Table 8.
Qualitative comparison of smart healthcare-related works.
| Use Case | Ref | Contribution | AI Role (At the Edge) |
AI Algorithm | Dataset | AI Placement | Employed Technology |
Platform | Metrics | Benefits AI-Edge |
Drawbacks | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Smart healthcare | DHM | [80] | Food recognition | Classification Storage | DRCNN | Food 101Image | Smartphone | Quantization, GPU accelerator | TensorFlow Lite | Accuracy loss values, computational power | Low response time | Loss of accuracy over time |
| [81] | Food recognition | Classification preprocessing | GoogLeNet | UEC-256 UEC-100 Food-101 | Smartphone | Pretrained CNN | Caffe | Response time, accuracy, computational power | Low response time | Loss of accuracy over time | ||
| AAL | [82] | Accurate and timely fall detection | Classification | LSTM/GRU | SisFall dataset | IoT, gateway (fog) | Virtualization | Docker HDFS-Apache Kafka-MongoDB Tensorflow | Accuracy, sensitivity, precision, inference | Scalability, flexibility | Memory consumption needs to be optimized Mobility not considered | |
| [83] | Online/offline monitoring elderly patients suffering from chronic disease | Prediction | NB-FA | Vital signs, behavioral data environmental data | Cloud, edge | Transfer learning | Weka, classifier, Spark job | Accuracy, sensitivity, precision, inference time | Accurate, fault-tolerant, fast decisions | High computational cost | ||
| [84] | Real-time fall detection | Preprocessing, prediction | LDA KNN SVM | SisFall datasets | Raspberry Pi 3 B + | Real-time test | Low-cost model | Response time | High accuracy, low response time | Accuracy and generalization still improved | ||
| [85] | Multimodal fall detection | Prediction | PCA linear regression MLP | SisFall data set | Mist, fog, cloud, edge | Not mentioned | Low-cost model | CC, MAE RMSE, RAE, RRSE response time | High accuracy, less inference time | Generalization needs to be solved | ||
| [86] | Real-time in-home health monitoring | Prediction | GCAE | MobiAct dataset | Cloud, edge | Federated learning | Not mentioned | Accuracy communication rounds scalability | Heterogeneity of data and communication cost solved | Data privacy issues | ||
| Smart healthcare | HAR | [87] | Real-time abnormal human activities | Prediction | PCA -CNN | UniMiB DATASET | Edge device | Transfer learning | Python 3.6 | Process time | Low energy consumption, less computational cost | Lack of security |
| [88] | Real-time, human activity recognition | Prediction | DRNN | WISDM dataset | Raspberry Pi3 (edge devices) | Virtualization | TensorFlow | Accuracy F1-score recognition time | Less recognition time, high accuracy | High computational cost | ||
| [89] | Energy-efficient, human-activity recognition | Training, prediction | CNN | Opportunity dataset, w-HAR dataset | Edge devices | Transfer learning | Not mentioned | Accuracy, precision, recall, weighted, F1-score | Less memory overhead, high accuracy | Stability not tested | ||
| [90] | Human activity recognition | classification | SVM | KTH Dataset Hollywood2 Action Dataset | Edge/cloud | Transfer learning Blockchain | TensorFlow | Accuracy | High accuracy multiclass classification | Less scalability | ||
| [91] | Multiaccess physical monitoring system | Classification | BDN | Real-world dataset | Wearable IoT | Transfer learning | Not mentioned | Accuracy data transmission time RMSE | Less energy consumption, high accuracy | Lack of data privacy, less scalability | ||
| [92] | Physical instance-based irregularity recognition | Classification | CNN LSTM | NTU RGB dataset | Fog nodes | Transfer learning | Python-Pillow, OpenCV, Numpy libraries | Rate of latency analysis | High accuracy, less latency | Environmental changes and model generalization not considered | ||
| Smart healthcare | LDP | [93] | Monitoring and predicting COVID-19 outspread | Prediction visualisation | FCM T-RNN SOM | - | Fog nodes | MATLAB-Ifogsim | Preprocessing | Latency time, response delay, accuracy, precision | reliability, high accuracy | Lack of security |
| [94] | Location-aware monitoring and preventing encephalitis | Prediction visualisation | FCM- T-RNN, SOM | Cloud, edge | UCI-repository data | Preprocessing | MATLAB | Latency time, response delay, accuracy, precision | Reliability, high accuracy, location aware, data management | Lack of security | ||
| [95] | Early detection of Kyasanur forest disease and control the disease outbreak | Classification | ANN | KFD dataset | Fog/cloud | Lightweight model | Not mentioned | Accuracy, sensitivity, specificity, RMSE MAE | High accuracy | High computational cost | ||
| [96] | Continuous monitoring and early detection of mosquito-borne disease | Classification | FNN, SNA graph | UCI-repository data | Fog node | Lightweight model | Not mentioned | Accuracy, sensitivity, specificity | High accuracy | Data integrity and security not considered | ||
| [97] | Automatic diagnosis of COVID-19 | Classification | K-MEANS -VGG16 | X-ray ultrasound datasets | Edge devices | Pretrained model | TensorFlow | RMSE, MAE | Cope with data heterogeneity | Less accuracy, lack of security | ||
| [99] | Remote COVID-19 diagnosis | classification | RF GAN GNB | Generated dataset | Fog nodes | Open-source language R iFogSim | Accuracy response time, recall | High accuracy | High energy consumption, lack of security | |||
| Smart healthcare | LDP | [99] | Remote COVID-19 diagnosis | Classification | Mobile-Net V2 | Chest CT scan image dataset | Transfer learning | Edge devices | TensorFlow | Sensitivity specificity precision F1-score | High accuracy, less response | Not tested for large datasets, accuracy needs to be improved |
| [100] | Low delay in prediction of health status of COVID-19 patients | Preprocessing prediction | eRF | COVID-19 dataset | Edge devices | Lightweight model | TensorFlow | Training time, accuracy, precision, recall, MAE, RMSE | High accuracy | High computational cost | ||
| DD | [101] | Early lung cancer diagnosis | Preprocessing, feature selection Classification | FCM, CS, SVM | (ELCAP) dataset | Fog nodes | Lightweight model | MATLAB 2013a | Accuracy, sensitivity, specificity, MCC, F-measure, ROC curves, computational cost | Less training time, high accuracy | High cost of model for fog implementation | |
| [102] | Intelligent monitoring of cardiomyopathy patients | Intelligent sensing | FHHO, FL | Real-world dataset | Fog nodes | – | Not mentioned | Execution time, accuracy, precision, recall, F-measure | High accuracy, low time cost | Lack of security, high energy consumption | ||
| [103] | Real-time monitoring patients with chronic diseases | Classification | NB-WOA | Clinical dataset, Physio Bank-MIMIC II database | Fog nodes, cloud | Transfer learning | Weka, Spark | Accuracy, recall, precision | Higher accuracy, high response time | High complexity of model, lack of security | ||
| [104] | Early heart disease prediction | data fusion prediction | CFS, KRF | UCI repository data | Fog nodes | Lightweight model | – | Accuracy, training time, scalability | Scalability, accuracy | Quality of the data depends on the number of sensors, improved accuracy is required | ||
| Smart healthcare | DD | [105] | Early detection of Parkinson’s Disease | Prediction | ANFIS GWO PSO | UCI University of California | Fog nodes | Distributed computing | TensorFlow | RMSE, MAE | High accuracy | Lack of security |
| [106] | Diabetic cardio disease prediction | Prediction | Rule-based clustering, CRA, ANFIS | (Heart disease, diabetes) dataset | Edge devices | Blockchain | Java | Purity NMI accuracy execution time | Efficient grouping medical data, high accuracy, secure data sharing, good training with uncertainty | Low accuracy | ||
| [107] | Remote cardiac patient monitoring | Classification | 1D-CNN | MIT-BIH Arrhythmia | Fog nodes (single-board computer), cloud | Transfer learning | Not mentioned | RMSE MAE CPU usage accuracy loss recall precision F1-score | High accuracy, low computational overhead, low resource usage, low response time | Scalability not considered | ||
| [108] | Timely disease diagnosis of health conditions | Data preprocessing classification | AE HMWWO | UCI-repository data | Edge devices | Lightweight model | Not mentioned | Latency, F-measure time complexity sensitivity | High sensitivity, improved accuracy Minimum time complexity and latency scalability | Small dataset used for evaluation, lack of data protection | ||
| [109] | Real-time physiological parameter detection | Preprocessing prediction, load balancing | RK-PCA HMM MoSHO SpikQ-Net | UCI repository data | Edge devices, fog nodes | Lightweight model | iFogSim | Execution, time accuracy, latency | Stability, scalability, low execution, time, low latency, low complexity | Lack of security | ||
| [110] | Real-time blood glucose | Prediction | GRU | (OhioT1DM ABC4D ARISES) datasets | Edge device (Smartphone) | Hardware accelerator | TensorFlow Lite | RMSE, MSE | Low energy consumption, good training with uncertainty | Less sensitivity |