This study is reported in accordance with the PRISMA 2020 statement for systematic reviews. A review protocol was not registered; however, the eligibility criteria, search strategy, and data-extraction items were specified a priori to support reproducibility.

We searched peer-reviewed literature published between 1 January 2020 and 31 May 2025 in IEEE Xplore, ScienceDirect (Elsevier), SpringerLink, MDPI, Hindawi/Wiley, PubMed, and Google Scholar. The final search was completed in May 2025. We also screened the reference lists of included articles to identify additional eligible studies within the same time window.

Search queries were constructed using controlled keywords and free-text terms and then adapted to the syntax of each database/publisher platform. The core Boolean expression was:

(“Internet of Medical Things” OR IoMT OR “medical IoT”) AND (fog OR edge OR “fog computing” OR “edge computing” OR cloud) AND (“chronic disease” OR diabetes OR “cardiovascular disease” OR “heart disease” OR arrhythmia OR stroke) AND (“machine learning” OR “deep learning” OR AI) AND (latency OR jitter OR throughput OR “quality of service” OR QoS OR energy OR power).

Search results were exported (where supported), de-duplicated, and then screened against the eligibility criteria.

Inclusion criteria:

Primary research articles proposing an IoMT, Fog, Cloud, or hybrid IoMT–Fog–Cloud architecture for chronic disease diagnosis or prognosis.

Figure 1 summarizes the selection process following PRISMA 2020. The search identified 42 records. After removing 12 records before screening (2 non-English and 10 outside 2020–2025), 30 records were screened. Five records were excluded because they did not use machine/deep learning. Twenty-five reports were sought for retrieval; six could not be retrieved. Nineteen reports were assessed for eligibility, and five were excluded (one conceptual framework and four without experimental evaluation). Fourteen studies were included in the final review.

Fourteen primary studies met the eligibility criteria. Of these, six implemented an explicit three-tier IoMT–Fog–Cloud (or close equivalent) architecture and reported at least one network/QoS metric; these six form the core set for the joint performance–QoS synthesis (Tables 3,4). To reflect recent algorithmic developments in multi-disease AI-IoMT, we also included two representative models—DACL and TasLA—that report strong diagnostic performance but provide limited network/QoS evaluation. Accordingly, they are compared mainly on AI design choices and diagnostic metrics (Table 5) and discussed in terms of likely deployment constraints. The six IoMT–Fog–Cloud frameworks were prioritized because they (1) instantiate a three-tier architecture, (2) report both diagnostic and network/QoS metrics (enabling trade-off analysis), and (3) provide architectural diversity (sensing modalities, Fog task allocation, and evaluation environments) with sufficient reporting detail for metric extraction. When several eligible studies addressed the same disease task, we favored papers with more complete metric reporting and higher citation impact. DACL and TasLA were selected because they introduce attention/fusion and optimization strategies that are increasingly used in medical AI and are plausible candidates for future IoMT deployments, even when QoS metrics are not yet routinely reported.

From each included study we extracted (i) diagnostic performance outcomes and (ii) system/network QoS outcomes when reported. Extracted items included disease type, dataset source, sample size, number of features, AI model(s), architectural placement (IoMT/edge, Fog, Cloud), validation strategy (when described), and the reported metrics. Because the studies differ substantially in datasets, validation protocols, and network-evaluation methods, we synthesized results narratively rather than conducting a meta-analysis. For consistency, we defined 14 key metrics: six diagnostic metrics (accuracy, precision, recall, F1-score, sensitivity, specificity) and eight system/network metrics (latency, jitter, throughput, bandwidth utilization, processing/execution time, network usage, energy consumption, power consumption).

Performance index:

Accuracy, F1-score, precision, recall, sensitivity, and specificity.

To support cautious interpretation, we recorded potential sources of optimistic diagnostic performance during data extraction, including small or highly curated datasets, class imbalance, reliance on benchmark or synthetic data, unclear patient-level train–test separation, and possible leakage/overfitting—particularly in simulation-based Fog/Cloud evaluations. These considerations are discussed in Section 4.2.

The three diabetes-oriented frameworks target either type 2 diabetes mellitus or gestational diabetes prediction. Pati et al. (1) propose an IoMT–Fog–Cloud architecture in which wearable sensors and mobile devices send data to Fog nodes that perform pre-processing and deep learning-based classification, with additional benchmarking models deployed in the Cloud. Their ANN with bagging and voting ensembles achieves high accuracy and F1-score while exploring multiple network configurations (single master, master-plus-multi-Fog, and Cloud-only), revealing clear latency and energy trade-offs (1).

El-Rashidy et al. (2) propose a Fog–Cloud IoMT framework for gestational diabetes prediction using EHR data (MIMIC-III) and an explainable deep neural network. The paper reports strong performance (ACC ≈ 0.957) and millisecond-scale latency in the Fog–Cloud deployment; however, the article was later retracted (21 Nov 2024), so its quantitative results should be interpreted cautiously and are not used as primary evidence for conclusions.

Yıldırım et al. (3) propose a Fog–Cloud architecture for type 2 diabetes monitoring using data from East Jakarta health clinics and public repositories. Their framework relies on fuzzy logic and machine learning classifiers, achieving moderate accuracy but providing detailed measurements of processing time, latency, jitter, and energy consumption across different network configurations (3).

Table 3 summarizes the key characteristics of these three diabetes frameworks, including sample size, number of features, AI models, diagnostic performance, and network behavior (1–3).

Sun et al. (4) introduce FogMed, a Fog-based framework for streaming ECG data and prognostic analysis. A stacked LSTM model operates at the Fog layer to detect atrial fibrillation from long-term recordings while the Cloud stores and aggregates data. This framework demonstrates that deep recurrent models can be deployed near the edge for real-time arrhythmia monitoring (4).

Pati et al. (5) extend their earlier diabetes architecture to cardiovascular disease. They use an IoMT–Fog–Cloud framework where Fog nodes host a deep neural network, combined with bagging and voting strategies, to classify coronary heart disease from multiple UCI datasets. The authors report strong accuracy and F1-score, while network simulations quantify latency, jitter, throughput, and power consumption under different Fog node counts (5).

Angel Nancy et al. (6) present a Fog-based smart cardiovascular disease prediction system that leverages a gated recurrent unit-based recurrent neural network (GNU-RNN). Their design explores large-scale synthetic and real datasets, achieving very high diagnostic performance (near-perfect accuracy and specificity) while maintaining low end-to-end latency when Fog nodes are used as intermediate processing hubs (6).

Table 3 also includes these cardiovascular systems to enable cross-disease comparison of sensor types, AI models, and network metrics (4–6).

To extend the comparison beyond disease-specific IoMT–Fog–Cloud frameworks, we reviewed two recent multi-disease AI-IoMT architectures and several related optimization-oriented studies. The DACL framework proposed by Nandagopal et al. (7) uses a three-tier AI-IoMT design: wearable sensors collect patient data; a deep auto-encoder imputes and normalizes missing values; Golden Flower Optimization performs feature selection; and a Collaborative Bias Integrated GAN (ColBGaN) performs multi-disease classification (e.g., diabetes, heart disease, stroke). Across the datasets evaluated, DACL reports lower classification error and improved accuracy compared with baseline machine learning and deep learning models (7).

TasLA, introduced by Khadidos et al. (8), presents a Tasmanian and Lichtenberg optimized attention-based deep convolution data-fusion model for IoMT smart healthcare. Experiments on benchmark healthcare datasets show higher classification accuracy and fewer false predictions than competing deep learning models, highlighting the importance of optimization and attention mechanisms in multi-source IoMT data fusion (8).

Complementary work by Selvarajan et al. (9) on excretory-organism recognition and by Kumar et al. (10) on IoMT-enabled arrhythmia classification illustrates the use of deep learning and optimization for other medical pattern-recognition tasks. Selvarajan (11) surveys modern optimization techniques relevant to engineering and AI-IoMT systems. Nigar et al. (12) present an IoMT-to-Cloud chronic disease diagnosis pipeline, while Xhaferra et al. (13) and Suleiman et al. (14) report recent IoMT–Fog–Cloud models for diabetes and heart-disease prediction, respectively.

Table 5 summarizes the high-level architecture, disease coverage, optimization strategies, and main reported benefits of DACL and TasLA (7, 8).

Overall, the included studies point to recurring architectural patterns that explain trade-offs between diagnostic performance and network/QoS behavior, and they help clarify why some disease domains report QoS metrics more consistently than others.

Across diabetes-oriented IoMT–Fog–Cloud frameworks, reported diagnostic performance generally lies in the high-80s to mid-90s. Ensemble strategies and explainable deep models can improve accuracy and F1-score while maintaining low-latency inference at the Fog layer (1–3). Importantly, several studies rely on public benchmark datasets (e.g., UCI) or mixed public–clinical sources, which can simplify preprocessing and reduce data noise compared with real-world clinical streams.

Cardiovascular and arrhythmia frameworks often report higher peak performance. FogMed demonstrates the feasibility of real-time atrial fibrillation detection from continuous ECG streams using Fog-layer LSTM processing (4). Other heart-disease frameworks report very high (sometimes near-perfect) accuracy/precision/recall under optimized configurations (5, 6). These results highlight the promise of Fog-assisted inference for time-sensitive conditions, but they must be interpreted in the context of dataset characteristics and validation methodology (Section 4.2).

Recent multi-disease AI-IoMT frameworks such as DACL and TasLA extend beyond single-disease prediction by combining deep feature learning, meta-heuristic optimization, and data-fusion strategies across heterogeneous inputs (7, 8). However, these studies typically emphasise algorithmic performance on curated datasets and provide limited reporting of IoMT network behavior (latency, jitter, energy) under realistic deployment constraints (7–11).

Several included studies report extremely high diagnostic metrics (>99% accuracy or near-perfect precision/recall) in specific experimental settings (6). While these values are reported as stated in the source publications, cross-study comparison of point metrics can be misleading when datasets, class balance, and validation protocols differ. Additionally, one gestational-diabetes framework in our initial extraction was later retracted (2), underscoring the importance of ongoing evidence appraisal when conducting technology reviews.

Potential sources of optimistic performance include: (1) small datasets or limited patient diversity, (2) substantial class imbalance, (3) reliance on benchmark or partially synthetic datasets, and (4) incomplete reporting of data-splitting procedures (e.g., patient-level separation) and hyperparameter tuning. In Fog/Cloud simulation studies, additional risks include unintended data leakage (feature normalisation or selection performed before splitting) and overfitting to a single network configuration.

To make these limitations more explicit, Tables 3,5 now report dataset source, validation strategy (as described by the original authors), and deployment realism (simulation vs. real-world). For future work, we recommend reporting patient-level splitting, nested cross-validation or external validation on independent cohorts, and calibration/uncertainty measures alongside accuracy and F1-score to better reflect clinical generalizability.

The IoMT–Fog–Cloud frameworks provide a rich set of network-level results, but direct comparability is limited because latency, jitter, throughput, and energy are often defined and measured differently across studies. Some papers report end-to-end latency, others report processing time at a specific layer, and several omit the measurement point, workload, or simulation/testbed configuration.

Despite these limitations, a consistent architectural trend emerges: moving inference closer to the patient (edge/Fog) reduces end-to-end delay and bandwidth demand, but may increase energy/power consumption and system complexity, especially when multiple Fog nodes are deployed to improve responsiveness (1–6). By contrast, Cloud-only approaches can simplify maintenance and centralise analytics but typically increase network delay and may be unsuitable for time-critical events (e.g., arrhythmia alarms).

To clarify definitions and measurement context, Table 4 summarizes the QoS metrics reported in each IoMT–Fog–Cloud study and notes, where available, the measurement point and evaluation environment. Future studies should specify metric definitions (e.g., end-to-end vs. layer-specific), system layer, workload, and testbed/simulator details to support fair benchmarking.

To make the design guidance more actionable, Table 2 maps each recommendation to evidence extracted from the reviewed implementations (e.g., Fog/Cloud task allocations and architectural choices associated with lower latency or reduced energy in Tables 3–5). Table 2 also lists practical engineering options—such as lightweight model design and deployment-oriented optimization (feature selection, attention, and training strategies reported in DACL and TasLA 7, 8)—for Fog/Edge execution. Finally, Table 2 includes a minimal reporting checklist (dataset provenance, validation strategy, and QoS definitions/measurement context) intended to improve reproducibility and support more reliable cross-paper comparisons.

Although the reviewed engineering frameworks demonstrate promising diagnostic and QoS performance, clinical readiness requires additional evidence beyond retrospective accuracy metrics. Key translational considerations include prospective validation in diverse patient cohorts, robustness to sensor noise and missing data, interpretability/traceability of model decisions, cybersecurity and privacy-by-design (e.g., encrypted transmission and secure model updates), and regulatory requirements for software as a medical device (SaMD) and connected devices.

Looking ahead, rapid advances in data integration and high-dimensional biomedical measurement may inform future IoMT analytics pipelines. For example, recent spatial multi-omic profiling and neuroinflammation mapping technologies highlight how multi-modal data fusion can uncover clinically relevant patterns at high resolution (15). Spatially resolved perturbation screens similarly illustrate scalable approaches to linking complex biological signals with outcomes (16). While these methods are not IoMT deployments, they reinforce the importance of rigorous, transparent data pipelines and may inspire future hybrid sensing frameworks that combine remote monitoring with complementary multi-modal clinical data sources.