HEAL-AI: Enabling proactive and personalized smart healthcare through hierarchical edge autonomous learning

Objectives

HEAL-AI introduces an integrated architecture that combines FL, neuro-symbolic reasoning, hierarchical contextual attention, and edge computing to meet the critical demands of next-generation smart healthcare. The main objectives of this work are as follows.

It aims to develop a neuro-symbolic FL framework that enables collaborative model training across distributed health nodes (e.g., hospitals, clinics, and wearable devices) without compromising patient privacy. The integration of symbolic medical knowledge with neural models enhances transparency and interpretability in clinical decision-making.

It aims to design a hierarchical contextual attention mechanism that dynamically prioritizes multimodal patient data—such as vitals, medical history, and behavioral patterns—based on real-time context and risk profiles. This enables the system to provide personalized insights tailored to each patient's unique health trajectory.

It aims to implement an energy-aware edge intelligence layer that incorporates model compression and adaptive computational offloading strategies. This component ensures efficient, low-latency inference on resource-constrained IoT and wearable healthcare devices without sacrificing model performance.

It aims to integrate a robust differential privacy mechanism within the FL process. This includes noise injection, secure aggregation, and adaptive privacy budgeting strategies to ensure GDPR compliance and protect sensitive health information at all stages of model training and deployment.

It aims to validate the HEAL-AI framework across diverse real-world and synthetic healthcare datasets, including MIMIC-III, PhysioNet Challenge 2019, eICU, MIMIC-IV, and a custom HEAL-Wearable dataset. Experimental results demonstrate significant improvements in diagnostic accuracy (+17.3%), reduction in inference latency (−32.6%), and increased explainability (+43.8%) over state-of-the-art AI healthcare assistants.

Core novelty of HEAL-AI

HEAL-AI introduces a unified federated neuro-symbolic learning framework that tightly integrates symbolic medical reasoning, hierarchical contextual attention, and adaptive edge intelligence within a privacy-preserving FL paradigm. Unlike prior works that combine these components in isolation, HEAL-AI provides a single end-to-end architecture that jointly optimizes interpretability, personalization, computational efficiency, and privacy across distributed healthcare environments.

Key contributions

The remainder of this paper is structured as follows: Section II discusses related work on neuro-symbolic AI, FL, and smart healthcare technologies. Section III introduces the system architecture and formulates the problem. Section IV details the components of the HEAL-AI framework. Section V outlines the experimental setup. Section VI presents the evaluation results and comparisons. Section VII concludes the paper and discusses potential future research directions.

System architecture overview

The HEAL-AI system is conceptualized as a distributed network of interconnected health nodes comprising hospitals, clinics, personal health devices, and wearable sensors, all collaborating within a privacy-preserving FL framework. Figure 1 illustrates the high-level architecture of the HEAL-AI system.

OPEN IN VIEWER

Figure 1.

HEAL-AI system architecture.

The HEAL-AI architecture features a distributed network of local health nodes—including hospitals, personal devices, and wearables—that process patient data locally while sharing only encrypted model updates. An edge computing layer handles real-time preprocessing and inference on resource-constrained devices using model compression and offloading. An FL coordinator securely aggregates updates with differential privacy to protect data. The neuro-symbolic core combines deep learning and medical knowledge for interpretable decision-making, while a hierarchical attention module prioritizes data streams based on clinical relevance. A dedicated privacy layer ensures regulatory compliance through secure aggregation and adaptive privacy controls. Together, these components enable continuous, personalized, and privacy-preserving healthcare.

Mathematical formulation

This section presents the comprehensive mathematical foundation underlying HEAL-AI's federated neuro-symbolic architecture. Each formulation addresses specific challenges in distributed healthcare AI, from data heterogeneity across medical institutions to privacy preservation and resource optimization on edge devices.

FL framework

HEAL-AI uses an FL approach to enable collaborative model training across distributed healthcare nodes (e.g., hospitals, clinics, and wearable devices) while maintaining data privacy. Each local node k ∈ { 1 , 2 , … , K } holds a private dataset:

D k = { ( x i k , y i k ) } for i = 1 t o n k

(1)

where x i k ∈ R d is a feature vector representing multimodal patient data and y i k is the corresponding label (e.g., diagnosis or outcome).

The global model f θ : R d → R c f _ θ : R d → R c f θ : R d → R c is trained by minimizing the weighted average of local loss functions:

F ( θ ) = Σ k ( ( n k ) n ) ⋅ F k ( θ ) , …

(2)

where n = Σ_k n_k and

F k ( θ ) = ( 1 n k ) ⋅ Σ i L ( f θ ( x i k ) , y i k ) …

(3)

The loss function L(·) can be cross-entropy for classification, MSE for regression, or focal loss for imbalanced classes.

To address data heterogeneity across institutions, HEAL-AI applies personalization using a regularization term:

F ~ k ( θ ) = F k ( θ ) + λ k ⋅ R ( θ , θ k ) , …

(4)

where R ( ⋅ ) can be L 2 n o r m : ∥ θ − θ k ∥ 2 or Kullback–Leibler (KL) divergence. The hyperparameter λ k balances global consistency and local adaptation.

Model updates follow a weighted aggregation scheme:

θ t + 1 = θ t + η ⋅ Σ k ∈ S t ( n k Σ j ∈ S t n i ) ⋅ Δ θ k t , …

(5)

where S_t is the subset of participating nodes in round t and Δθ_kt is the local update at node k.

This design ensures scalable, privacy-aware learning with clinical relevance, giving more influence to larger, data-rich institutions.

Advanced neuro-symbolic integration

HEAL-AI introduces an advanced neuro-symbolic integration strategy that merges the strengths of deep neural networks with symbolic reasoning to ensure accurate yet interpretable clinical decisions. The overall architecture is represented by the following function:

f N S ( x ) = S ( N ( x ) , K , C )

(6)

Here, N(x) is the neural component that transforms input patient data into a hidden representation, K is the medical knowledge base containing structured clinical rules and ontologies, C captures the current clinical context, and S is the symbolic reasoning engine. The neural network N(x) processes multimodal data through a deep architecture defined as

h = N ( x ) = σ L ( W L ⋅ σ { L − 1 } ( W { L − 1 } … σ 1 ( W 1 x + b 1 ) … + b { L − 1 } ) + b L ) …

(7)

where W₁ and b₁ are layer-specific weights and biases, respectively, and σ₁ is the activation function such as the rectified linear unit (ReLU) or softmax depending on the task. The symbolic reasoning component applies medical rules to the neural representation using

y ^ = S ( h , K , C ) = R ( h , { ϕ 1 , ϕ 2 , … , ϕ M } , C )

(8)

where are predefined clinical logic rules applied to the latent representation h. To connect neural outputs to symbolic atoms (e.g., symptoms or treatments), HEAL-AI uses a probabilistic softmax-based mapping:

P ( A i | h ) = exp ( f i ( h ) ) / Σ _ ( j = 1 ) ∧ ( N A ) exp ( f i ( h ) )

(9)

This assigns confidence scores to symbolic concepts based on neural evidence. For clinical inference, a differentiable theorem-proving mechanism is used:

P ( G | K , h ) = Σ _ { r ∈ R _ G } P ( G | b o d y ( r ) , K , h ) × P ( r | K , h )

(10)

Here, G is a clinical goal (e.g., diagnosis), R_G is the relevant set of inference rules, and P(r|K, h) determines the confidence in each rule, computed as

P ( r | K , h ) = σ ( w _ r + α _ r ⋅ E v i d e n c e S t r e n g t h ( h , r ) )

(11)

with w r as the prior confidence and α_r as the weight for neural evidence. Finally, the knowledge base evolves over time as new data and rules emerge using

K { t + 1 } = K t ∪ E t ∪ L ( D t ) ∖ O t

(12)

where E t is the expert-curated updates, L ( D t ) is the rules learned from patient data, and O t is the deprecated or obsolete entries. This hybrid system allows HEAL-AI to make real-time, interpretable, and context-aware decisions by combining the scalability of deep learning with the logic and traceability of symbolic AI.

Hierarchical contextual attention mechanism

The hierarchical contextual attention mechanism in HEAL-AI dynamically prioritizes multimodal healthcare data based on clinical relevance, enabling efficient and interpretable processing of high-dimensional inputs. It operates across three levels. At the modality level, attention weights (α_d) are assigned to each data type (e.g., vital signs, labs, and images) using

α d = S o f t m a x ( v d T t a n h ( W d [ c ; h d ] + b d ) )

(13)

where c is the patient-specific context and h_d is the encoded modality representation.

At the temporal level, attention (β_{d,t}) highlights important time steps in longitudinal data:

β { d , t } = S o f t m a x ( v t T t a n h ( W t [ c ; h { d , t } ; T e m p o r a l C o n t e x t t ] + b t ) )

(14)

At the feature level, individual features within each time step are weighted via

γ { d , t , f } = S o f t m a x ( v f T t a n h ( W f [ c ; h { d , t , f } ; C l i n i c a l P r i o r { d , f } ] + b f ) )

(15)

The context vector c aggregates patient history, risk scores, environment, and social determinants:

c = f c ( [ p ; e ; r ; C l i n i c a l H i s t o r y ; S o c i a l D e t e r m i n a n t s ] ) …

(16)

The final patient representation z integrates all levels:

z = Σ d α d Σ t β { d , t } Σ f γ { d , t , f } h { d , t , f }

(17)

To adapt over time, the context is updated recurrently:

c t = G R U ( c { t − 1 } , [ z { t − 1 } ; Δ x t ; a { t − 1 } ; C l i n i c a l E v e n t s t ] )

(18)

This layered attention allows HEAL-AI to focus on what matters most for each patient, improving both prediction accuracy and clinical transparency.

Edge intelligence optimization

The edge intelligence optimization framework addresses the critical challenge of deploying sophisticated AI models on resource-constrained devices in healthcare settings, including wearable sensors, mobile devices, and embedded medical equipment. This optimization is formulated as a multi-objective constrained optimization problem that balances model performance with computational efficiency.

The primary optimization objective seeks to minimize both the task-specific loss and computational resource consumption:

min θ e F e ( θ e ) + λ c C ( θ e ) subject to C ( θ e ) ≤ B

(19)

In this formulation, θ e ∈ R p e represents the optimized parameters of the edge model (where p e ≪ p compared to the full model); F e ( θ e ) is the task-specific loss function evaluated on edge-relevant data; C ( θ e ) quantifies the computational complexity through metrics such as floating-point operations (FLOPs), memory footprint, and energy consumption; λ c > 0 is a regularization parameter controlling the trade-off between accuracy and efficiency; and B represents the strict resource budget of the target edge device.

The model compression pipeline integrates multiple complementary techniques to achieve optimal size–performance trade-offs:

θ e = Q ( P ( F ( D ( θ , D e ) ) ) )

(20)

This sequential compression process applies knowledge distillation ( D ), structured pruning ( P ), low-rank factorization ( F ), and quantization ( Q ) in a carefully orchestrated sequence to maximize compression while preserving medical accuracy. Each operation is designed to exploit different aspects of model redundancy common in over-parameterized neural networks.

The knowledge distillation process transfers knowledge from a large, accurate teacher model to a compact student model suitable for edge deployment:

L K D ( θ e ) = ( 1 − λ ) L C E ( f θ e ( x ) , y ) + λ L K L ( f θ e ( x ) , softmax ( f θ ( x ) / T ) ) + β L M S E ( h e , h t )

(21)

This comprehensive distillation loss combines multiple knowledge transfer mechanisms: the cross-entropy loss L C E ensures that the student model learns from true labels y, the KL divergence L K L transfers the teacher's soft predictions using temperature T to soften probability distributions and reveal the teacher's uncertainty, and the mean squared error L M S E matches intermediate representations h e and h t from student and teacher models, respectively. The weighting parameters λ and β control the relative importance of different knowledge sources.

For adaptive computational offloading in multitier edge computing environments, we formulate an optimization problem that considers both device capabilities and clinical urgency:

min a { ( 1 − ω ) E ( a ) + ω T ( a ) + μ P ( a ) }

(22)

subject to multiple operational constraints:

T ( a ) ≤ T m a x ( PatientRisk , ClinicalUrgency )

(23)

E ( a ) ≤ E b u d g e t ( BatteryLevel , PowerProfile )

(24)

P ( a ) ≤ P m a x ( PrivacySetting , DataSensitivity )

(25)

In this multi-objective formulation, a ∈ { 0 , 1 } N is a binary decision vector indicating whether each of N model components is executed locally ( a i = 0 ) or offloaded to fog/cloud resources ( a i = 1 ). The objective function balances energy consumption E ( a ) , execution latency T ( a ) , and privacy risk P ( a ) through weights ω ∈ [ 0 , 1 ] and μ ≥ 0 . The constraints ensure that latency remains within clinically acceptable bounds based on patient risk level, energy consumption stays within device capabilities, and privacy protection meets healthcare regulatory requirements.

The latency function T ( a ) models the total time for distributed computation:

T ( a ) = max i [ a i ( T i c o m m + T i r e m o t e ) + ( 1 − a i ) T i l o c a l ] + T s y n c

(26)

This formulation captures the maximum execution time across all components, where T i c o m m represents communication time for offloading component i, T i r e m o t e is the remote execution time, T i l o c a l is the local execution time, and T s y n c accounts for synchronization overhead in distributed execution.

The energy consumption model considers both computational and communication costs:

E ( a ) = ∑ i = 1 N [ a i ( E i c o m m + E i i d l e ) + ( 1 − a i ) E i c o m p ]

(27)

Here, E i c o m m represents the energy cost of transmitting data for component i, E i i d l e is the energy consumed during waiting for remote computation results, and E i c o m p is the energy required for local computation of component i.

Privacy-preserving mechanisms

The privacy-preserving framework in HEAL-AI implements formal differential privacy guarantees to protect sensitive patient information during FL while maintaining the utility of collaborative model training. This approach is essential for compliance with healthcare privacy regulations such as HIPAA, GDPR, and emerging AI governance frameworks.

For each local model update Δ θ k ∈ R p computed by healthcare node k, we apply a carefully calibrated noise addition mechanism:

Δ θ k ~ = Δ θ k + N ( 0 , σ 2 C 2 I p )

(28)

This additive noise mechanism introduces Gaussian noise N ( 0 , σ 2 C 2 I p ) with zero mean and covariance matrix σ 2 C 2 I p (where I p is the p × p identity matrix), ensuring that the noise is independent across parameters. The clipping bound C > 0 represents the L2 sensitivity of the model update, computed as the maximum possible L2 norm of the difference between updates on neighboring datasets (differing by one patient record).

The noise scale σ is carefully calibrated to achieve ( ε , δ ) -differential privacy guarantees:

σ = C 2 ln ( 1.25 / δ ) ε

(29)

This calibration ensures that the privacy mechanism satisfies the formal definition of ( ε , δ ) -differential privacy, where ε > 0 controls the privacy loss (smaller values provide stronger privacy) and δ ∈ ( 0 , 1 ) represents the probability of privacy failure (typically set to 1 / n , where n is the dataset size).

To optimize the fundamental privacy–utility trade-off in healthcare applications, we implement an adaptive privacy budget allocation strategy that allocates more privacy budget to parameters that have greater impact on model utility:

ε i = ε ⋅ S i ∑ j = 1 p S j

(30)

The parameter-specific sensitivity score S i quantifies the importance of parameter θ i for model performance:

S i = | ∂ L ∂ θ i | ⋅ ∥ θ i ∥ 2 ⋅ ClinicalImportance i

(31)

This sensitivity measure combines the gradient magnitude | ∂ L / ∂ θ i | (indicating how much the loss changes with respect to parameter i ), the parameter magnitude ∥ θ i ∥ 2 (larger parameters typically have more influence), and ClinicalImportance i , which encodes domain knowledge about the medical relevance of different model components.

To further enhance privacy protection while maintaining model utility, we implement a sophisticated layer-wise adaptive noise mechanism that recognizes that different layers in neural networks have varying sensitivity to noise:

σ l = σ 0 ⋅ β d l ⋅ LayerSensitivity l AvgSensitivity

(32)

In this formulation, σ l is the noise scale for layer l, σ 0 is the base noise scale, β ∈ ( 0.8 , 0.95 ) is a decay factor that typically reduces noise for deeper layers (which often learn more abstract, less sensitive features), d l represents the depth of layer l relative to the output layer, and the sensitivity ratio ensures that layers with higher sensitivity to perturbations receive proportionally more noise for equivalent privacy protection.

The federated aggregation process incorporates secure multiparty computation to prevent the server from observing individual model updates:

θ t + 1 = θ t + η ∑ k = 1 K n k n SecureAggregate ( Δ θ k ~ )

(33)

The SecureAggregate function implements cryptographic protocols that enable the server to compute the weighted average of model updates without accessing individual contributions, providing an additional layer of privacy protection beyond differential privacy.

To track privacy consumption over multiple rounds of FL, we employ the advanced composition theorem for differential privacy:

ε t o t a l = 2 T ln ( 1 / δ ) ⋅ ε r o u n d + T ⋅ ε r o u n d ⋅ ( e ε r o u n d − 1 )

(34)

This composition bound ensures that the total privacy loss ε t o t a l over T communication rounds remains within acceptable limits, where ε r o u n d is the privacy budget allocated per round. The composition theorem allows for principled privacy budget management throughout the entire FL process.

Federated neuro-symbolic architecture

The core fed neuro-symbolic architecture and the HEAL-AI learning process unify the disciplines of neural and symbolic learning to enable collective learning across a federate of health nodes and maintain interpretability. In a decentralized healthcare system, this structure addresses the fundamental issue of systematizing the explainability of symbolic systems with the learning ability of neural networks. Figure 2 represents an architecture of the proposed system in greater detail, with an emphasis placed on bidirectional information flow between the neural and symbolic parts of the FL framework.

OPEN IN VIEWER

Figure 2.

HEAL-AI federated neuro-symbolic architecture.

As demonstrated in Figure 2, a synergy of distributed learning with interpretable reasoning can potentially achieve it by allowing healthcare institutions to collaborate, thereby improving their diagnostic precision through privacy and security of patient data and producing clinically interpretable outcomes. Raw multimodal data are encoded in the special encoders of the neural network and on the graph of medical knowledge and clinical guidelines of the symbolic network, which provide intelligible decision assistance. The FL protocol in HEAL-AI follows a sophisticated client–server architecture that incorporates advanced privacy preservation mechanisms, adaptive optimization strategies, and Byzantine fault tolerance. The protocol is designed to handle the heterogeneous nature of medical data across different healthcare institutions while ensuring robust convergence under non-IID data distributions.

HEAL-AI introduces a composite loss function to enhance model performance in federated healthcare environments. It combines core task loss with additional regularizations to manage client drift, enforce smooth parameter transitions, and improve fairness across demographic groups. Domain adaptation is achieved through adversarial learning to handle data variations across institutions, while consistency and clinical constraint terms ensure robust, medically aligned predictions. Client selection is refined using data quality, device resources, diversity, and reliability metrics, prioritizing nodes that contribute meaningful updates. Learning rates are adjusted dynamically based on data size and variance to balance training contributions. Gradient clipping and stability checks further ensure safe optimization during decentralized learning. Privacy is maintained through Rényi differential privacy with structured noise injection that preserves gradient correlations. Malicious client behavior is mitigated by scoring updates based on cosine similarity and verifying with zero-knowledge proofs. Aggregation excludes unreliable updates to ensure robustness against Byzantine threats. Global momentum is adjusted using variance reduction to stabilize learning, while a curvature correction mechanism integrates second-order information to accelerate convergence without high computational cost. HEAL-AI's neuro-symbolic framework blends deep neural networks with structured medical knowledge to offer explainable and trustworthy AI for healthcare. The knowledge base includes factual, probabilistic, temporal, and causal layers tailored to clinical contexts. Neural encoders process diverse data types—text, imaging, vitals, genomics—using advanced models like BERT, CNN-LSTM, and Graph Transformers. These embeddings are fused via hierarchical cross-modal attention, incorporating clinical priors to strengthen decision relevance. Symbolic reasoning performs probabilistic and causal inference to generate transparent, medically grounded predictions. The system ensures bidirectional consistency between neural outputs and symbolic rules, enforcing alignment through additional losses that promote explainability, causality, and temporal coherence. This comprehensive architecture enables personalized, privacy-preserving, and clinically reliable AI support in distributed healthcare networks. The hierarchical contextual attention mechanism in HEAL-AI implements a sophisticated multi-scale attention architecture that dynamically adapts to patient-specific risk profiles, temporal contexts, and clinical scenarios. Figure 3 demonstrates the intricate attention hierarchy and its adaptive behavior across different clinical contexts.

OPEN IN VIEWER

Figure 3.

HEAL-AI hierarchical contextual attention mechanism.

As illustrated in Figure 3, the attention mechanism operates at multiple granularities, enabling a fine-grained focus on clinically relevant information while maintaining global context awareness. The hierarchical structure allows the system to adaptively prioritize different aspects of patient data based on the current clinical scenario, with the top level focusing on modality selection, the middle level on temporal attention, and the bottom level on feature-specific attention.

HEAL-AI incorporates clinical relevance weighting to ensure attention is aligned with expert guidelines and patient-specific conditions. It dynamically scores each modality based on how applicable clinical guidelines are to a patient and how relevant each data type is to those guidelines.

The system also employs an uncertainty-aware attention mechanism, which combines model and data uncertainty to down-weight unreliable inputs, enhancing trustworthiness in predictions. Additionally, causal attention constraints enforce biologically and clinically valid attention paths based on known cause–effect relationships between variables.

An adaptive temperature mechanism fine-tunes attention sharpness depending on the model's confidence and contextual uncertainty. Meanwhile, a temporal bias mechanism models time decay and clinical periodicity (like medication schedules) to capture critical time-sensitive patterns in patient data. HEAL-AI adopts a meta-learning framework that enables rapid adaptation to new patients while ensuring clinical safety and fairness. The system learns optimal model parameters that generalize across individuals and fine-tunes those using patient-specific data and constraints. For patients with limited data, the model uses similarity-based transfer learning, drawing knowledge from clinically and demographically similar patients. This is guided by a multimodal similarity function combining demographics, medical conditions, and genomic profiles. A continual adaptation module updates patient models over time using online learning. It considers prediction confidence, urgency of clinical context, and safety to avoid risky updates. This ensures that the model remains adaptive and relevant while minimizing harmful drift. Lastly, a novelty detection module identifies unfamiliar patterns using historical comparisons, entropy, clinical divergence, and data distribution shifts. This safeguards against inaccurate responses to unseen or rare patient conditions.

Table 1 provides a comprehensive comparison of different model compression techniques applied to HEAL-AI, demonstrating the effectiveness of each approach across multiple critical metrics.

OPEN IN VIEWER

Table 1.

Comprehensive comparison of model compression techniques for HEAL-AI edge deployment.

Compression technique	Model size (MB)	Inference time (ms)	Energy (mJ)	Memory (MB)	AUC MIMIC-III	AUC HEAL-Wearable	Clinical safety
None (full model)	87.6	245.8	612.5	156.3	0.912	0.843	0.978
Knowledge distillation (KD)	32.3	52.7	131.6	58.9	0.903	0.835	0.971
Quantization (8-bit) (Q)	22.1	36.9	93.2	39.1	0.904	0.836	0.973
Pruning (90% sparse) (P)	11.2	32.8	82.1	19.8	0.889	0.817	0.952
Structured sparsification (SS)	13.8	25.3	68.9	24.7	0.895	0.825	0.963
Low-rank factorization (LRF)	14.5	31.6	79.2	26.8	0.901	0.832	0.968
Neural architecture search (NAS)	18.9	29.7	73.5	32.1	0.905	0.834	0.970
KD + Q	9.2	24.5	62.3	17.4	0.897	0.830	0.965
KD + P	7.6	21.3	55.8	14.2	0.885	0.813	0.948
KD + SS	8.4	19.7	51.6	15.8	0.891	0.820	0.957
KD + LRF	8.9	22.8	58.4	16.5	0.899	0.828	0.962
KD + NAS	10.3	20.5	53.2	18.9	0.901	0.831	0.966
KD + Q + SS	4.2	12.3	31.8	8.7	0.886	0.815	0.951
KD + Q + SS + LRF	3.4	8.7	22.5	6.9	0.882	0.812	0.946

As demonstrated in Table 1, the comprehensive compression pipeline achieves remarkable efficiency gains while maintaining clinical safety. The best-performing combination (KD + Q + SS + LRF) delivers 25.8× compression ratio with 28.2× faster inference and 27.2× lower energy consumption, while maintaining 96.7% of the original accuracy on MIMIC-III, 96.3% on the HEAL-Wearable dataset, and 96.7% clinical safety score. The progressive degradation pattern shows that knowledge distillation provides the best foundation for maintaining performance under compression, while the addition of quantization and structured sparsification yields the most favorable efficiency–accuracy trade-offs.

The HEAL-AI offloading framework dynamically decides where to execute different components of the healthcare model—locally, at the edge, or in the cloud—based on real-time factors such as patient risk, device capability, energy levels, network conditions, and clinical urgency. Each model component is evaluated for its complexity, energy needs, privacy impact, and clinical importance. A dependency graph maps inter-component relationships, while clinical impact scores help prioritize components that are safety-critical. The system uses a multi-objective optimization strategy to minimize latency, energy usage, and cost while maximizing accuracy, privacy, and reliability. It dynamically adjusts weightings for each objective depending on the current healthcare scenario—e.g., a critical patient gets higher weight for latency and accuracy, while low-battery situations prioritize energy efficiency. The final strategy is optimized using dynamic programming and then validated through a clinical risk assessment module. If needed, adjustments are made to prioritize patient safety, ensuring a “safety-first” decision. Overall, the framework ensures that computational tasks are smartly distributed across layers (device, edge, cloud) while adhering to strict medical standards and maximizing system efficiency.

Figure 4 clearly illustrates the superior performance of HEAL-AI's adaptive offloading compared to baseline approaches. The adaptive strategy successfully navigates the energy–latency trade-off space, achieving optimal points that static strategies cannot reach. Particularly notable is the algorithm's ability to maintain low latency for high-risk patients (red points) while optimizing energy efficiency for stable patients (blue points). The Pareto frontier demonstrates that clinical context-aware optimization achieves better trade-offs than one-size-fits-all approaches, with the adaptive method reducing energy consumption by up to 65% for low-risk scenarios while maintaining sub-100 ms latency for critical cases.

OPEN IN VIEWER

Figure 4.

Energy–latency trade-off analysis for HEAL-AI offloading strategies.

HEAL-AI integrates an intelligent energy optimization framework that proactively manages power consumption while preserving clinical performance. It predicts future energy demands using a hybrid model that combines physics-based calculations, machine learning (via LSTM), and anticipated clinical workload. This allows the system to estimate energy use based on device usage patterns, patient risk, historical energy trends, and task-specific clinical needs. To further optimize energy use, a dynamic precision scaling mechanism is employed. It uses reinforcement learning to adjust the computational precision (bit-width) of operations based on current energy levels, task complexity, and clinical urgency. A built-in safety mechanism ensures full precision for high-risk patients, preventing dangerous performance degradation. Bit-width selection is guided by real-time clinical risk assessment. Patients with higher risk or critical tasks are automatically assigned higher precision levels, while others receive energy-efficient settings that still meet clinical safety thresholds. The system also features an adaptive sampling controller. It modifies the data collection frequency in real-time by considering patient risk, current energy stress, presence of novel patterns, critical monitoring periods (e.g., medication windows), and physiological variability. This ensures that data is collected more frequently when clinically important, while reducing unnecessary sampling during stable or low-risk periods—extending device battery life without compromising care quality. Together, these predictive and adaptive mechanisms enable HEAL-AI to operate efficiently in resource-constrained healthcare environments, especially during continuous monitoring or mobile edge deployments.

Privacy preservation and security framework

HEAL-AI integrates an advanced differential privacy mechanism using adaptive budget allocation, correlated noise injection, and clinical utility preservation. The system uses Rényi Differential Privacy with a clinical-aware noise matrix that prioritizes medically relevant data. Gradient sensitivity, convergence dynamics, and clinical importance guide the level of noise per parameter. Clinical correlations are preserved using structured noise based on medical data covariances and domain knowledge.

Privacy budgets are allocated by weighing clinical and gradient importance. More budget is reserved during high-risk or high-learning intervals. The trade-off between privacy and clinical performance is quantified in Table 2.

OPEN IN VIEWER

Table 2.

Privacy–utility trade-off in HEAL-AI federated learning.

Method	Accuracy (%)	AUC	F1 score	Clinical consistency	Membership inference (%)	Attribute inference (%)	Property inference (%)
No DP	91.2	0.912	0.897	0.943	89.4	67.8	45.2
10.0	89.8	0.901	0.885	0.931	68.7	42.3	28.9
5.0	88.6	0.893	0.876	0.924	59.1	34.6	21.4
2.0	86.5	0.876	0.861	0.912	47.3	23.7	14.8
1.0	83.7	0.852	0.839	0.895	38.2	16.9	9.7
0.5	79.2	0.821	0.807	0.871	28.4	11.2	6.1
0.1	67.4	0.732	0.718	0.798	12.6	4.8	2.3
Adaptive 1.0	85.3	0.867	0.852	0.908	32.1	14.2	8.4
Adaptive 0.5	81.8	0.841	0.825	0.889	22.7	8.9	4.8
Adaptive 0.1	71.2	0.759	0.743	0.823	9.8	3.1	1.7
Clinical-aware 1.0	86.7	0.879	0.864	0.921	35.8	15.7	9.2
Clinical-aware 0.5	83.4	0.856	0.841	0.903	25.3	10.4	5.6
Clinical-aware 0.1	73.8	0.781	0.767	0.847	11.2	3.8	2.1

The secure aggregation protocol includes threshold secret sharing with medical verification, Byzantine-resilient filtering, and zero-knowledge proof integration. Local updates are validated for clinical consistency before aggregation. Domain knowledge helps identify malicious updates.

Verifiable computation mechanisms produce proof logs and clinical audit trails to ensure transparency and trust.

HEAL-AI uses a hierarchical key management structure based on role and patient consent. Session keys are derived contextually. Homomorphic encryption allows secure aggregation of encrypted updates.

Finally, attribute-based access control ensures that only justified clinical roles can access sensitive data, while a machine learning-based anomaly detector monitors and flags suspicious behavior based on access frequency, duration, and type.

Experimental setup

To evaluate HEAL-AI across a range of healthcare applications, five diverse datasets were used, each targeting specific scenarios of federated healthcare AI.

The MIMIC-III dataset provides detailed critical care data from 46,520 patients across 58,976 ICU admissions at Beth Israel Deaconess Medical Center between 2001–2012, including physiological measurements, laboratory tests, medication records, clinical notes, demographics, and outcomes. In this study, MIMIC-III was used for in-hospital mortality prediction, length-of-stay estimation, diagnosis classification, and medication recommendation. For all tasks, 24–48 h observation windows following ICU admission were employed, with evaluation performed at the end of each window. Records with missing outcome labels, ICU stays shorter than 24 h, or excessive missing measurements after preprocessing were excluded.

The PhysioNet/Computing in Cardiology Challenge 2019 dataset focuses on early sepsis detection, containing hourly clinical data from 40,336 patients across three U.S. hospital systems. The prediction target is sepsis onset, defined according to Sepsis-3 criteria, with models trained to predict sepsis up to 6 h prior to clinical onset using hourly rolling evaluation windows. Patients with ambiguous sepsis labels, insufficient pre-onset observation time, or incomplete vital sign records were excluded.

The eICU Collaborative Research Database comprises high-resolution data from 200,859 ICU admissions at 208 hospitals across the United States between 2014–2015. This dataset was used for cross-hospital mortality prediction and clinical risk stratification, enabling evaluation under heterogeneous institutional conditions. Prediction targets were evaluated using the first 24 h of ICU data, and admissions with missing discharge outcomes, duplicated stays, or fewer than 24 h of recorded observations were excluded to ensure consistent endpoint definition and fair comparison across institutions.

The custom HEAL-Wearable Dataset was collected from 1247 adult participants aged 18–75 over a six-month monitoring period using commercially available wearable devices. Participants were recruited from urban hospitals, suburban clinics, and rural healthcare centers to ensure demographic and environmental diversity. The dataset includes continuous physiological signals such as heart rate, heart rate variability, physical activity, and sleep patterns, along with environmental measurements (air quality indices, weather conditions, and coarse-grained geolocation) and behavioral data (dietary patterns, exercise routines, and sleep habits). Clinical labels were assigned based on physician-verified monthly clinical assessments, laboratory test panels, and confirmed health outcomes, including hospitalizations and medication adherence records. Missing data were handled using forward filling for short-duration signal gaps, while samples with prolonged data loss were excluded from analysis. Participants with incomplete informed consent, device malfunction exceeding 30% of the monitoring period, or missing clinical follow-up information were excluded. This dataset supports rigorous evaluation of edge intelligence and remote health monitoring under realistic real-world conditions.

Lastly, the MIMIC-IV dataset (https://www.kaggle.com/datasets/mehrnooshazizi/mimic-iv-dataset) extends MIMIC-III (https://www.kaggle.com/datasets/asjad99/mimiciii) with 11 years of hospital data from 523,740 admissions and 382,278 patients (2008–2019), offering improved data quality, additional emergency department and ward admissions, chest X-ray reports, and social determinants of health (e.g., insurance status, discharge disposition). It is primarily used for temporal validation and assessing HEAL-AI's adaptability to contemporary clinical environments.

To realistically emulate FL in healthcare settings, we implemented a detailed simulation environment with institution-based partitioning that reflects real-world clinical heterogeneity. Specifically, data was divided among 12 simulated healthcare institutions representing diverse operational profiles: 3 large academic medical centers (35% of data) with comprehensive diagnostic capabilities and broad patient populations; 5 community hospitals (45%) handling general medical cases; 2 specialty centers (12%) focusing on diseases such as cardiology and oncology; and 2 rural health centers (8%) with limited resources and primary care focus. These partitions enabled us to simulate practical deployment conditions for federated models.

To handle data heterogeneity, we modeled non-IID distributions using the formulation:

P ( D k ) = α × P g l o b a l ( D ) + ( 1 − α ) × P l o c a l ( k ) ( D ) …

(35)

Here, P ( D k ) represents the data distribution at institution k, P g l o b a l ( D ) is the global population distribution, P l o c a l ( k ) ( D ) reflects institution-specific characteristics, and α is a tunable parameter ranging between 0 and 1 that controls the level of distribution shift. We explored three levels of heterogeneity:

Mild heterogeneity: α = 0.8

Moderate heterogeneity: α = 0.5

Severe heterogeneity: α = 0.2

The institution-specific distribution P l o c a l ( k ) ( D ) is further defined as a product of conditional distributions based on institution type, local demographics, and case mix restrictions:

P l o c a l ( k ) ( D ) = ∏ P b i a s ( k ) ( X j | I n s t i t u t i o n T y p e k , D e m o g r a p h i c s k , C a s e R e s t r i c t i o n k ) …

(36)

This formulation captures real-world clinical variability among healthcare institutions, enabling a realistic simulation of FL performance under non-IID data conditions.

All experiments were conducted using PyTorch on NVIDIA A100 GPUs with CUDA support. Neural encoders included 1D-CNN–LSTM models for time-series data, transformer-based encoders for clinical text, and graph neural networks for relational data, each with 3–6 layers and hidden dimensions ranging from 128 to 512. FL was performed for 150 communication rounds with 30–50% client participation per round, local batch size of 64, and Adam optimizer with learning rates in the range of 1e−4 to 3e−4. Differential privacy noise scales were set according to the specified ε values, and gradient clipping thresholds were dynamically adjusted per round.

All datasets were obtained from officially maintained and publicly available repositories. Preprocessing included removal of physiologically implausible values, normalization of continuous variables, temporal aggregation into fixed observation windows, and imputation of missing values using forward filling or population statistics where appropriate. Records with insufficient observation length, missing outcome labels, or excessive missing data were excluded to ensure reliable evaluation.

Baseline methods for comparison

We compared HEAL-AI against eight state-of-the-art AI and FL models. The Centralized Deep Learning (CDL) model served as an upper-bound benchmark, featuring a multimodal architecture with attention mechanisms and domain-specific encoders: 1D-CNN combined with LSTM for time-series data, TabNet (https://www.kaggle.com/datasets/mohamed3abdelrazik/tabnet-dataset) for structured tabular data, and BioBERT for processing clinical text.

FedAvg was used as the baseline FL method, enhanced with client sampling and learning rate scheduling. FedProx addressed non-IID challenges by adding a proximal term to the loss function, with the proximal coefficient (µ) tuned using a grid search over the values [0.001, 0.01, 0.1, 1.0]. FedNova mitigated objective inconsistency by normalizing local updates prior to aggregation.

We also evaluated domain-specific models such as HARNESS, a hierarchical attention-based recurrent neural network for EHRs; MedicalQA, a neuro-symbolic model leveraging medical knowledge graphs for reasoning; HealthFog, an edge-cloud integrated framework optimized for constrained environments; and GraphCare, which uses graph neural networks to model inter-patient relationships for clinical predictions.

All baseline models were implemented in accordance with their original published descriptions, using the reported architectures, training procedures, and hyperparameter settings wherever such details were available. For baseline methods with incomplete implementation specifications, commonly adopted defaults consistent with prior literature were applied to ensure reasonable and stable performance. All models were trained and evaluated within the same experimental pipeline to maintain consistency in data preprocessing, evaluation metrics, and computational environment; however, this unified implementation may introduce potential bias. To improve transparency and reproducibility, future revisions will include the public release of baseline implementations, detailed hyperparameter configurations, and experimental scripts. The reported improvements in explainability, particularly relative to MedicalQA, arise from differences in evaluation criteria, as HEAL-AI incorporates symbolic rule coverage and attention coherence metrics that are not explicitly modeled in purely neural or graph-based baseline methods.

All baseline models were trained using identical data splits, preprocessing pipelines, and evaluation metrics. Hyperparameters were selected based on reported values in original publications or tuned using validation sets within fixed computational budgets. Training time, number of parameters, and hardware usage were monitored to ensure comparable resource allocation across methods.

Comprehensive performance evaluation

All experimental results reported in this section are presented as the mean and standard deviation (mean ± SD) computed over five independent runs with different random seeds to ensure robustness and stability of the observed performance. Where applicable, 95% confidence intervals (CIs) were estimated using bootstrap resampling. Statistical significance of performance differences between HEAL-AI and baseline methods was evaluated using paired t-tests, with p < 0.05 considered statistically significant. This statistical reporting framework was applied consistently across diagnostic accuracy, early-detection lead time, inference latency, energy consumption, and explainability evaluations to support reliable and reproducible comparison.

The evaluation metrics used in this study were selected to reflect both clinical relevance and system-level feasibility in real-world healthcare deployments. Accuracy, AUC, and F1 score quantify diagnostic reliability and robustness, particularly under class imbalance commonly observed in critical care and remote monitoring datasets. Early-detection lead time is included to capture timeliness in life-critical conditions such as sepsis, where earlier prediction directly impacts patient outcomes. Inference latency and energy consumption evaluate the practicality of deploying HEAL-AI on resource-constrained edge devices, which is essential for real-time and continuous healthcare monitoring. Finally, explainability metrics assess the transparency and clinical interpretability of model outputs, which are required for clinician trust, validation, and adoption in decision support systems.

Table 3 compares centralized, federated, edge-based, graph-based, and neuro-symbolic learning methods on four healthcare datasets: MIMIC-III, PhysioNet/Computing in Cardiology Challenge 2019, eICU, and the HEAL-Wearable dataset. Performance is evaluated using diagnostic accuracy (Acc), area under the ROC curve (AUC), early-detection lead time (Early, hours) for time-critical tasks such as sepsis detection, and F1 score for remote monitoring scenarios. The Improvement row reports relative gains of HEAL-AI over the strongest baseline for each metric, highlighting its effectiveness in accuracy, timeliness, and generalization across heterogeneous clinical and wearable data sources.

OPEN IN VIEWER

Table 3.

Comprehensive diagnostic performance comparison across datasets and medical conditions.

Method	MIMIC-III		PhysioNet 2019		eICU		HEAL-Wearable
	Acc (%)	Early (h)	AUC	Early (h)	Acc (%)	AUC	F1 score
CDL	79.2	4.8	0.821	2.7	77.8	0.804	0.742
FedAvg	81.5	5.3	0.835	3.1	80.2	0.823	0.768
FedProx	83.7	5.7	0.842	3.3	82.5	0.836	0.781
FedNova	84.1	5.9	0.846	3.5	83.1	0.841	0.786
HARNESS	85.9	6.8	0.866	4.2	84.1	0.851	0.798
MedicalQA	87.2	6.3	0.854	3.9	85.3	0.847	0.805
HealthFog	83.4	7.4	0.840	4.6	81.9	0.832	0.773
GraphCare	86.8	6.1	0.871	4.0	84.9	0.858	0.812
HEAL-AI	91.2	8.5	0.912	5.9	89.8	0.889	0.867
Improvement	+4.6%	+25.0%	+4.7%	+40.5%	+5.3%	+3.6%	+6.8%

The bold and italicized values represent the superior performance of HEAL-AI across all datasets and metrics compared to all baseline.

The comparative results in Table 3 indicate that HEAL-AI consistently outperforms baseline methods across all datasets due to its joint integration of multimodal hierarchical attention, neuro-symbolic reasoning, and personalized federated optimization. The hierarchical attention mechanism enables more effective temporal and contextual feature selection from heterogeneous clinical data, while symbolic reasoning incorporates structured medical knowledge that improves decision consistency in complex conditions such as sepsis and multi-morbidity. In contrast, baseline methods rely primarily on statistical correlations or single-modality representations, limiting their ability to capture nuanced clinical relationships. However, these performance gains come with increased model complexity and modest communication overhead compared to simpler federated or centralized baselines, particularly under stricter privacy constraints. This highlights an inherent trade-off between predictive performance, interpretability, and system efficiency, which HEAL-AI addresses through adaptive edge optimization and privacy-aware training.

Table 3 reveals several critical findings that demonstrate HEAL-AI's clinical superiority. The system achieves a 4.6% improvement in diagnostic accuracy over the best-performing baseline (MedicalQA) on MIMIC-III, which represents a significant clinical advancement given the difficulty of the critical care prediction tasks. More importantly, HEAL-AI enables identification of sepsis 5.9 h earlier than standard clinical protocols on the PhysioNet 2019 dataset (https://www.kaggle.com/datasets/farjanayesmin/the-physionet-challenge-2019-dataset), representing a 40.5% improvement over HARNESS. This early-detection capability could potentially save lives, as each hour of delay in sepsis treatment increases mortality risk by 7.6% according to clinical studies.

The performance gains are consistent across all datasets, indicating robust generalization capabilities. On the eICU database, which represents diverse multicenter data, HEAL-AI maintains a 5.3% accuracy improvement, demonstrating effective handling of cross-institutional data heterogeneity. The custom HEAL-Wearable dataset shows a 6.8% F1 score improvement, validating the system's effectiveness for continuous remote monitoring applications.

Figure 5 provides a detailed breakdown of diagnostic performance across specific medical conditions, revealing where HEAL-AI's neuro-symbolic approach provides the greatest clinical value.

OPEN IN VIEWER

Figure 5.

HEAL-AI computational efficiency analysis across edge devices.

Figure 6 demonstrates that HEAL-AI's advantages are most pronounced for complex medical conditions that benefit from multimodal data integration and temporal reasoning. For sepsis detection, the combination of vital signs, laboratory values, and clinical context processed through the hierarchical attention mechanism enables more nuanced pattern recognition than single-modality approaches. The neuro-symbolic integration proves particularly valuable for conditions like delirium, where clinical guidelines provide structured knowledge that enhances pattern recognition in subtle behavioral and cognitive changes.

OPEN IN VIEWER

Figure 6.

(a–e) HEAL-AI performance analysis across diverse healthcare datasets.

The top row of scatter plots compares HEAL-AI with baseline models (CDL, FedAvg, FedProx, HARNESS, HealthFog, etc.) across four key metrics: MIMIC-III accuracy, PhysioNet 2019 AUC, average early-detection lead time (hours), and HEAL-Wearable F1 score. HEAL-AI consistently outperforms all baselines (highlighted with red markers). The bottom bar chart summarizes HEAL-AI's superiority across all datasets with quantified improvements: +4.6% in MIMIC-III accuracy, +4.7% in PhysioNet AUC, +5.3% in eICU accuracy, and +6.8% in wearable F1 score. These results demonstrate HEAL-AI's robust generalization, superior diagnostic precision, and real-time applicability across federated healthcare environments.

Table 4 compares the computational efficiency of centralized, federated, edge-based, and neuro-symbolic healthcare AI methods when deployed on three representative edge platforms: smartphones, wearable devices, and IoT sensors. Performance is evaluated using end-to-end inference latency (milliseconds) and energy consumption per inference (millijoules), which are critical metrics for real-time and battery-constrained healthcare applications. The final row (vs. HealthFog) reports the relative percentage reduction achieved by HEAL-AI compared to the most efficient baseline (HealthFog), highlighting HEAL-AI's suitability for low-latency, energy-efficient deployment in resource-constrained edge healthcare environments.

OPEN IN VIEWER

Table 4.

Comprehensive analysis of inference latency and energy consumption across edge devices.

Method	Smartphone		Wearable		IoT sensor
	Latency (ms)	Energy (mJ)	Latency (ms)	Energy (mJ)	Latency (ms)	Energy (mJ)
CDL	254.7	638.2	412.3	205.8	687.5	89.2
FedAvg	249.2	621.5	403.6	201.4	675.8	87.6
FedProx	256.8	642.1	415.9	207.4	692.3	90.8
FedNova	251.4	627.3	407.2	203.1	681.9	88.4
HARNESS	217.3	543.1	352.6	176.1	586.4	76.3
MedicalQA	283.5	708.2	460.1	229.8	767.2	98.5
HealthFog	168.4	421.2	273.2	136.4	455.6	61.2
GraphCare	232.1	581.7	376.8	188.3	627.4	81.9
HEAL-AI	113.6	284.1	184.3	92.1	306.9	41.7
vs. HealthFog	−32.6%	−32.5%	−32.5%	−32.5%	−32.6%	−31.9%

The bold and italicized values represent the superior efficiency of HEAL-AI on resource-constrained edge devices (smartphones, wearables, IoT sensors).

Table 4 demonstrates HEAL-AI's exceptional efficiency gains through its comprehensive edge intelligence framework. The 32.6% reduction in inference latency compared to HealthFog (the most efficient baseline) is achieved through multiple optimization techniques working synergistically: knowledge distillation reduces model complexity while preserving accuracy, quantization decreases memory bandwidth requirements, and adaptive computation offloading optimizes the distribution of workload between edge and cloud resources.

The energy efficiency improvements are particularly critical for wearable devices, where HEAL-AI achieves 92.1 mJ per inference compared to 136.4 mJ for HealthFog. This 32.5% reduction translates to significantly extended battery life for continuous monitoring applications. For IoT sensors with limited power budgets, the 31.9% energy reduction enables deployment in previously infeasible scenarios such as remote patient monitoring in underserved areas.

Figure 5 reveals HEAL-AI's superior scalability characteristics essential for large-scale healthcare deployment. The sublinear growth in convergence time results from the intelligent client selection strategy that prioritizes high-quality data sources and the adaptive aggregation mechanism that efficiently handles heterogeneous contributions. Communication costs grow more slowly than baseline methods due to the compression techniques and selective parameter updates. Model accuracy remains stable even as network size increases, demonstrating robust handling of increasing data heterogeneity across diverse healthcare institutions.

This multi-panel figure evaluates HEAL-AI's performance in terms of latency and energy efficiency across smartphones, wearables, and IoT sensors. The top-left panel shows inference latency distributions, highlighting HEAL-AI's significantly lower latency (red stars). The top-right histogram compares energy consumption distributions, with HEAL-AI achieving the lowest energy usage. The bottom-left scatter plot reveals a strong negative correlation between HEAL-AI's latency and energy consumption across devices. The bottom-right bar chart quantifies average improvements: HEAL-AI achieves 32.6% latency reduction and 32.3% energy savings across all device categories, confirming its suitability for real-time, resource-constrained healthcare environments.

Table 5 reports a structured comparison of explainability performance across centralized, federated, edge-based, and neuro-symbolic healthcare AI models. Explanations were evaluated by five board-certified clinicians using a predefined rubric covering overall explainability quality, clinical relevance, decision confidence, and attention coherence, each scored on a 1–10 scale. Rule coverage (%) quantifies the proportion of model predictions supported by explicit symbolic or rule-based reasoning. The final row (vs. MedicalQA) reports relative percentage improvements achieved by HEAL-AI over the strongest explainability baseline, highlighting the benefits of neuro-symbolic integration for transparent and clinically interpretable decision support.

OPEN IN VIEWER

Table 5.

Comprehensive explainability assessment through clinical expert evaluation.

Method	Explainability score (1–10)	Clinical relevance (1–10)	Decision confidence (1–10)	Attention coherence (1–10)	Rule coverage (%)
CDL	3.7	4.2	3.8	3.9	0.0
FedAvg	3.9	4.3	4.0	4.1	0.0
FedProx	4.1	4.4	4.1	4.2	0.0
FedNova	4.0	4.5	4.2	4.3	0.0
HARNESS	5.8	6.3	5.9	6.1	15.2
MedicalQA	6.2	6.7	6.5	6.8	78.4
HealthFog	4.5	5.1	4.7	4.9	8.7
GraphCare	5.5	6.0	5.7	6.3	23.1
HEAL-AI	8.9	9.1	9.2	8.7	94.3
vs. MedicalQA	+43.5%	+35.8%	+41.5%	+27.9%	+20.3%

The bold and italicized values represent the superior interpretability and transparency of HEAL-AI as judged by clinical experts.

Table 5 demonstrates HEAL-AI's exceptional explainability capabilities, with clinical experts rating its explanations significantly higher than competing methods. The 43.5% improvement in overall explainability score compared to MedicalQA stems from the integration of multiple explanation modalities: symbolic reasoning paths provide logical justification, hierarchical attention weights highlight relevant data features, and counterfactual analysis shows how different inputs would change predictions.

The high rule coverage (94.3%) indicates that most predictions are backed by explicit symbolic reasoning, providing clinicians with interpretable decision pathways that align with medical training and clinical guidelines. This transparency is crucial for clinical adoption, as physicians need to understand and validate AI recommendations before incorporating them into patient care decisions.

Explainability was evaluated by five board-certified clinicians using a predefined rubric assessing relevance, clarity, and clinical consistency. Evaluations were conducted in a blinded manner with respect to model identity. Inter-rater reliability was assessed using intra-class correlation coefficients, yielding ICC = 0.82, indicating strong agreement. Representative explanation examples are provided in Figure 7.

OPEN IN VIEWER

Figure 7.

HEAL-AI clinical explainability and interpretability analysis.

Figure 7 illustrates HEAL-AI's clinical explainability and interpretability performance. The left panel presents a multi-dimensional evaluation of explainability metrics assessed by clinical experts, including overall explainability score, clinical relevance, decision confidence, and attention coherence across 200 test cases. HEAL-AI consistently achieves ratings in the “Excellent” range (≥8.5), significantly outperforming baseline models. The right panel demonstrates symbolic reasoning capability through explicit rule coverage analysis, where HEAL-AI attains 94.3% rule coverage. This high coverage indicates that the majority of model predictions are supported by interpretable, logic-based reasoning aligned with established clinical guidelines, thereby enhancing transparency, trust, and clinical usability.

The left panel presents a multidimensional explainability of clinical metrics—explainability metrics such as decision confidence and attention span evaluated by experts over 200 test cases. HEAL-AI significantly outperforms all baselines, achieving scores in the “Excellent” range (≥8.5) across all dimensions. The right panel shows symbolic reasoning capability via explicit rule coverage analysis. HEAL-AI attains 94.3% rule coverage, classified as high, demonstrating superior interpretability through logical justifications. In contrast, most baselines exhibit limited or no symbolic reasoning, reinforcing HEAL-AI's advantage in explainable and trustworthy medical AI.

Privacy preservation and security analysis

The privacy and security evaluation in this study assumes a semi-honest (honest-but-curious) adversarial model, which is commonly adopted in federated healthcare learning scenarios. The adversary is assumed to have access to global model parameters and aggregated model updates exchanged during federated training, but no direct access to raw patient data stored at participating institutions. Membership inference, attribute inference, and model inversion attacks were implemented following established evaluation protocols, with attack success rate and reconstruction similarity used as quantitative leakage metrics. Attackers were assumed to possess auxiliary population-level statistics and model architecture knowledge, but not individual-level ground-truth labels or direct access to local training data. This threat model reflects realistic risks faced by federated healthcare systems deployed across semi-trusted institutions.

Table 6 compares centralized, federated, edge-based, and neuro-symbolic healthcare AI models under explicit privacy and security threat models, including membership inference, attribute inference, and model inversion attacks. Privacy leakage (%) reports the proportion of sensitive information recoverable by an adversary, where lower values indicate stronger protection. Attribute inference success rate (%) measures an attacker's ability to infer private patient attributes from shared updates. Model inversion similarity score quantifies reconstruction fidelity between recovered and original patient features. Utility (AUC) reflects predictive performance retained under privacy constraints, while communication overhead (%) captures the additional cost introduced by privacy-preserving mechanisms. HEAL-AI results without differential privacy (no DP) are included to isolate the effect of privacy mechanisms relative to standard FL baselines.

OPEN IN VIEWER

Table 6.

Comprehensive privacy protection analysis under multiple attack scenarios.

Method	Privacy leakage (%, lower better)	Attribute inference success rate (%)	Model inversion similarity score	Utility (AUC)	Communication overhead (%)
CDL	76.3	68.4	0.847	0.821	—
FedAvg	62.4	54.7	0.623	0.835	0.0
FedProx	59.8	52.1	0.598	0.842	2.3
FedNova	58.6	50.9	0.587	0.846	3.1
HARNESS	54.2	47.8	0.542	0.866	5.7
MedicalQA	48.7	43.2	0.489	0.854	4.2
HealthFog	51.3	45.6	0.521	0.840	3.8
GraphCare	52.8	46.3	0.534	0.871	6.2
HEAL-AI (no DP)	45.2	39.8	0.441	0.912	7.5
HEAL-AI ( ε = 10.0 )	31.2	28.6	0.298	0.901	8.1
HEAL-AI ( ε = 5.0 )	18.7	17.4	0.182	0.893	8.7
HEAL-AI ( ε = 1.0 )	4.3	4.1	0.041	0.852	9.6
HEAL-AI ( ε = 0.5 )	1.8	1.6	0.019	0.821	11.2

Table 6 demonstrates HEAL-AI's exceptional privacy protection capabilities across multiple attack vectors. With differential privacy at ε = 1.0 , HEAL-AI achieves only 4.3% privacy leakage compared to 76.3% for the non-private centralized approach, representing a 94.4% reduction in privacy risk. This protection comes with acceptable utility degradation, maintaining 93.4% of the original model's AUC (0.852 vs. 0.912).

The protection against attribute inference attacks is particularly important in healthcare, where attackers might attempt to infer sensitive patient characteristics such as disease status, genetic predispositions, or behavioral patterns. HEAL-AI's 4.1% success rate for attribute inference attacks at ε = 1.0 provides strong protection against such privacy violations.

The mathematical foundation for this privacy protection is based on the differential privacy guarantee:

P [ A ( D ) ∈ S ] P [ A ( D ′ ) ∈ S ] ≤ e ε

(37)

where A is the privacy mechanism that adds calibrated noise to model updates, D and D ′ are neighboring datasets differing by one patient record, and S is any possible output set. This guarantee ensures that the probability of any specific output changes by at most a factor of e ε when any individual patient's data are added or removed from the training set. For ε = 1.0 , this provides a strong privacy guarantee while maintaining acceptable utility.

Figure 8 illustrates the fundamental privacy–utility trade-off that healthcare organizations must navigate when deploying AI systems.

OPEN IN VIEWER

Figure 8.

Privacy–utility trade-off analysis with clinical acceptability regions.

Figure 8 identifies the optimal operating point for healthcare applications where privacy protection and clinical utility are both acceptable. The analysis reveals that ε = 1.0 provides the best balance, achieving strong privacy protection (4.3% leakage) while maintaining clinically acceptable accuracy (AUC = 0.852). Values below ε = 0.5 provide excellent privacy but may compromise clinical safety due to reduced accuracy, while values above ε = 5.0 may not provide sufficient privacy protection for sensitive healthcare data.

While the reported differential privacy guarantees are mathematically valid at each federated training round, cumulative privacy loss over long-term and continuous deployment scenarios must be carefully considered. In realistic healthcare settings, FL models are often updated over many rounds, leading to privacy budget consumption governed by advanced composition effects. HEAL-AI explicitly tracks privacy budget usage across rounds using differential privacy composition principles; however, privacy degradation over prolonged training horizons remains a practical challenge that warrants further investigation. The reported 94.4% reduction in privacy risk is intended to demonstrate relative improvement over centralized non-private training rather than to claim superiority over all privacy-preserving FL approaches. Direct comparisons with alternative privacy-aware federated methods, such as secure-aggregation-only FL and adaptive clipping-based differential privacy frameworks under long-term training conditions, are left as an important direction for future work.

Comprehensive ablation studies

To understand the contribution of individual components to HEAL-AI's overall performance, we conducted extensive ablation studies across multiple performance dimensions. Table 7 reports the impact of systematically removing or isolating major architectural components of HEAL-AI, including neuro-symbolic integration, hierarchical attention, FL, edge intelligence, differential privacy, adaptive offloading, and knowledge distillation. Diagnostic accuracy (%) and early-detection time (hours) quantify clinical performance, explainability score (1–10) reflects clinician-rated interpretability, privacy leakage (%) measures vulnerability under inference attacks, and energy consumption (mJ) captures edge-device efficiency. Comparisons highlight explicit performance trade-offs, demonstrating that HEAL-AI's full configuration achieves balanced gains across clinical accuracy, interpretability, privacy preservation, and energy efficiency, while partial or isolated configurations degrade one or more dimensions.

OPEN IN VIEWER

Table 7.

Comprehensive ablation study analyzing the contribution of individual HEAL-AI components across accuracy, early-detection capability, explainability, privacy protection, and energy efficiency.

Configuration	Diagnostic accuracy (%)	Early-detection time (h)	Explainability score (1–10)	Privacy leakage (%)	Energy (mJ)
Full HEAL-AI	91.2	8.5	8.9	4.3	92.1
w/o neuro-symbolic integration	85.6	6.1	4.2	4.8	89.7
w/o hierarchical attention	87.3	5.8	6.7	4.1	91.3
w/o federated learning	83.9	7.2	8.5	76.3	88.4
w/o edge intelligence	90.8	5.3	8.8	4.2	205.8
w/o differential privacy	92.6	8.7	9.0	45.2	91.8
w/o adaptive offloading	90.9	6.8	8.7	4.1	156.3
w/o knowledge distillation	91.0	8.3	8.6	4.5	284.1
Only neural component	80.5	4.9	3.5	5.2	85.2
Only symbolic component	72.3	3.8	7.8	2.1	45.7
Traditional attention only	86.1	5.4	5.9	4.6	90.8
Static computation distribution	90.6	7.1	8.7	4.4	142.6

Table 7 reveals several critical insights about HEAL-AI's architectural design. The neuro-symbolic integration contributes 6.1% to diagnostic accuracy and dramatically improves explainability (from 4.2 to 8.9), demonstrating the value of combining pattern recognition with logical reasoning. The hierarchical attention mechanism contributes 4.3% to accuracy and significantly enhances early-detection capabilities, highlighting its role in identifying subtle clinical patterns.

FL's contribution is particularly evident in the privacy dimension, where its removal increases privacy leakage from 4.3% to 76.3%. This massive increase demonstrates that FL provides inherent privacy protection even before differential privacy mechanisms are applied. The edge intelligence framework primarily impacts energy efficiency, reducing consumption by 123.7% (from 205.8 to 92.1 mJ) through optimized computation distribution and model compression.

The comparison between neural-only and symbolic-only components reveals their complementary nature: the neural component excels at pattern recognition but provides limited explainability, while the symbolic component offers excellent interpretability but lower accuracy. Their integration achieves the best of both approaches.

Real-world deployment case studies

To validate HEAL-AI's effectiveness in practical healthcare settings, we conducted three comprehensive case studies that represent common deployment scenarios and clinical use cases.

Case study 1: Early sepsis detection in critical care

We conducted a retrospective analysis using 347 confirmed sepsis cases from the PhysioNet 2019 dataset, comparing HEAL-AI's performance against standard clinical protocols and existing early-warning systems. The analysis focused on the critical time window before sepsis onset when interventions are most effective.

Clinical performance: HEAL-AI identified early signs of sepsis with a median lead time of 5.9 h before clinical diagnosis (95% CI: 4.2–7.8 h), significantly outperforming the current standard of care, which typically identifies sepsis 1.2 h before clinical recognition. The system achieved the following:

Sensitivity: 89.5% (95% CI: 85.8–92.7%)

Specificity: 92.3% (95% CI: 89.1–94.8%)

Positive predictive value: 87.2% (95% CI: 83.4–90.3%)

Negative predictive value: 94.1% (95% CI: 91.6–96.1%)

Clinical value analysis: The early-detection capability translated to meaningful clinical benefits. Using established clinical outcome models, the 4.7-h average improvement in detection time could potentially achieve mortality reduction calculated as

Mortality Reduction = 1 − ( 1.076 ) − Δ t = 1 − ( 1.076 ) − 4.7 = 30.2 %

(38)

where Δ t = 4.7  h represents the improvement in detection time and 7.6% per hour represents the established mortality increase per hour of delayed sepsis treatment based on clinical literature. This calculation demonstrates the potential life-saving impact of early-detection capabilities.

Explainability assessment: Clinical expert evaluation of HEAL-AI's explanations revealed high clinical relevance:

Actionability: 8.7/10 (explanations directly supported clinical decision-making)

Clinical consistency: 9.1/10 (explanations aligned with established sepsis pathophysiology)

Trust and confidence: 8.9/10 (clinicians expressed confidence in AI recommendations)

The hierarchical attention mechanism consistently highlighted clinically relevant patterns including subtle heart rate variability changes, early lactate elevation trends, and temperature instability patterns that preceded more obvious clinical manifestations.

Discussion

The comprehensive evaluation of HEAL-AI highlights its transformative potential in shaping the future of intelligent, secure, and equitable healthcare systems. One of the core findings is the effectiveness of combining FL with neuro-symbolic AI, which results in a synergistic improvement in both model interpretability and diagnostic performance. Compared to purely neural models, HEAL-AI achieved a 4.6% improvement in diagnostic accuracy and a substantial 43.5% increase in explainability, addressing long-standing challenges in making clinical AI decisions more transparent and trustworthy. Additionally, the integration of a hierarchical contextual attention mechanism significantly enhanced early sepsis detection by 40.5%, demonstrating the vital role of personalized, context-aware reasoning in real-time clinical decision-making. From a deployment standpoint, HEAL-AI's efficient edge-cloud architecture delivered a 32.6% reduction in latency and energy consumption, proving its viability in resource-constrained environments such as rural clinics or in-home patient monitoring. This capability promotes healthcare equity by extending advanced AI-powered diagnostic tools to underserved populations. Furthermore, the privacy-preserving framework employed by HEAL-AI, leveraging differential privacy and secure aggregation techniques, achieved strong privacy guarantees with only a 4.3% leakage rate at ε = 1.0, establishing a practical and ethical foundation for federated AI deployment in sensitive healthcare environments. Despite these advancements, several limitations and emerging challenges were identified. The reliance on manually curated symbolic rules introduces a knowledge maintenance bottleneck, as updating and scaling medical reasoning rules to match rapidly evolving clinical knowledge remains labor-intensive. HEAL-AI's distributed training framework also demands significant computational, storage, and communication resources, which may hinder its scalability across institutions with limited infrastructure. Variations in data quality across institutions further impact the robustness of the global model, highlighting the need for automatic data quality scoring mechanisms to appropriately weight contributions from different clients. Regulatory and ethical considerations also pose nontrivial challenges, particularly with compliance to evolving standards such as the Food and Drug Administration’s (FDA) Software as a Medical Device (SaMD), GDPR, and national data governance laws, as well as addressing fairness, bias mitigation, and clinician oversight in decision-making. Additionally, real-world clinical validation through prospective trials is critical to fully understanding the system's impact on patient outcomes, clinician trust, and integration into clinical workflows. 22Future research will focus on overcoming these limitations through a series of targeted innovations. Automated knowledge integration using natural language processing will help update and expand symbolic reasoning modules by mining clinical guidelines, research literature, and observational data. Causal FL methods will be explored to estimate treatment effects in a privacy-preserving way, enabling personalized medicine at scale. Training healthcare-specific multimodal foundation models that can handle diverse data types—such as imaging, genomic sequences, clinical notes, and wearable sensor data—will further enhance the system's capabilities while preserving interpretability. Privacy will be further strengthened through the exploration of quantum-secure techniques, including quantum encryption and key distribution, to future-proof FL against emerging threats. To promote meaningful clinician engagement, future iterations of HEAL-AI will incorporate human–AI collaboration frameworks, enabling clinicians to interpret, validate, and refine AI decisions via intuitive explanation interfaces and feedback mechanisms. Additionally, blockchain-backed infrastructure using smart contracts can help enforce data quality, ensure transparent auditability, and manage incentive structures in federated training environments. Finally, designing lightweight, offline-capable, and culturally adaptive versions of HEAL-AI will be key to extending its reach into low-resource global health settings. Collectively, these directions aim to evolve HEAL-AI into a clinically validated, globally deployable, and ethically grounded AI framework that supports sustainable and trustworthy digital transformation in healthcare.

Recent research highlights emerging directions that can further inform the evolution of intelligent healthcare systems. Policy-driven digital health interventions demonstrate how regulatory frameworks, environmental factors, and governance mechanisms influence the effectiveness of AI-enabled healthcare solutions, reinforcing the importance of privacy compliance and policy awareness in system design. ³² In parallel, recent ensemble-based cardiovascular prediction models, including hybrid ensemble learning and bagging–boosting approaches, show improved robustness and generalization for cardiovascular risk assessment.^33,34 These advances suggest future research opportunities for extending HEAL-AI by incorporating policy-aware optimization and ensemble learning strategies within its federated neuro-symbolic architecture, particularly for chronic disease monitoring and cardiovascular decision support.

While HEAL-AI integrates multiple advanced components spanning FL, neuro-symbolic reasoning, edge intelligence, and privacy-preserving mechanisms, this work focuses on architectural integration and empirical validation rather than exhaustive optimization of each individual sub-domain. The development and experimentation were conducted using modular implementations built upon established open-source frameworks, with domain-specific components validated independently prior to system integration. At the time of submission, the source code and trained models are undergoing institutional review for release due to healthcare data governance constraints. The authors commit to releasing a reproducible codebase, model configurations, and experimental protocols in a public repository upon acceptance to support transparency and independent verification.