns weights of wT=0.85w_T=0.85 and wM=0.15w_M=0.15 to runtime and SRAM, respectively. Configurations with incomplete objective records are excluded before Pareto filtering. The resulting quantity is a relative proxy rather than physical energy consumption or MCU-measured latency. 5 Results The predictive performance of the classifier was first evaluated on the held-out test set. On the complete test set of 2,004 images, the MobileNetV3-Small model achieved an accuracy of 66.77% with a test loss of 1.0641. 5.1 LLM Proposals Across Stakeholder Profiles The top-five XAI methods proposed by GPT-4.1 mini and Gemini 2.0 Flash were compared across the Clinical, TinyML, and Balanced stakeholder profiles. Table 3 reports the ranked proposals and their execution scope. Table 3: Top-five XAI methods proposed by GPT-4.1 mini and Gemini 2.0 Flash across the stakeholder profiles. Methods are listed in rank order; † denotes host-only execution. Profile Model Ranked proposals Clinical GPT-4.1 mini 1) LR-CAM; 2) TopK-CAM; 3) Binary-CAM; 4) CAM; 5) GradCAM† Gemini 2.0 Flash 1) LR-CAM; 2) CAM; 3) GradCAM†; 4) GradCAM++†; 5) Score-CAM† TinyML GPT-4.1 mini 1) CAM; 2) LR-CAM; 3) Binary-CAM; 4) TopK-CAM; 5) Micro-CAM Gemini 2.0 Flash 1) CAM; 2) LR-CAM; 3) Binary-CAM; 4) TopK-CAM; 5) Micro-CAM Balanced GPT-4.1 mini 1) LR-CAM; 2) TopK-CAM; 3) Binary-CAM; 4) CAM; 5) GradCAM++† Gemini 2.0 Flash 1) LR-CAM; 2) CAM; 3) TopK-CAM; 4) Micro-CAM; 5) GradCAM† • TinyML: GPT-4.1 mini and Gemini 2.0 Flash proposed the same five methods in the same rank order: CAM, LR-CAM, Binary-CAM, TopK-CAM, and Micro-CAM. All five methods were classified in the method knowledge base as single-pass, on-device methods and passed the deterministic MCU-feasibility stage. • Clinical: Both models ranked LR-CAM first. GPT-4.1 mini proposed four on-device methods and one host-only method, whereas Gemini 2.0 Flash proposed two on-device methods and three host-only methods. After deterministic MCU-feasibility filtering, LR-CAM, TopK-CAM, Binary-CAM, and CAM remained from the GPT-4.1 mini proposals, while LR-CAM and CAM remained from the Gemini proposals. • Balanced: Both models proposed LR-CAM, CAM, and TopK-CAM. GPT-4.1 mini also proposed Binary-CAM and the host-only GradCAM++, whereas Gemini 2.0 Flash also proposed Micro-CAM and the host-only GradCAM. The principal difference between the on-device proposal sets was therefore Binary-CAM in the GPT-4.1 mini shortlist and Micro-CAM in the Gemini shortlist. Four methods from each model passed deterministic MCU-feasibility filtering. CAM-family methods were the most frequently proposed methods across the three profiles in the observed runs. The two models produced identical rankings for the TinyML profile but different method sets and rankings for the Clinical and Balanced profiles. Representative qualitative outputs for the five methods jointly proposed by both models under the TinyML profile are shown in Fig. 2. These examples illustrate the visual effects of the different CAM transformations; they are not presented as a clinical validation of the resulting explanations. Figure 2: Representative outputs for the five XAI methods jointly proposed by GPT-4.1 mini and Gemini 2.0 Flash under the TinyML profile. The leftmost image is the original input, followed by CAM, LR-CAM, Micro-CAM, Binary-CAM, and TopK-CAM explanations. 5.2 Runtime and Method-Level Overhead Analysis Runtime was measured in the experimental execution environment using a batch of 16 test images. Two untimed warm-up iterations were performed before each measurement. Each timing operation was then repeated five times, and the mean runtime was recorded. All evaluated configurations used the shared-forward execution model: a single model forward pass and base CAM computation were shared with the configuration-specific post-processing operation. Table 4 reports the mean total shared-forward runtime for representative configurations and the corresponding increase relative to CAM within the same stakeholder profile. Table 4: Mean shared-forward runtime for representative configurations. Measurements were obtained for a batch of 16 images after two untimed warm-up iterations and were repeated five times. Incremental overhead was calculated relative to CAM within the same profile. All values are reported in milliseconds. Goal Method Total runtime Increase vs. CAM Clinical CAM 45.38848 0.00000 Binary-CAM (τ=0.80τ=0.80) 45.39458 0.00610 Binary-CAM (τ=0.90τ=0.90) 45.39354 0.00506 TinyML CAM 43.20912 0.00000 Binary-CAM (τ=0.60τ=0.60) 43.22048 0.01136 Binary-CAM (τ=0.50τ=0.50) 43.22030 0.01118 Micro-CAM (s=5s=5) 43.41238 0.20326 Balanced CAM 43.96584 0.00000 Binary-CAM (τ=0.60τ=0.60) 43.97222 0.00638 Binary-CAM (τ=0.50τ=0.50) 43.97226 0.00642 TopK-CAM (k=0.10k=0.10) 44.09098 0.12514 The CAM baseline runtime differed slightly across the three profile-specific timing runs. Comparisons were therefore made only relative to the CAM measurement obtained within the same run. Among the representative configurations, the largest measured increase was 0.20326 ms for Micro-CAM with s=5s=5 under the TinyML profile, followed by 0.12514 ms for TopK-CAM with k=0.10k=0.10 under the Balanced profile. The reported Binary-CAM configurations added between 0.00506 and 0.01136 ms relative to their corresponding CAM baselines. These batch-level measurements show that the shared model forward pass and base CAM computation accounted for most of the total measured runtime in the experimental environment. They should not be interpreted as direct latency measurements on the target Cortex-M7 hardware. Figure 3 complements the representative total-runtime measurements by showing the mean incremental post-processing overhead of each explanation family relative to CAM, averaged across its evaluated parameter settings. Figure 3: Mean incremental post-processing overhead of each explanation family relative to the CAM baseline, averaged across the evaluated parameter settings. The reported values exclude neither the shared forward pass nor the base CAM computation from the total-runtime measurements; the plotted increment represents the configuration-specific increase relative to CAM. 5.3 Fidelity and Stability Evaluation The evaluation included 67 unique configurations under each of the three deployment goals and used the complete held-out test set of 2,004 images. For every configuration, the logit-based deletion AUC, insertion AUC, AOPC, and composite fidelity values were identical across A2_BALANCED, A2_CLINICAL, and A2_TINYML. This occurred because the same trained model, test set, explanation configuration, and fidelity-evaluation procedure were used for all three profiles. SSIM-based stability was evaluated in separate profile-specific runs. Table 5 therefore reports the mean SSIM across the three profiles together with the corresponding minimum and maximum values. The highest composite fidelity score of 0.9397 was achieved by CAM, LR-CAM with d=1d=1, and Micro-CAM with sizes 7–10. Table 5 reports s=10s=10 as a representative configuration. These configurations produced a deletion AUC of 0.9745, an insertion AUC of 1.9960, and an AOPC of 1.3302. However, the individual fidelity metrics did not identify a single dominant configuration. Micro-CAM with s=6s=6 achieved the highest insertion AUC of 2.0021, whereas TopK-CAM with k=0.46k=0.46 achieved both the lowest deletion AUC of 0.9342 and the highest AOPC of 1.3703. The latter nevertheless produced a lower composite fidelity score and substantially lower SSIM-based stability. These findings show that the fidelity metrics capture complementary aspects of attribution behaviour and should therefore be interpreted jointly. Table 5: Representative fidelity and stability results across the Clinical, TinyML, and Balanced profiles. Fidelity metrics were identical across profiles, while SSIM is reported as the mean and range across the three profile-specific runs. Higher values are preferable except for deletion AUC. Bold values indicate the best observed result for each metric. Configuration Del. ↓ Ins. ↑ AOPC ↑ Fid. ↑ SSIM mean [range] ↑ CAM 0.9745 1.9960 1.3302 0.9397 0.8205 [0.8196, 0.8216] LR-CAM (d=1d=1) 0.9745 1.9960 1.3302 0.9397 0.8193 [0.8166, 0.8211] Micro-CAM (s=10s=10) 0.9745 1.9960 1.3302 0.9397 0.8201 [0.8174, 0.8236] Micro-CAM (s=6s=6) 1.1000 2.0021 1.2046 0.7665 0.8663 [0.8640, 0.8698] TopK-CAM (k=0.46k=0.46) 0.9342 1.8683 1.3703 0.9209 0.6566 [0.6533, 0.6614] Binary-CAM (τ=0.50τ=0.50) 1.0201 1.8406 1.2850 0.7840 0.8252 [0.8200, 0.8296] Micro-CAM (s=4s=4) 1.2111 1.9926 1.0934 0.6042 0.8780 [0.8751, 0.8797] Binary-CAM (τ=0.90τ=0.90) 1.1506 1.6877 1.1549 0.5100 0.9271 [0.9252, 0.9291] TopK-CAM (k=0.04k=0.04) 1.1803 1.6527 1.1255 0.4477 0.9434 [0.9428, 0.9439] Micro-CAM (s=1s=1) 1.3724 1.4381 0.9333 0.0502 1.0000 [1.0000, 1.0000] The tested configurations also show that stronger spatial compression or sparsification can increase SSIM-based stability while reducing fidelity. Reducing the Micro-CAM resolution from s=10s=10 to s=1s=1 decreased composite fidelity from 0.9397 to 0.0502 while increasing mean SSIM from 0.8201 to 1.0000. Similarly, reducing the TopK-CAM retained ratio from k=0.46k=0.46 to k=0.04k=0.04 decreased fidelity from 0.9209 to 0.4477 while increasing mean SSIM from 0.6566 to 0.9434. Increasing the Binary-CAM threshold from τ=0.50τ=0.50 to τ=0.90τ=0.90 produced the same pattern, reducing fidelity from 0.7840 to 0.5100 while increasing mean SSIM from 0.8252 to 0.9271. A high SSIM value should therefore not be interpreted independently as evidence of a high-quality explanation. Strongly compressed or nearly invariant attribution maps may appear highly stable while retaining little informative spatial variation. Fidelity and stability must be considered jointly to distinguish informative and stable explanations from stable but low-fidelity attribution maps. 5.4 Pareto-Optimal Trade-Off Analysis The joint trade-off among fidelity, stability, and deployment cost was examined using three-objective Pareto dominance. Composite fidelity and SSIM-based stability were maximized, whereas the deployment-cost proxy was minimized. Although fidelity and stability were available for all 67 configurations under each profile, the Pareto analysis required complete fidelity, stability, and deployment-cost records. After this completeness requirement was applied, 55 configurations were eligible under TinyML and 45 configurations were eligible under each of the Clinical and Balanced profiles. Three-objective Pareto filtering retained 16 of the 55 eligible configurations for TinyML, 12 of the 45 configurations for Clinical, and 13 of the 45 configurations for Balanced. In addition to this three-objective analysis, Fig. 4 presents a two-dimensional visualization of deployment cost against stability. The horizontal axis represents the deployment-cost proxy, the vertical axis represents SSIM-based stability, and point colour encodes composite fidelity. The highlighted cross markers are determined using all three objectives rather than only the two displayed axes. CAM appeared in the Pareto-optimal set under all three profiles at the lowest deployment-cost value of 0.150. It combined the highest composite fidelity score of 0.9397 with an SSIM-based stability value of approximately 0.82, making it a low-cost, high-fidelity reference configuration. Binary-CAM variants occupied the low-deployment-cost region of each Pareto-optimal set, with cost values of approximately 0.166–0.171. Across the Pareto-optimal Binary-CAM configurations, increasing the binary threshold generally increased stability from approximately 0.82 to 0.93 while reducing fidelity from 0.7840 to 0.5100. These configurations therefore represented low-cost alternatives with different fidelity–stability balances. TopK-CAM variants occupied an intermediate deployment-cost region. Across the three profiles, the Pareto-optimal TopK-CAM configurations had deployment-cost values ranging from approximately 0.46 to 0.62 and stability values between approximately 0.84 and 0.94. Lower retained ratios generally produced higher stability but lower fidelity. For example, TopK-CAM with k=0.04k=0.04 achieved a stability value close to 0.94 but a fidelity score of 0.4477, whereas TopK-CAM with k=0.12k=0.12 retained a higher fidelity score of 0.7578 with lower stability. This pattern illustrates the trade-off introduced by increasingly sparse activation selection. (a) TinyML (b) Clinical (c) Balanced Figure 4: Two-dimensional visualizations of the three-objective Pareto analysis for the (a) TinyML, (b) Clinical, and (c) Balanced profiles. The horizontal and vertical axes represent the deployment-cost proxy and SSIM-based stability, respectively, while point colour represents composite fidelity. Cross markers identify configurations that are non-dominated when deployment cost, fidelity, and stability are considered jointly. The dashed lines connect these three-objective Pareto configurations in ascending deployment-cost order and are included only as visual guides; they do not represent a separately calculated two-objective Pareto frontier. The TinyML Pareto-optimal set additionally included Micro-CAM configurations with sizes 4–7. Their deployment-cost values ranged from approximately 0.33 to 0.75 and represented different compromises among spatial compression, fidelity, and stability. Micro-CAM configurations had complete timing and deployment-cost records only under the TinyML profile and were therefore eligible for the Pareto analysis only in that profile. Their absence from the Clinical and Balanced Pareto-optimal sets should not be interpreted as a dominance-based rejection under those profiles. Several LR-CAM configurations produced the maximum SSIM value of 1.000. Among these configurations, the three-objective Pareto procedure retained d=4d=4 for TinyML, d=8d=8 for Clinical, and d=7d=7 for Balanced. Each of these configurations had a composite fidelity score of only 0.0502. Their presence in the Pareto-optimal sets should therefore not be interpreted as evidence of superior explanation quality. Instead, they illustrate that an almost invariant attribution map can maximize SSIM-based stability while retaining little informative spatial variation. This finding reinforces the need to consider fidelity, stability, and deployment cost jointly. The Pareto-optimal sets shared a common low-cost structure based on CAM, Binary-CAM, and TopK-CAM, but they were not identical. Micro-CAM appeared only in the TinyML analysis because complete deployment-cost records were available only for that profile, and the retained LR-CAM configuration differed across the three profiles. No single configuration simultaneously optimized all three objectives. The resulting non-dominated sets instead provide multiple candidates for the subsequent stakeholder-specific decision stage. 5.5 Final Pareto-Constrained Configuration Selection Following the three-objective Pareto analysis, the non-dominated configurations for each stakeholder profile were ranked using the profile-specific goal score and supplied to GPT-4.1 mini for final selection. The LLM was restricted to selecting exact configuration identifiers from the corresponding Pareto-optimal set. It was permitted to select one primary configuration for the Clinical profile and two complementary primary configurations plus one fallback for the Balanced and TinyML profiles. The LLM output was subsequently subjected to deterministic validation. For the Balanced and TinyML profiles, the complementary primary configuration was required to retain a composite fidelity score of at least 0.70, improve SSIM-based stability by at least 0.02 relative to the high-fidelity anchor, and remain within 0.05 of the anchor’s normalized deployment-cost value. The fallback was required to be distinct from the primary configurations, provide higher fidelity than the complementary primary configuration, and remain within the same low-cost interval. Every primary and fallback configuration was also required to belong to the corresponding three-objective Pareto-optimal set. GPT-4.1 mini produced a valid selection for all three profiles on the first attempt. All selected configurations passed deterministic validation, and no corrective retry was required. Table 6 summarizes the resulting selections. Table 6: Final Pareto-constrained configurations selected by GPT-4.1 mini for each stakeholder profile. Every primary and fallback configuration belongs to the corresponding three-objective Pareto-optimal set. Profile Primary configurations Fallback Clinical CAM – Balanced CAM; Binary-CAM (τ=0.60τ=0.60) Binary-CAM (τ=0.50τ=0.50) TinyML CAM; Binary-CAM (τ=0.60τ=0.60) Binary-CAM (τ=0.50τ=0.50) Under the Clinical profile, GPT-4.1 mini selected CAM as the single final configuration. CAM achieved the highest composite fidelity score of 0.9397, an SSIM-based stability value of 0.8203, and the minimum deployment-cost value of 0.150. Although other Pareto-optimal configurations achieved higher stability, they did so with lower fidelity. CAM was therefore selected as the fidelity-oriented, low-cost configuration for the Clinical profile. Under the Balanced profile, CAM was selected as the high-fidelity, low-cost anchor. Binary-CAM with τ=0.60τ=0.60 was selected as the complementary primary configuration. It increased stability from 0.8196 for CAM to 0.8481 while retaining a composite fidelity score of 0.7312 and a deployment-cost value of approximately 0.1672. Binary-CAM with τ=0.50τ=0.50 was selected as the fallback because it provided a higher fidelity score of 0.7840, a stability value of 0.8200, and a similar deployment-cost value of approximately 0.1673. The same configuration structure was selected under the TinyML profile. CAM provided a composite fidelity score of 0.9397, a stability value of 0.8216, and the minimum deployment-cost value of 0.150. Binary-CAM with τ=0.60τ=0.60 was selected as the complementary primary configuration, providing a higher stability value of 0.8562 while retaining a fidelity score of 0.7312 and a deployment-cost value of approximately 0.1683. Binary-CAM with τ=0.50τ=0.50 was selected as the fallback, providing a higher fidelity score of 0.7840, a stability value of 0.8296, and a deployment-cost value of approximately 0.1680. CAM was selected across all three profiles as the common high-fidelity, low-cost anchor. Under the Balanced and TinyML profiles, Binary-CAM with τ=0.60τ=0.60 provided a complementary increase in stability while satisfying the specified fidelity and deployment-cost constraints. Binary-CAM with τ=0.50τ=0.50 provided a nearby Pareto-valid fallback with higher fidelity than the complementary primary configuration. The final LLM selections therefore retained distinct operating options within the low-deployment-cost region while remaining subject to explicit and deterministically validated selection constraints. 6 Discussion The results expose a clear limitation of stability-only evaluation. Micro-CAM with s=1s=1 achieved an SSIM of 1.000 but a composite fidelity of only 0.0502, whereas CAM achieved the highest fidelity of 0.9397 with an SSIM of approximately 0.82. Binary-CAM and TopK-CAM showed the same pattern: stronger sparsification increased stability while reducing fidelity. SSIM must therefore be interpreted together with fidelity to avoid favoring nearly invariant but weakly informative attribution maps. The deployment measurements also distinguish spatial compression from runtime efficiency. Binary-CAM retained the same estimated additional SRAM as CAM (27.5625 kB) but introduced only 0.00506–0.01136 ms of additional batch runtime in the reported configurations. Micro-CAM reduced the SRAM estimate to 9.0 kB but produced a larger increase of 0.20326 ms for s=5s=5. Thus, reduced representation size did not necessarily correspond to lower measured post-processing time. Three-objective Pareto filtering retained 12 of 45 eligible Clinical configurations, 13 of 45 Balanced configurations, and 16 of 55 TinyML configurations. CAM was Pareto-optimal under all profiles, combining the minimum deployment-cost proxy of 0.150 with the highest fidelity of 0.9397. LR-CAM configurations with SSIM equal to 1.000 also remained Pareto-optimal, but their fidelity of 0.0502 confirms that Pareto membership alone does not indicate explanation quality. Micro-CAM appeared only in the TinyML analysis because complete timing and cost records for this family were available only for that profile. GPT-4.1 mini selected CAM as the Clinical configuration and selected CAM with Binary-CAM at τ=0.60τ=0.60 for Balanced and TinyML, with Binary-CAM at τ=0.50τ=0.50 as the fallback. All selected configurations were Pareto-valid and passed deterministic validation on the first attempt; no corrective retry was required. This preserves a clear division between LLM-guided selection and deterministic enforcement of Pareto membership, numerical constraints, and output validity. The runtime and deployment-cost results remain comparative rather than hardware-level measurements. Timing was obtained in the experimental execution environment, not on the target Cortex-M7, and the deployment-cost objective used predefined SRAM estimates rather than measured energy. Moreover, the heatmaps were not assessed using lesion masks or clinician evaluation. The conclusions are therefore limited to the implemented fidelity metrics, SSIM perturbations, timing measurements, and deployment-cost proxy. Future work should repeat the timing analysis on physical MCU hardware, obtain parameter-specific memory and energy measurements, evaluate the heatmaps using lesion annotations and dermatologists, and examine sensitivity across repeated LLM runs and additional model architectures. 7 Conclusion This study presented a human-centered framework for selecting XAI configurations under TinyML deployment constraints. Across 67 evaluated configurations, CAM provided the highest composite fidelity and the minimum deployment-cost proxy, while stronger compression and sparsification often increased SSIM stability at the expense of fidelity. Three-objective Pareto filtering retained 12 Clinical, 13 Balanced, and 16 TinyML configurations from their eligible sets. LLMs generated profile-specific method proposals, and GPT-4.1 mini selected final configurations only from the validated Pareto sets. Feasibility, Pareto membership, numerical selection constraints, and output validity remained deterministic. The results constitute a traceable proof of concept; physical MCU measurements and clinical evaluation remain necessary before deployment. Acknowledgements This work was supported by UK Research and Innovation through the EPSRC National Edge AI Hub for Real Data [grant number EP/Y028813/1]. References [1] N. N. Alajlan and D. M. Ibrahim, “Tinyml: Enabling of inference deep learning models on ultra-low-power iot edge devices for ai applications,” Micromachines, vol. 13, no. 6, p. 851, 2022. [2] R. Kallimani, K. Pai, P. Raghuwanshi, S. Iyer, and O. L. López, “Tinyml: Tools, applications, challenges, and future research directions,” Multimedia Tools and Applications, vol. 83, no. 10, p. 29 015–29 045, 2024. [3] M. Neseem, “Ai at the edge: Efficient deep learning for resource-constrained environments,” Ph.D. dissertation, Brown University, 2024. [4] K. Rasheed, A. Qayyum, M. Ghaly, A. Al-Fuqaha, A. Razi, and J. Qadir, “Explainable, trustworthy, and ethical machine learning for healthcare: A survey,” Computers in Biology and Medicine, vol. 149, p. 106043, 2022. [5] A. Bussone, S. Stumpf, and D. O’Sullivan, “The role of explanations on trust and reliance in clinical decision support systems,” in 2015 international conference on healthcare informatics. IEEE, 2015, p. 160–169. [6] D. Shin, “The effects of explainability and causability on perception, trust, and acceptance: Implications for explainable ai,” International journal of human-computer studies, vol. 146, p. 102551, 2021. [7] R. Rosenbacke, Å. Melhus, M. McKee, and D. Stuckler, “How explainable artificial intelligence can increase or decrease clinicians’ trust in ai applications in health care: systematic review,” Jmir Ai, vol. 3, p. e53207, 2024. [8] A. Chaddad, J. Peng, J. Xu, and A. Bouridane, “Survey of explainable ai techniques in healthcare,” Sensors, vol. 23, no. 2, p. 634, 2023. [9] B. Zhou, A. Khosla, A. Lapedriza, A. Oliva, and A. Torralba, “Learning deep features for discriminative localization,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, p. 2921–2929. [10] R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, and D. Batra, “Grad-cam: Visual explanations from deep networks via gradient-based localization,” in Proceedings of the IEEE international conference on computer vision, 2017, p. 618–626. [11] M. T. Ribeiro, S. Singh, and C. Guestrin, “” why should i trust you?” explaining the predictions of any classifier,” in Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, 2016, p. 1135–1144. [12] V. Petsiuk, A. Das, and K. Saenko, “Rise: Randomized input sampling for explanation of black-box models,” arXiv preprint arXiv:1806.07421, 2018. [13] M. Attioui and M. Lahby, “Sustainable traffic congestion forecasting through lightweight explainable ai and tinyml edge deployment: A casablanca case study,” Sustainability, vol. 18, no. 9, p. 4439, 2026. [14] C. Majdoubi, S. E. Mendili, Y. Gahi, and K. El-Khatib, “Avi-shield: An explainable tinyml cross-platform threat detection framework for aviation mobile security,” Information, vol. 17, no. 1, p. 21, 2025. [15] O. Vermesan, V. Piuri, F. Scotti, A. Genovese, R. D. Labati, and P. Coscia, “Explainability and interpretability concepts for edge ai systems,” in Advancing edge artificial intelligence. River Publishers, 2024, p. 197–227. [16] B. Leichtmann, C. Humer, A. Hinterreiter, M. Streit, and M. Mara, “Effects of explainable artificial intelligence on trust and human behavior in a high-risk decision task,” Computers in Human Behavior, vol. 139, p. 107539, 2023. [17] V. Lai, Y. Zhang, C. Chen, Q. V. Liao, and C. Tan, “Selective explanations: Leveraging human input to align explainable ai,” Proceedings of the ACM on Human-Computer Interaction, vol. 7, no. CSCW2, p. 1–35, 2023. [18] S. Spreeuwenberg, “Choose for ai and for explainability,” in OTM Confederated International Conferences” On the Move to Meaningful Internet Systems”. Springer, 2019, p. 3–8. [19] R. Bommasani, D. A. Hudson, E. Adeli, R. Altman, S. Arora, S. von Arx, M. S. Bernstein, J. Bohg, A. Bosselut, E. Brunskill et al., “On the opportunities and risks of foundation models,” arXiv preprint arXiv:2108.07258, 2021. [20] OpenAI, “Gpt-4 technical report,” arXiv preprint arXiv:2303.08774, 2023. [21] C. El Zeinaty, W. Hamidouche, G. Herrou, D. Menard, and M. Debbah, “Can llms revolutionize the design of explainable and efficient tinyml models?” in 2025 International Joint Conference on Neural Networks (IJCNN). IEEE, 2025, p. 1–10. [22] P. Tschandl, C. Rosendahl, and H. Kittler, “The ham10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions,” Scientific data, vol. 5, no. 1, p. 180161, 2018. [23] A. Howard, M. Sandler, G. Chu, L.-C. Chen, B. Chen, M. Tan, W. Wang, Y. Zhu, R. Pang, V. Vasudevan et al., “Searching for mobilenetv3,” in Proceedings of the IEEE/CVF international conference on computer vision, 2019, p. 1314–1324. [24] F. Dehrouyeh, I. Shaer, S. Nikan, F. B. Ajaei, and A. Shami, “Pruning-based tinyml optimization of machine learning models for anomaly detection in electric vehicle charging infrastructure,” in ICC 2025-IEEE International Conference on Communications. IEEE, 2025, p. 1195–1200. [25] —, “Tinyml-enabled resource-efficient framework for real-time network securiy in electric vehicle charging networks,” IEEE Transactions on Network Science and Engineering, vol. 13, p. 5092–5109, 2025. [26] I. Lamaakal, C. Yahyati, Y. Maleh, K. El Makkaoui, and I. Ouahbi, “Explainable kolmogorov–arnold networks for zero-shot human activity recognition on tinyml edge devices,” Machine Learning and Knowledge Extraction, vol. 8, no. 3, p. 55, 2026. [27] M. Sabih, M. Yayla, F. Hannig, J. Teich, and J.-J. Chen, “Robust and tiny binary neural networks using gradient-based explainability methods,” in Proceedings of the 3rd Workshop on Machine Learning and Systems, 2023, p. 87–93. [28] S. Gulati, P. Shah, L. Pathak, C. Trivedi, R. Gupta, and S. Tanwar, “Nanoxai: Explainable ai for anomaly detection in electromagnetic nano-networks of cavs,” in 2025 Joint European Conference on Networks and Communications & 6G Summit (EuCNC/6G Summit). IEEE, 2025, p. 103–108. [29] R. Arthi and y. S. Krishnaveni, “Optimized tiny machine learning and explainable ai for trustable and energy-efficient fog-enabled healthcare decision support system,” International Journal of Computational Intelligence Systems, vol. 17, no. 1, p. 229, 2024. [30] C. Majdoubi, S. El Mendili, and Y. Gahi, “Enhancing mobile threat detection through tinyml and auditable ml,” in 2025 IEEE 8th Congress on Information Science and Technology (CiSt). IEEE, 2025, p. 145–152. [31] Y. Patel, K. Vanjara, M. Thakore, K. Vyas, L. Pathak, R. Gupta, and S. Tanwar, “Xai-enhanced lightweight multi-class attack classification in smart homes with 5g-based secure communication framework,” in 2025 IEEE International Conference on Electronics, Computing and Communication Technologies (CONECCT). IEEE, 2025, p. 1–6. [32] D. Trivedi, S. Bhatia, S. Bhatt, L. Pathak, K. Shastri, K. Vyas, D. Mehta, R. Gupta, and S. Tanwar, “Defeat: Explanable tinyml-based framework for anomaly detection in nano-sensor traffic in 6g-enabled smart grid,” in 2025 5th Intelligent Cybersecurity Conference (ICSC). IEEE, 2025, p. 43–48. [33] S. Shah, K. Vasava, P. Joshi, D. Kansara, L. Pathak, R. Gupta, A. Kumari, and S. Tanwar, “Cosmolite: Lightweight xai-based framework for secure next-generation space sensor networks,” in IEEE INFOCOM 2025-IEEE Conference on Computer Communications Workshops (INFOCOM WKSHPS). IEEE, 2025, p. 1–6. [34] A. M. Uddin, K. M. Ferdouse, and A. M. Sakib, “Edge-deployable tinyml framework for breast cancer detection with explainable ai from dce-mri,” Available at SSRN 6225231. [35] Y. Rong, T. Leemann, T.-T. Nguyen, L. Fiedler, P. Qian, V. Unhelkar, T. Seidel, G. Kasneci, and E. Kasneci, “Towards human-centered explainable ai: A survey of user studies for model explanations,” IEEE transactions on pattern analysis and machine intelligence, vol. 46, no. 4, p. 2104–2122, 2023. [36] J. Kim, H. Maathuis, and D. Sent, “Human-centered evaluation of explainable ai applications: a systematic review,” Frontiers in Artificial Intelligence, vol. 7, p. 1456486, 2024. [37] N. Scharowski, S. A. Perrig, M. Svab, K. Opwis, and F. Brühlmann, “Exploring the effects of human-centered ai explanations on trust and reliance,” Frontiers in Computer Science, vol. 5, p. 1151150, 2023. [38] M. Matejová and J. Paralič, “A multi-criteria decision-making approach for the selection of explainable ai methods,” Machine Learning and Knowledge Extraction, vol. 7, no. 4, p. 158, 2025. [39] T. Vermeire, T. Laugel, X. Renard, D. Martens, and M. Detyniecki, “How to choose an explainability method? towards a methodical implementation of xai in practice,” in Joint European Conference on Machine Learning and Knowledge Discovery in Databases. Springer, 2021, p. 521–533. [40] M. Langer, D. Oster, T. Speith, H. Hermanns, L. Kästner, E. Schmidt, A. Sesing, and K. Baum, “What do we want from explainable artificial intelligence (xai)?–a stakeholder perspective on xai and a conceptual model guiding interdisciplinary xai research,” Artificial intelligence, vol. 296, p. 103473, 2021. [41] J. Stodt, C. Reich, and M. Knahl, Demystifying XAI: requirements for understandable XAI explanations. IOS Press, 2024. [42] A. Zhukov, J. Benois-Pineau, R. Giot, and R. Bourqui, “Explainable ai in image classification tasks: How to choose the best?” in Emotional Data Applications and Regulation of Artificial Intelligence in Society. Springer, 2025, p. 49–70. [43] A. B. Arrieta, N. Díaz-Rodríguez, J. Del Ser, A. Bennetot, S. Tabik, A. Barbado, S. García, S. Gil-López, D. Molina, R. Benjamins et al., “Explainable artificial intelligence (xai): Concepts, taxonomies, opportunities and challenges toward responsible ai,” Information fusion, vol. 58, p. 82–115, 2020. [44] A. Holzinger, “From machine learning to explainable ai,” in 2018 world symposium on digital intelligence for systems and machines (DISA). IEEE, 2018, p. 55–66. [45] M. Velmurugan, C. Ouyang, C. Moreira, and R. Sindhgatta, “Developing a fidelity evaluation approach for interpretable machine learning,” arXiv preprint arXiv:2106.08492, 2021. [46] M. Miró-Nicolau, A. Jaume-i Capó, and G. Moyà-Alcover, “Assessing fidelity in xai post-hoc techniques: A comparative study with ground truth explanations datasets,” Artificial Intelligence, vol. 335, p. 104179, 2024. [47] D. Alvarez-Melis and T. S. Jaakkola, “On the robustness of interpretability methods,” arXiv preprint arXiv:1806.08049, 2018. [48] W. Samek, A. Binder, G. Montavon, S. Lapuschkin, and K.-R. Müller, “Evaluating the visualization of what a deep neural network has learned,” IEEE transactions on neural networks and learning systems, vol. 28, no. 11, p. 2660–2673, 2016. [49] S. Hooker, D. Erhan, P.-J. Kindermans, and B. Kim, “A benchmark for interpretability methods in deep neural networks,” Advances in neural information processing systems, vol. 32, 2019. [50] Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image quality assessment: from error visibility to structural similarity,” IEEE transactions on image processing, vol. 13, no. 4, p. 600–612, 2004. [51] J. Peng, C. Shi, E. Laugeman, W. Hu, Z. Zhang, S. Mutic, and B. Cai, “Implementation of the structural similarity (ssim) index as a quantitative evaluation tool for dose distribution error detection,” Medical physics, vol. 47, no. 4, p. 1907–1919, 2020. [52] K. Deb, “Multi-objective optimisation using evolutionary algorithms: an introduction,” in Multi-objective evolutionary optimisation for product design and manufacturing. Springer, 2011, p. 3–34. [53] F. Doshi-Velez and B. Kim, “Towards a rigorous science of interpretable machine learning,” arXiv preprint arXiv:1702.08608, 2017. [54] P. Tschandl, C. Rinner, Z. Apalla, G. Argenziano, N. Codella, A. Halpern, M. Janda, A. Lallas, C. Longo, J. Malvehy et al., “Human–computer collaboration for skin cancer recognition,” Nature Medicine, vol. 26, no. 8, p. 1229–1234, 2020. [55] G. Comanici, E. Bieber, M. Schaekermann, I. Pasupat, N. Sachdeva, I. Dhillon, M. Blistein, O. Ram, D. Zhang, E. Rosen et al., “Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities,” arXiv preprint arXiv:2507.06261, 2025. IEEEbiographynophoto Zeinab Dehghani Zeinab Dehghani is an MRes student in Computer Science at the University of Hull. Her research focuses on trustworthy and explainable AI, LLM-guided XAI design, and AI assurance for resource-constrained and safety-critical environments. She can be contacted at zeinab.dehghani068@gmail.com. IEEEbiographynophoto Dhavalkumar Thakker is a Professor of Artificial Intelligence (AI) and the Internet of Things (IoT) at the University of Hull, where he leads a group focused on Responsible Artificial Intelligence. His research emphasizes AI Explainability, AI Safety, and Fairness. With nearly two decades of experience, Dhavalkumar has been at the forefront of innovative solutions through funded projects. His interdisciplinary research spans Generative AI and the applications of Edge computing alongside IoT technologies. He has a track record in leveraging AI for Social Good, notably in Smart Cities, Digital Health, and the Circular Economy. Contact him at D.Thakker@hull.ac.uk IEEEbiographynophoto Koorosh Aslansefat is an assistant professor of computer science at the University of Hull, HU6 7RX Hull, U.K., affiliated with the Dependable Intelligent System Group. His research interests span artificial intelligence safety, Markov modeling, and real-time dependability analysis. Aslansefat received his PhD in computer science from the University of Hull. He is a Member of IEEE. Contact him at K.Aslansefat@hull.ac.uk IEEEbiographynophoto Kuniko Paxton is a Postdoctoral Edge AI Researcher at the University of Hull and a member of the National Edge AI Hub, focusing on the development of efficient and responsible machine learning methods for resource-constrained and decentralized environments. Contact her at k.paxton@hull.ac.uk IEEEbiographynophoto Bhupesh Kumar Mishra is a Lecturer in AI and Data Science at the University of Hull. His research interests include Explainable AI, Edge Computing, and the Internet of Things. Dr. Mishra received his PhD in Relied Item Optimization in Disaster Scenarios using computational algorithms from the University of the West of Scotland, UK. Contact him at Bhupesh.Mishra@hull.ac.uk IEEEbiographynophoto Baseer Ahmad is a Lecturer at the University of Hull. His research interests include intelligent predictive maintenance, the Internet of Things, and embedded electronic systems. Dr. Ahmad received his PhD in Predictive Maintenance from the University of the West of Scotland, UK. Contact him at Baseer.Ahmad@hull.ac.uk. IEEEbiographynophoto Rameez Raja Kureshi is a Lecturer and Program Director at the University of Hull, UK. His research interests include Edge AI, fairness, Cyber Security, and the Internet of Things. Dr. Kureshi received his PhD in Artificial Intelligence and IoT from the University of Bradford, UK. He is a Senior Fellow at Advance HE. Contact him at R.Kureshi@hull.ac.uk.