Source-linked AI summary
HealthFog: An Ensemble Deep Learning based Smart Healthcare System for Automatic Diagnosis of Heart Diseases in Integrated IoT and Fog Computing Environments
Shreshth Tuli, Nipam Basumatary, Sukhpal Singh Gill, Mohsen Kahani, Rajesh Chand Arya, Gurpreet Singh Wander, Rajkumar Buyya
TL;DR
HealthFog addresses the gap between accurate deep-learning diagnosis and the low-latency requirements of IoT healthcare, including heterogeneous heart-patient data volumes. It integrates ensemble deep learning with edge-fog-cloud resources through FogBus, and evaluates configurable operation modes across accuracy and quality-of-service metrics. The paper reports a lightweight automatic heart-disease diagnosis service, while identifying file-based input and separate worker-node training as current boundaries.
Problem
Existing healthcare-fog models do not adequately integrate deep-learning accuracy with edge-level low latency while handling both little and big heart-patient data.
Method
HealthFog deploys ensemble deep learning across integrated IoT, edge, fog, and cloud resources using FogBus for automatic heart-disease diagnosis.
Results
HealthFog provides a lightweight automatic heart-patient diagnosis system evaluated with FogBus across accuracy, response time, network bandwidth, energy consumption, latency, jitter, and execution time.
Takeaways & Limitations
The configurable fog service supports different fog-computation scenarios and user requirements by prioritizing quality of service or prediction accuracy.
Takeaways & Limitations
HealthFog currently uses file-based input data and trains separate models at each worker node before combining them with bagging.
Abstract
from arXiv · showhide
Cloud computing provides resources over the Internet and allows a plethora of applications to be deployed to provide services for different industries. The major bottleneck being faced currently in these cloud frameworks is their limited scalability and hence inability to cater to the requirements of centralized Internet of Things (IoT) based compute environments. The main reason for this is that latency-sensitive applications like health monitoring and surveillance systems now require computation over large amounts of data (Big Data) transferred to centralized database and from database to cloud data centers which leads to drop in performance of such systems. The new paradigms of fog and edge computing provide innovative solutions by bringing resources closer to the user and provide low latency and energy-efficient solutions for data processing compared to cloud domains. Still, the current fog models have many limitations and focus from a limited perspective on either accuracy of results or reduced response time but not both. We proposed a novel framework called HealthFog for integrating ensemble deep learning in Edge computing devices and deployed it for a real-life application of automatic Heart Disease analysis. HealthFog delivers healthcare as a fog service using IoT devices and efficiently manages the data of heart patients, which comes as user requests. Fog-enabled cloud framework, FogBus is used to deploy and test the performance of the proposed model in terms of power consumption, network bandwidth, latency, jitter, accuracy and execution time. HealthFog is configurable to various operation modes that provide the best Quality of Service or prediction accuracy, as required, in diverse fog computation scenarios and for different user requirements.
1. Introduction
HealthFog addresses the need for heart-disease diagnosis that combines deep-learning accuracy with fog-enabled low latency and efficient IoT data handling. It proposes an ensemble deep-learning healthcare service deployed across integrated IoT, edge, fog, and cloud resources.
- Cloud computing can be unsuitable for real-time healthcare because centralized processing introduces high time delay.
- Heart problems are difficult to detect early, while limited doctor availability constrains conventional diagnosis.
- Heart-patient data includes both little data processed at fog nodes and big data processed at cloud data centers, reaching 250 MB per minute or more.
- The proposed model integrates deep-learning accuracy with edge computing's low latency for real-time healthcare services.
- HealthFog is a lightweight fog service for automatic heart-disease diagnosis using deep learning, IoT devices, and FogBus.
- The paper proposes a generic architecture and lightweight ensemble deep-learning system for deploying heart-patient diagnosis on fog computing resources.
2. Related Work
Related work shows progress in fog-based healthcare monitoring, security, scheduling, and prediction, but existing systems often optimize only selected metrics or lack integrated ensemble deep learning. HealthFog targets simultaneous real-time processing, resource efficiency, and heart-disease diagnosis.
- Several related systems use specialized approaches, including ECG monitoring, graph-based representation learning, FHIR-based prediction, and cloud-based smart-home sensing.
- The related-work comparison is presented in Table 1 as a comparison of existing models with HealthFog.
- Prior healthcare systems address monitoring, scheduling, security, or prediction, but their evaluations commonly consider only limited performance dimensions.
- HealthFog extends edge healthcare architectures by using distributed deep-learning models in an ensemble to improve prediction accuracy for critical heart patients.
- Existing work identifies low response time, low energy consumption, resource utilization, and real-time heart-disease diagnosis as unresolved healthcare-fog challenges.
3. Background Technologies
FogBus supports integrated Fog-Cloud application deployment and connects IoT sensors with gateway, broker, and worker nodes. Aneka supports distributed cloud application development and dynamic acquisition of virtual resources.
- FogBus provides structured communication and platform-independent execution for integrated Fog-Cloud environments.
- FogBus connects healthcare sensors to gateways and fog workers, while broker nodes manage resources and task initiation.
- FogBus uses blockchain, authentication, and encryption to support data integrity, privacy, security, reliability, and robustness.
- Aneka provides APIs for developing and deploying distributed applications on cloud virtual resources.
- Aneka's dynamic provisioning acquires virtual-machine resources from Infrastructure-as-a-Service cloud providers and integrates them into existing systems.
4. System Architecture
HealthFog integrates Edge-Fog-Cloud hardware and software components to process heart-patient data and deliver fast, accurate healthcare results. Its resource manager selects broker, worker, or cloud processing, while ensemble and preprocessing modules support diagnosis.
- Architecture: HealthFog manages heart-patient data and diagnoses disease severity through structured Edge-Fog-Cloud integration.The architecture integrates diverse hardware instruments through software components for fast and accurate result delivery.
- Hardware components: Body Area Sensor Networks collect medical, activity, and environmental measurements and transfer patient data to gateway devices.Medical sensors include ECG, EEG, EMG, oxygen, temperature, respiration, and glucose sensors.
- Resource management: The broker receives gateway requests, secures communications, and arbitrates task placement among broker, worker, and cloud resources.Worker nodes process allocated tasks and may host deep learning, filtering, analytics, and storage components.
- Cloud integration: HealthFog uses Cloud Data Centers for heavy-load or location-independent processing when fog resources are insufficient or tasks are latency tolerant.Cloud resources provide greater availability for heavy workloads.
- Software components: Preprocessing filters, reduces, encrypts, and extracts key data components before automated diagnosis and database storage for retraining.The pipeline uses PCA with SPIHT and SVD-based encryption to prepare health feature vectors.
- Software components: The ensembling module combines predictions from different models by voting, while FogBus supports persistent communication among edge devices.Gateway devices can send data to the broker, another worker, or the cloud according to arbitration results.
5. HealthFog Design
HealthFog preprocesses heart-patient features and trains an ensemble deep neural network for binary disease classification. At diagnosis time, distributed workers combine predictions to provide a majority class, with ensemble use configurable for latency-critical requests.
- Application workflow: The system receives sensor data and returns whether a patient has heart disease, together with confidence in the claim.Its application includes preprocessing, ensemble deep learning, and a gateway interface.
- Data preprocessing: Heart-patient inputs require domain-specific preprocessing, including normalization of skewed age and blood-pressure distributions.The source data includes graphical pulse-oximeter or ECG measurements that must be converted into model features.
- Model training: The ensemble deep neural network is trained on labeled Cleveland Dataset data and then predicts binary classes for real-time inputs.The trained model maps heart-patient data to disease or no-disease predictions.
- Model training: Training, validation, and testing use a 70:10:20 split, with validation tuning the model and testing assessing performance on new data.Trained models can be stored across processing-capable nodes through a common database.
- Ensemble diagnosis: During diagnosis, a 13-feature patient vector is multicast to workers, whose predictions are combined by bagging into a majority class.Users can disable ensemble processing when latency is critical.
6. Implementation
The implementation uses Python preprocessing and ensemble components with SciKit-learn's BaggingClassifier to train multiple deep neural networks and aggregate their predicted classes.
- Software implementation: Python implements the preprocessing and ensemble deep learning components, while preprocessing normalizes fields using dataset maxima, minima, and distributions.Normalization is based on the observed scale and distribution of each field parameter.
- Ensemble implementation: SciKit-learn's BaggingClassifier randomly distributes training data among deep neural-network classifiers and outputs the majority predicted class.The model accepts the base classifier type and number of classifiers as inputs.
- Model configuration: The deep neural network uses 13 input features and 2 output classes for binary heart-disease classification.The architecture contains three hidden layers with fully connected layers of 20, 20, and 10 nodes.
- Model configuration: ReLU is used as the activation function in the implemented neural network.The model parameters specify the activation function alongside layer sizes and class outputs.
- Result delivery: Worker or broker results are ensembled and forwarded to the Android gateway, with module interactions represented in Figure 7.The gateway receives the final result after prediction aggregation.
7. Performance Evaluation
The authors deployed HealthFog on a real FogBus-based fog framework to evaluate heart-disease diagnosis across accuracy, response-time, network, and energy dimensions. The evaluation used heterogeneous gateway, broker, worker, and cloud hardware with dedicated monitoring tools.
- Evaluation setup: HealthFog was implemented and deployed on actual fog devices using FogBus for a real-world heart-problem detection application.The evaluation examines accuracy and response times alongside network and energy overheads.
- Hardware configuration: The gateway device was a Samsung Galaxy S7 running Android 9.It served as the mobile gateway in the deployed system.
- Hardware configuration: The broker/master node used a Dell XPS 13 with an Intel i5-7200 CPU, 8 GB DDR4 RAM, Windows 10, and Apache HTTP Server 2.4.34.This node hosted the master-side deployment components.
- Hardware configuration: Worker nodes used Raspberry Pi 3B+ devices with quad-core ARM Cortex-A53 CPUs, 1 GB LPDDR2 SDRAM, Wi-Fi, Raspbian Stretch, and Apache HTTP Server 2.4.34.The setup represents embedded edge processing hardware.
- Hardware configuration: The public-cloud component used a Microsoft Azure B1s virtual machine with 1 vCPU, 1 GB RAM, 2 GB SSD, and Windows Server 2016.The cloud VM provided the public-cloud processing environment.
- Measurement: Performance data came from Microsoft Performance Monitor, NMON, Microsoft Network Monitor 3.4, and vnStat across master, cloud, broker, and Raspberry Pi nodes.The tools measured system parameters and network-bandwidth consumption.
7.2. Dataset
HealthFog is evaluated on Cleveland heart-patient data to predict whether a patient has heart disease. Experiments examine dataset partitioning, accuracy, and how edge-node count affects training and test performance.
- Dataset: The Cleveland database encodes heart-disease diagnosis as target value 0 for absent disease and 1 for present disease.The dataset includes patient parameters such as major vessels and thalassemia status.
- Dataset: The dataset was split into 70% training, 10% validation, and 20% testing subsets.Validation was used for model-parameter adjustment, while the final subset measured performance.
- Experiments: Accuracy was evaluated on 1,807 examples, using 1,355 for training and 452 for testing.Training examples were distributed equally across worker and broker nodes, with comparisons across edge-node counts and ensemble settings.
- Accuracy: Training accuracy increases as the number of edge nodes increases because each node receives fewer samples and can overfit them across multiple epochs.Figure 9 reports training accuracy as a function of broker-plus-worker edge nodes.
- Accuracy: Test accuracy decreases with more edge nodes because smaller training subsets reduce the model’s ability to generalize.Figure 10 examines test accuracy as the number of edge nodes changes.
7.5. Prediction Confidence
HealthFog assigns confidence to binary heart-disease predictions from the separation between the two class probabilities. Confidence is higher for correct predictions, while low-confidence cases are flagged for medical consultation; the section also examines timing across fog configurations.
- Prediction Confidence: Confidence is computed from p0 and p1, the probabilities of no disease and heart disease, with p0 + p1 = 1 and a range of 0–100.Equal probabilities produce 0 confidence, while probabilities of (0.9, 0.1) produce 80 confidence.
- Prediction Confidence: The model’s confidence is higher on correctly classified test datapoints than on incorrectly classified datapoints.Figure 11 compares the complete test set with correctly and incorrectly predicted subsets.
- Prediction Confidence: 49.7% is the maximum confidence observed for an incorrect prediction, so confidence below 50% prompts consultation with a doctor.The paper treats predictions below this threshold as potentially unreliable.
- Timing Characteristics: Arbitration time is nearly 115 ms for direct Broker or Cloud submission and increases when additional edge workers require load checking.Figure 12 compares Broker-only, single-worker, two-worker, and Cloud configurations.
- Timing Characteristics: Latency is defined as the combination of communication time and queuing delay in the evaluated fog configurations.Figure 13 compares latency across the reported scenarios.
- Timing Characteristics: Jitter is higher in the Broker-only configuration because the Broker also performs arbitration, resource management, and security checks.Figure 14 presents jitter on a logarithmic vertical scale.
- Timing Characteristics: Ensemble prediction increases execution time because workers must determine the majority class across predictions.Cloud execution time is lowest because of greater resource availability, while Raspberry Pi workers have lower clock frequencies.
7.7. Network Bandwidth Usage Characteristics
Network and power usage increase as HealthFog uses more worker nodes. Ensemble operation has the highest network demand because input data is sent to all workers, while cloud deployment consumes more power than edge deployment.
- Network Bandwidth Usage Characteristics: Network usage increases with more worker nodes because heartbeat packets, security checks, and cloud data transfers increase.Figure 16 compares bandwidth usage across the evaluated scenarios.
- Network Bandwidth Usage Characteristics: Ensemble prediction produces the highest network bandwidth consumption because data is sent to every worker node.The ensemble configuration distributes the input across all workers.
- Power Consumption: Power consumption increases as the number of worker nodes increases.The study compares power use among the Cloud Data Center, Broker laptop, and Raspberry Pi workers.
- Power Consumption: Cloud deployment consumes much more power than edge deployment because the Cloud Data Center has substantially higher power consumption than Broker or worker nodes.Figure 17 reports power consumption across the configurations.
7.9. Analysis with Related Work
Compared with related healthcare fog models, HealthFog combines edge resources, deep-learning prediction, ensemble operation, and configurable FogBus deployments. Its recommendations trade accuracy against latency, network use, energy, and workload demands.
- Analysis with Related Work: Many related healthcare models do not use resources near the network edge and therefore have higher latency from cloud-based computation and data transfer.The comparison references Figure 13’s latency analysis.
- Analysis with Related Work: Other prior works lack integrated deep-learning models and consequently provide lower disease-detection accuracy than HealthFog’s approach.The paper identifies this limitation across several related healthcare systems.
- HealthFog Comparison: HealthFog extends FogBus with ensemble deep learning and configurable deployments spanning accuracy, response time, network usage, and power usage.The framework synchronizes fog nodes to support ensemble execution and user-specific requirements.
- Recommendations: Worker nodes are recommended for latency-critical, lightweight, or energy-constrained tasks because their proximity provides low result-delivery time.The recommendation targets applications requiring rapid responses or limited energy.
- Recommendations: Bagging should be disabled under energy or network constraints, whereas enabling it provides better accuracy when those constraints are absent.This recommendation expresses the accuracy-versus-resource trade-off for ensemble operation.
- Recommendations: The Cloud Data Center configuration is required for heavy, latency-tolerant tasks that cannot complete successfully on resource-constrained edge workers.The boundary is defined by task workload and latency tolerance.
8. Conclusions and Future Work
HealthFog integrates deep learning and IoT in edge computing devices to provide fog-based automatic heart-disease diagnosis, validated with FogBus. Future work targets cost-optimal execution, direct sensor input, improved ensemble models, and broader application domains.
- Conclusions: HealthFog integrates deep learning into edge computing devices for automatic heart-disease analysis using IoT data.The system provides healthcare as a fog service and manages heart-patient data from different IoT devices.
- Future Work: Future work proposes cost-optimal execution under different QoS characteristics and fog-cloud cost models.
- Future Work: HealthFog currently uses file-based input, with future integration planned for direct sensor data acquisition.
- Future Work: Future extensions include more intelligent ensemble models, other fog applications, and healthcare domains such as diabetes, cancer, and hepatitis.
- Availability: The HealthFog implementation, experiment scripts, and results were released as open-source software.