Artificial intelligence-enhanced CSI-based Wi-Fi sensing for non-contact vital sign monitoring: a systematic review

Overview of health monitoring systems

Health monitoring systems have revolutionized healthcare by offering continuous methods to track vital signs and detect potential health issues. These systems range from wearable devices to ambient sensors, providing real-time data on various physiological parameters. Traditional hospital-based methods typically rely on multiple wired sensors, which restrict patient mobility and comfort (Zhang et al., 2023). These systems often require manual checks, increasing staff workload. Moreover, subtle patient movements can easily affect these devices, leading to inaccurate readings and multiple false alarms (Liu et al., 2021).

To address the limitations of traditional devices, recent advancements in healthcare monitoring have embraced Wi-Fi-based sensing technologies, which show promising results for non-contact vital signs monitoring. For instance, Wang, Yang & Mao (2020b) demonstrated that Wi-Fi can accurately monitor breathing and heart rates. However, challenges remain in data privacy, real-world accuracy, and healthcare workflow integration. Therefore, addressing these issues is crucial for the widespread adoption of advanced health monitoring systems.

CSI and its role in health monitoring

When a transmitter (TX) and a receiver (RX) communicate wirelessly through Wi-Fi, various environmental factors, including walls, windows, obstacles, and even human bodies, can affect the signals. Even the smallest change in surroundings can alter how signals propagate in the medium. CSI essentially describes the characteristics of the communication medium between the transmitter and receiver, providing detailed insights into movements and changes in the environment with each packet. CSI sensitivity to micro movements makes it a perfect fit for sensing vital signs such as respiration rate and heart rate. For instance, subtle chest movements during the breathing cycle—inhaling and exhaling—affect signal propagation, and are therefore captured by CSI. Thus, Wi-Fi signals are a promising tool for non-invasive health monitoring. By analyzing subtle changes in Wi-Fi signal patterns caused by human presence and movement, researchers have developed innovative methods for detecting vital signs and activities. CSI extracted from Wi-Fi signals has been successfully used to monitor respiratory and heart rates. For instance, Zhang et al. (2023) achieved 96.887% accuracy for estimating breathing rate and 94.708% accuracy for heart rate using CSI data. Similarly, Zeng et al. (2019) demonstrated respiration monitoring with 98% detection rate at distances up to 9 m.

More advanced applications include sleep monitoring and apnea detection. Yu et al. (2021) developed Wi-Fi sleep, achieving 81.8% accuracy in four-stage sleep classification, while Yang et al. (2021) created a system for detecting sleep apnea with over 95% accuracy.

Wi-Fi-based health monitoring offers several advantages, such as:

• Non-invasive and contactless monitoring, ideal for long-term patient observation.

• Utilizing existing Wi-Fi infrastructure, reducing implementation costs.

• Ability to operate through walls and furniture, allowing for discreet monitoring.

• High accuracy in readings and predictions, aiding early intervention in critical cases.

• Potential for muti-person monitoring, as demonstrated by Gao et al. (2020) for respiration tracking.

However, researchers have yet to fully address challenges such as sensitivity and environmental changes. Ongoing research focuses on improving accuracy of readings and expanding the range of detectable health parameters to enable better health assessments and early disease detection.

Wi-Fi CSI overview

Research questions

The aim of the research questions is to identify a well-established plan to retrieve articles restricted only from the areas of consideration. Table 1 outlines the research questions identified in this article.

Search strategy

To ensure a comprehensive and inclusive search of the literature, we developed a multi-layered keyword-based strategy using Boolean logic across six major databases: IEEE Xplore, ScienceDirect, SpringerLink, ACM Digital Library, Google Scholar, and ResearchGate. We focused on keyword searches in titles, abstracts, and keywords, which is appropriate given the topic’s engineering and computer science orientation. While PubMed and MeSH indexing are highly valuable for biomedical and public health research, many studies on Wi-Fi-based sensing and signal-level data processing are not indexed there. Therefore, we prioritized repositories that more comprehensively cover technical research. At the same time, we recognize the clinical relevance of our topic and the importance of including healthcare-related work. Although PubMed was not part of our main search strategy due to its limited coverage of wireless sensing technologies, we sought to compensate by using broader repositories such as Google Scholar and ResearchGate, which include research across both technical and medical disciplines. We also incorporated clinically related keywords, such as “heart rate,” “respiration rate,” “health monitoring,” and “patient monitoring,” to capture articles that bridge both fields. This approach helped us address the overlap between clinical applications and signal-level AI techniques. More details about the search logic, keywords used, and database selection are provided in Appendix.

Articles selection criteria

Choosing the right search terms is an important step in conducting a thorough and wide-ranging systematic review. In this study, we carefully selected search terms that included key concepts such as Wi-Fi signals and CSI. The goal was to strike a balance, not too generic to avoid unrelated articles and not too narrow to not miss related articles (Cooper et al., 2018). The primary keywords were derived from the research questions and refined through several iterations. Searches were then conducted using the six major databases mentioned earlier. A full list of search strings, keyword combinations, and databases searched as well as the overall search process are outlined in the Appendix.

Screening of articles

The results of the individual searches returned a total of 8,573 publication records: (IEEEXplore = 5,473, Springer Link = 250, ScienceDirect = 1,412, Google Scholar = 312, ACM Digital Library = 576, ResearchGate = 550). Following title and abstract screening, and the application of our inclusion/exclusion criteria (Table S3), 45 studies were included in the final analysis. A comprehensive description of the screening and exclusion process is available in the Appendix, along with figures and tables showing study selection by database, timeline, and final inclusion outcomes.

Taxonomy for study categorization and quality assessment

• Taxonomy table: to facilitate comparative analysis, with classification dimensions (see Table 2).

• Sensing technique: indicates if the study was relying on extracting CSI amplitude, phase, or both.

• Use case: describes the health parameter being monitored by the study, such as respiration rate, heart rate, and sleep apnea.

• AI integration level: specifies the degree of AI involvement, ranging from no AI integration to traditional machine learning or deep learning approaches.

• Evaluation: defines the complexity of the environment in which the system was tested, including the number of people considered and whether the system was running in real-time.

• Metrics reported: lists the metrics used to evaluate the system, including Mean Absolute Error (MAE), accuracy, recall, precision, specificity, and F1-score.

Table 3 assesses the quality of each study using seven criteria. Sample size defines the size of the dataset used. Validation method reflects the process used to train and test the data in AI studies, and how vital signs ground truth was obtained in non-AI articles. Benchmarking indicates whether the study compared its performance to previous work. Dataset, code, and algorithms availability, notes whether any of these resources are publicly accessible or included in the article. Reproducibility assesses how replicable the system’s performance is, considering factors such as dataset availability and the sufficiency of the data provided. Measurement method specifies the device, sensor, or method used to identify the ground truth data. Finally, overall quality ranks the study into one of three levels based on the information provided.

Data extraction and analysis

After the final primary studies that met the quality assessment standards were identified, data were extracted to help answer the research questions. These studies were organized into two comparative tables: one focusing on AI-based healthcare monitoring using CSI, and the other on deterministic CSI-based healthcare monitoring systems. In both tables, the first three columns summarize the essential details of each publication.

In both tables, the “Estimation Method” column refers to the fundamental algorithms or techniques used for to evaluate the data. The “Preprocessing Phase” column identifies the first steps that data undergoes before being fed into the system. The approach used to ensure that the credibility of study results is presented under “Validation Strategy.” The “Dataset” column details the source of the data, while the “Results” column summarizes the main conclusions of the research. The “Limitations” column points to the restrictions or difficulties encountered during the study. This table structure facilitates a systematic comparison of various strategies in CSI-based monitoring across various applications.

CSI-based healthcare monitoring (AI-based approach)

Tables 4 and 5 present a comparative overview on how researchers have leveraged Wi-Fi CSI to detect and measure crucial physiological parameters such as respiration and heart rates. Table 4 contain articles utilizing AI-based methods and Table 5 includes those employing non-AI techniques. With this division, we can observe the evaluation of CSI-based vital sign monitoring from traditional signal processing approaches to more advanced machine learning algorithms. By examining these studies, we gain insights into the accuracy of non-invasive, contactless health monitoring systems and can compare the results achieved using both non-AI and AI methods.

Armenta-Garcia et al. (2021) proposed BReML, a system that combines Fast Fourier Transform (FFT) method for estimating breathing rate with machine learning and compares three models: Support Vector Machines (SVM), K-Nearest Neighbor (K-NN), and Quadratic Discriminant Classifier (QDC). BReML filters CSI amplitude using Hampel, Savitzky Golay (SG), and Butterworth band-pass filters. Rather than using all subcarriers, it selects the 10 most sensitive ones to reduce computational power. FFT is then applied to extract additional features for the AI models, including an initial estimation of the breathing rate. Then Discrete Wavelet Transform (DWT) is also applied to generate extra features, which are fed into multiple classification models for comparison. The classification consists of five classes covering 12 to 20 breaths per minute (brpm) in steps of two. The proposed system achieved high accuracy, with 99.09%, 99.21%, and 99.65% for K-NN, SVM, and QDC, respectively. However, the system was not tested on subjects outside the training data and did not estimate the heart rate.

To extend the research in Armenta-Garcia et al. (2021) and estimate both breathing and heart rate, Armenta-Garcia et al. (2023) proposed a system following similar concept of Armenta-Garcia et al. (2021). However, they collected more data to train the system for heart and breathing rate estimation and made both the dataset and the system publicly available for further research. As for the class division, they divided respiration rate into five classes: <=10, 11–13, 14–16, 17–19, and >=20 with only three classes for heart rate: <= 60, 61–79, and >= 80. Nevertheless, the proposed system achieved an accuracy of 99.18%, 90.71%, and 91.35% using K-NN, SVM, and QDC, respectively, for breathing rate. On the other hand, the system did not perform well when estimating the heart rate with accuracy of 56.55%, 55.06%, and 55% for K-NN, SVM, and QDC, respectively. Additionally, the heart rate classification used only three broad categories with wide ranges, making it unsuitable for continuous health monitoring use.

Mosleh et al. (2022) explored employing two off-the-shelf routers in their experimental setup. They adopted a machine learning approach by introducing an innovative algorithm, BreatheSmart, which comprises seven layers that leverage the LSTM network. BreatheSmart processes both the phase and amplitude information of each subcarrier and subsequently assigns the data to one of the predefined respiration pattern classes. Data are acquired from three transmitting and three receiving antennas, utilizing all 56 subcarriers within the 2.4 GHz network spectrum. The results revealed BreatheSmart’s ability to correctly predict respiration pattern classification with up to 99.97% accuracy. However, the evaluation was conducted under controlled conditions, with significant hardware setup, limiting its applicability in diverse scenarios. Additionally, BreatheSmart does not provide specific measurements for respiration rate, because it categorizes data into a limited number of nine distinct classes.

To summarize, the articles in Table 4 all leverage AI classifiers to assign respiration rate to one of the predefined classes. Reference (Mosleh et al., 2022) stands out with an impressive accuracy of 99.97% using an LSTM-based model. Unlike other studies, it uses both CSI amplitude and phase for estimating the breathing rate, rather than relying on one only. In addition, it is the only study that does not use conventional filters, such as Hampel filter, to remove noise and outliers, relying only on FFT. However, it is noteworthy that the system was evaluated in a controlled environment with a breathing simulator, which yields constant patterns unlike human respiration. Moreover, the nine classes had varying steps, with some classes much further apart and no intermediate class. Both Armenta-Garcia et al. (2021) and Armenta-Garcia et al. (2023) used Hampel, Savitzky Golay, and band-pass filters and experimented with SVM, K-NN, and QDC. Despite this, Armenta-Garcia et al. (2021) achieved higher accuracy of 99.65% with QDC. Nevertheless, the study did not address heart rate estimation, whereas (Armenta-Garcia et al., 2023) attempted to estimate pulse rate but achieved poor results, with only 55.65% accuracy using K-NN.

Wi-Fi signal-based monitoring systems using a deterministic approach

Guan et al. (2023) advances takes breathing estimation in proposing a new method to estimate multi-person breathing rates simultaneously. The proposed system utilizes the variance of CSI amplitude to determine the most sensitive subcarrier across different antennas. FFT is then applied on the CSI phase to produce a frequency spectrum. To distinguish between different individuals, the study utilizes Density-Based Spatial Clustering to separate each person’s data, so that each cluster represents one individual. Thereby, the system finds the breathing rate of each person by identifying the maximum frequency within each cluster. The system demonstrated good performance, with a 97.04% accuracy; however, the accuracy drops significantly as the distance between the targets and transceivers increases. In addition, the study did not consider two individuals with the same breathing rate, so it categorizes them as one person.

Gui et al. (2023) explored a setup in which a router served as the TX and a PC equipped with three antennas functioned as the RX. They introduced a system designed to monitor respiration rate and detect sleep-related activities utilizing Wi-Fi CSI data. The system uses the peak-to-peak interval method to calculate respiration rates in the absence of movement, deriving the rate as 60 divided by the average of peak-to-peak intervals. To enhance accuracy, Gui et al. (2023) proposed a new dynamic subcarrier selection method and peak identification technique. Upon detecting an activity, the system suspends respiration rate monitoring to identify specific activities such as lying on the spine, prone, or in lateral positions. Upon settling into a new position, the system recalibrates by repeating the subcarrier selection process to ensure optimal selection for each posture. Experimental results demonstrated promising performance, with estimation errors of 0.13 and 0.16 breaths per minute for supine and prone sleeping postures, respectively. However, while the proposed system exhibited robust performance, it did not operate in real time and did consider multiple persons.

Guo et al. (2021) proposed a system for emergency detection based on three vectors: position, behavior, and breathing condition. Position refers to the patient’s location within the household. Behavior encompasses the patient’s movements, such as falling. Breathing indicates whether respiration is abnormal, including apnea, dyspnea, or tachypnea. Position and behavior data are derived from keypoint maps generated via a neural network algorithm trained using camera data sourced from the OpenPose library. Breathing condition, notably apnea detection, is determined using filtered CSI data via the DWT method. The system leverages both phase and amplitude components of the CSI data, as they offer complementary information. To construct the keypoint map, data are collected from two receivers positioned in different locations to ensure comprehensive coverage. Patient movement is required for accurate keypoint map generation, as the method relies on the dynamic component of the CSI. In tests involving the same volunteer and within the same room, the system achieved impressive results, accurately matching 80.7% of points exactly to the original keypoint maps, with 99.3% showing close alignment. Furthermore, apnea detection accuracy reached 91.2%. However, when tested with different volunteers who were not part of the neural network training or when obstructed by a wall, accuracy dropped significantly.

Overall, the articles in Table 5 relied on manual estimation of vital signs, implemented in one of three ways. First, the FFT method, as followed by Gu et al. (2019), Guan et al. (2023), Zhang et al. (2022), selects the frequency corresponding to the highest peak within the normal respiration or heart rate range. Second, peak counting over a specific interval, used by Guo et al. (2021), Ali et al. (2023), relies on identifying inhalation and exhalation phases, with each peak representing one breathing cycle. Third, the peak-to-peak interval method, conceptually similar to peak counting, calculates the average interval between consecutive peaks (time from one breath to the next). Dividing 60 by this average interval yields the number of breaths per minute. This method, followed by Gui et al. (2023), Wang, Yang & Mao (2020c), Zhang et al. (2022), has achieved the most success in estimating breathing rate; for instance, Gui et al. (2023) reported an average estimation error of 0.13 breaths per minute (brpm) in the supine position.

In addition to using common filters such as Hampel and Butterworth low-pass filters, Gui et al. (2023) also proposed a new subcarrier selection method based on the relative difference using FFT. They also introduced additional steps to remove false peaks. However, the system did not estimate heart rate or operate in real time, limiting its efficiency for real life use. Furthermore, Gui et al. (2023) did not account for multiple people breathing in the same room, unlike (Guan et al., 2023), which applied a clustering algorithm after FFT to group data corresponding to each individual. Although (Guan et al., 2023) achieved high accuracy of 97.04%, the system could misclassify multiple people with the same breathing rates as one person.

Regarding heart rate estimation, both Gu et al. (2019) and Zhang et al. (2022) went for FFT-based method for estimating the heart rate. Both studies used Hampel filter and achieved an accuracy of 94% even though (Gu et al., 2019) used Butterworth band-pass filter and only amplitude while (Zhang et al., 2022) used DWT and considered both phase and amplitude. Still, Zhang et al. (2022) managed to achieve a slightly better result of 94.3%.

Principal findings

Across the included studies, approximately 70% were classified as having moderate quality, primarily due to testing with a sufficient sample size and using reliable methods to obtain ground truth, such as the Neulog Respiration Monitor Belt logger sensor (NUL-236), rather than relying on error-prone methods like manual counting. High-quality studies also increased reproducibility by using public datasets, sharing their dataset or code, and comparing their systems’ performance to previous studies. In contrast, low-quality studies typically did not compare their results to previous studies, did not report their sample size or used minimal samples, or employed less accurate methods for measuring ground truth. None of the articles attempted to validate performance against publicly available datasets. Non-AI studies more often compared their system performance with older studies; over 55% of non-AI articles did so, compared to only 25% of AI articles. Out of the 45 reviewed articles, only four shared their dataset, and just one shared the code as well. More than 75% of the studies did not disclose algorithms, code, or data, highlighting the need for stronger community standards around openness in CSI-based vital signs monitoring.

As presented in Table 2, the taxonomy of studies in CSI-based vital signs sensing reveals key trends and observations that highlight the current state of the field and provide insight for future of research in this field. In terms of the sensing techniques, using both amplitude and phase is the most common approach, especially in AI-based studies monitoring respiration rate, as it provides a holistic view and enables AI to analyze complex patterns in both amplitude and phase simultaneously. Deterministic approaches more commonly use amplitude alone for estimating heart rate and respiration rate and phase achieved similar accuracies of 94.5% and 94.3%, respectively. Conversely, an AI-based study using both amplitude and phase showed higher accuracy (98.5%) for heart rate estimation. Other applications, such as predicting sleep stage and sleep apnea, also utilized both amplitude and phase. Studies predicting heart rate and breathing rate with AI only investigated a single-person context, whereas deterministic studies researched various contexts, including multi-person, single-person, and single-person in real time. Regarding reported metrics, only four deterministic articles reported both MAE and accuracy; the remainder reported either MAE or Accuracy only. Among AI studies, only one did not report accuracy, while 41.66% reported accuracy, recall, precision, and F1-score, with the rest reporting either accuracy, MAE, or both. Figure 2 displays a graphical summary of the categorization of studies.

Our literature survey on AI-based health monitoring systems using Wi-Fi signal shows considerable improvements in accuracy, real-time operation, and multi-person monitoring. Heart rate estimation has achieved accuracies of up to 96% (Sun, Bian & Li, 2024), while breathing rate estimation has reached very high precision, with errors of 0.13 brpm (Gui et al., 2023). Multi-person tracking systems have successfully monitored up to four people simultaneously, achieving accuracies above 86% (Zeng et al., 2019). Subsequent developments in AI approaches, including deep learning models (Yu et al., 2021; Kumar et al., 2022; Mosleh et al., 2022), have further fine-tuned overall system performance, enabling highly sophisticated health assessments and increasing robustness across diverse environments. These advancements pave the way for less invasive, more user-friendly monitoring systems, with the possibility for use in sleep monitoring (Yu et al., 2021) and early detection of respiratory problems (Yang et al., 2021). Taken together, these findings suggest an optimistic trend toward developing efficient, individualized health monitoring systems that could radically impact both hospital and outpatient care.

With these results in mind, we present a comprehensive summary of the studies leveraging AI techniques for CSI-based vital signs monitoring. Table 6 offers an in-depth analysis of the AI algorithms employed, the results achieved, the datasets used, and the unique contribution of each study. The table highlights the diverse approaches applied to vital signs monitoring using CSI, with the “Contribution” column showcasing novel aspects such as new methodologies, multi-person tracking, and real-time capabilities. This detailed overview helps us better understand the progress made in applying AI to Wi-Fi-based vital signs monitoring and identifies areas for future research. Figure 3 illustrates the most commonly used AI models in Wi-Fi vital signs monitoring.

Table 7 presents an AI technique comparison framework for Wi-Fi-based health monitoring, summarizing all models from the surveyed articles and highlighting their respective strengths and weaknesses.

To synthesize the reported performance across different use cases, Table 8 summarizes the minimum, maximum, and average values of the reported metrics for each use case. Heart rate estimation had the lowest accuracy when first explored through CSI, at 56.55%. However, more recent studies have achieved accuracies up to 94.71%, with an average of 84.22% across five studies. Respiration rate detection demonstrated better overall performance, with a minimum accuracy of 78.6%, a maximum of 99.65%, and the highest average among all use cases at 95.7%. Fewer studies investigated sleep apnea, sleep stage, and respiration cycle detection, reporting average accuracies of 93.4%, 81.8%, and a MAE of 0.27 s, respectively.

To visually illustrate the relationship between healthcare use cases and the evaluation metrics reported in the studies, we created a snake diagram connecting the use cases to their corresponding metrics, as shown in Fig. 4. Studies monitoring respiration rate are strongly associated with accuracy and MAE, with 25 and 20 studies reporting these metrics, respectively. However, metrics such as precision, recall, and specificity appear less frequently. For heart rate estimation, accuracy remains the most commonly reported metric, while MAE and the other matrixes are less frequently included. Among the smaller use cases, articles reporting respiratory cycles are unique in reporting only MAE in seconds, whereas the remaining use cases report accuracy, F1-score, Precision, and Recall.

Comparison to prior work

Several recent studies have examined Wi-Fi-based sensing for healthcare applications, but they differ significantly from this work in scope, methodology, and focus. For instance, Soto et al. (2022) present a survey on vital signs monitoring using Wi-Fi CSI, reviewing approximately 34 articles (not explicitly listed) without following a systematic protocol such as PRISMA. Their inclusion criteria and article selection process are not clearly defined, and the survey does not cover articles that use AI approaches to enhance estimation. Similarly, Miao et al. (2025) introduce a general overview of Wi-Fi sensing approaches for human activity recognition. Their review is not PRISMA-compliant, includes no database search methodology, and focuses more on wireless signal processing and sensing theory than healthcare-specific applications. Ahmad, Ullah & Choi (2024) offer a narrative explanation of contactless Wi-Fi sensing in healthcare, listing emerging trends and application scenarios rather than systematically reviewing recent methodologies or implementation results. Their work does not distinguish between CSI-based and RSS/radar-based systems, nor does it classify studies by performance, signal type, or clinical feasibility.

In contrast, our systematic review strictly follows PRISMA and Kitchenham standards, with clearly defined inclusion and exclusion criteria, screening procedures, and documented reasons for study inclusion. We conducted a systematic multi-database search (IEEE Xplore, ScienceDirect, SpringerLink, ACM Digital Library, ResearchGate, and Google Scholar), downloading 8,573 articles, from which 45 primary studies were selected following rigorous screening. This review focuses on CSI-based heart and respiratory rate monitoring in particular, contrasting AI-based and non-AI-based methods while emphasizing their reliability, performance, and feasibility for clinical or home settings. The included studies, published between 2019 to 2024, provide a contemporary synthesis of the field. Additionally, we present structured tabular summaries, performance metrics, setup requirements, and limitations, allowing healthcare professionals and engineers to interpret and apply the findings. This level of systematic depth and clinically relevant focus is currently absent in existing literature, highlighting the novelty and value of our work. Table 9 provides a comparison between our systematic review and prior works.

Methodological limitations and reproducibility concerns

A number of methodological concerns were identified across surveyed literature, which restrict the ability to compare, reproduce, and reliably interpret results in Wi-Fi-based health monitoring studies.

• •

  Statistical significance and robustness

  Only a small proportion of studies reported formal measures of statistical significance, such as p-values, confidence intervals, or variance across repeated runs. In most cases, results were presented as single-point performance data without any statistical analysis, making it difficult to determine whether observed differences are meaningful or a result of chance.

• •

  Replication and external validation

  Reproducibility is a major issue. As Table 3 shows, only 10 studies provided publicly available code, algorithms, or datasets, while 35 studies shared no implementation resources. Even fewer studies tested their models on independent, external datasets, relying instead on in-house or proprietary datasets that were not released publicly. This lack of transparency prevents independent validation and fair comparisons within the filed. Consequently, our search did not include clinical databases such as PubMed or Embase, which may contain reports of early-stage trials or clinical evaluations. This omission limits our ability to assess the clinical readiness or real-world deployments of CSI-based sensing.

• •

  Unpredictability of evaluation methods

  As Tables 2 and 3 show, the approaches used to measure system performance vary widely across studies.

• •

  Metrics: Accuracy was the most frequently reported metric (29 studies), followed by MAE in 23 studies coming in second. Five studies reported F1-score as well as precision and recall as primary metrics. Figure 5 illustrates the frequency of these reported metrics. A clear trend emerges: non-AI studies tend to report accuracy and MAE, while AI-based studies tend to report accuracy and F1-score, and sometimes precision and recall, to account for class imbalance and better reflect overall model performance.

• •

  Validation methods: Seven studies used k-fold cross-validation or similar resampling methods, whereas most relied on a single train-test split. Using only one split increases the risk of overfitting and can exaggerate reported performance.

• •

  Testing environments: Most systems were tested in controlled conditions with a single user. Only 10 studies directly measured performance in multi-user scenarios or more dynamic, realist environments, an important step toward measurement of real-world applicability in the healthcare setting.

• •

  Benchmarking practices

  Among the 45 studies, 22 engaged in benchmarking, directly comparing their results against findings reported in previous literature. This practice helps position new approaches within the broader research landscape and provides a relative measure of performance improvements over existing methods. However, benchmarking alone does not fully confirm a model’s generalizability, as many of these studies did not evaluate their systems on entirely independent datasets or in substantially different real-world environments. The remaining studies relied solely on internal validation, which, while useful for development, can overestimate real-world performance when not supplemented with external testing.

The lack of reliable evaluation frameworks, limited use of high-quality statistical reporting, and restricted availability of code and data combine to form a fragmented evidence base. To advance this technology toward clinical application, the research community needs to adopt uniform evaluation methods, prioritize reproducibility through open-source principles, and embrace statistical rigor when reporting performance.

Limitations and recommendations

Despite the accuracy and potential of using AI-enhanced Wi-Fi sensing non-invasive health monitoring systems, a range of technical and practical challenges remain. These limitations are interrelated and directly inform our proposed recommendations for future research.

• •

  High computational demands and edge deployment constraints

  Most AI-based solutions (e.g., Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTMs)) are computationally intensive and rely on powerful GPUs (Graphics Processing Units) or multi-core CPUs (central processing units) to analyze high-resolution CSI data in real time. This poses challenges for portable or embedded systems such as home gateways, low-power Internet of Things devices, or wearable hubs. The computational overhead affects latency, increases cost, and reduces viability in resource-limited clinical environments. Research should focus on lightweight, optimized models through pruning, quantization, knowledge distillation, and architectural simplification, enabling high performance on low-power, edge-based systems without costly hardware.

• •

  Energy efficiency limitations

  Constant CSI acquisition and high-rate sampling can be very energy-intensive, particularly for systems intended to monitor patients continuously over long intervals. This results in frequent battery charging, higher maintenance demands, and lower patient compliance in mobile or battery-based healthcare devices. It also wastes energy and bandwidth by processing data during inactivity. Adaptive sensing and event-based monitoring, where data are gathered more frequently when an anomaly is detected, providing a roadmap toward using less power without loss of monitoring capabilities of critical interest.

• •

  Integration gaps with healthcare IT systems

  Although numerous studies have demonstrated technical feasibility in controlled environments, few show interoperability with hospital electronic health record (EHR) systems, remote patient monitoring dashboards or telemedicine platforms. Without integration into existing healthcare IT workflows, these systems remain stand-alone prototypes that cannot contribute to continuous patient history or automated clinical decision support. Interoperability requires standardized data formats, secure Application Programming Interfaces (APIs), and compliance with regulations like the Health Insurance Portability and Accountability Act and General Data Protection Regulation. Using Standardized APIs, secure communication protocols, and compliance frameworks will help fill this gap so that CSI-generated health data can be transported, stored and retrieved within current healthcare infrastructure.

• •

  Security and privacy risks

  The transmission of sensitive physiological data over Wi-Fi raises risks of interception, tampering, or unauthorized access. Breaches may occur not only from weak or absent encryption but also from the passive nature of monitoring, where explicit consent may not be obtained. Security-by-design principles, such as robust encryption, authentication, secure storage, and anonymization, are essential for compliance and building patient trust.

• •

  Lack of high-quality, standardized datasets

  Most studies use small, self-generated datasets with limited demographic diversity, environmental variation, or sensing conditions. Public datasets, when available, tend to be incomplete, restricted to amplitude or phase data, or often partially pre-processed. This limits reproducibility and hinders cross-study comparison, forcing researchers to recreate data collection pipelines. Standardized, well-documented datasets would establish a common foundation for rapid innovation, reproducibility, and stringent performance testing between research teams.

• •

  Limited scope of monitored vital signs

  Current research focuses on respiration and heart rate, omitting other important measures, such as blood pressure, oxygen saturation, or arrhythmia. This narrow scope limits clinical applicability, particularly in use cases such as intensive care unit monitoring where multiple vitals must be measured simultaneously. Future work should focus on the development of algorithms and sensing methods that would be able to derive more pertinent physiological data concerning CSI data, and in that way, these systems could be upgraded into holistic health monitoring systems.

• •

  Environmental sensitivity and performance degradation

  CSI-based monitoring is strongly influenced by environmental factors such as furniture formations, wall compositions, or Wi-Fi signal fluctuations. These can degrade performance, increase false positives, or cause missed detections. Although some techniques freeze measurements during disturbances, they fail under frequent or persistent changes. Environment-adaptive calibration and robust signal processing methods are needed to guarantee stable performance with a large change in environmental conditions.

• •

  Multi-person monitoring

  Overlapping CSI patterns of multiple individuals remain a major challenge. Current methods perform poorly in multi-user or unsynchronized movement scenarios, limiting applicability in hospital wards, senior care facilities, or family homes. Research into clustering, angle-of-arrival estimation, and deep learning-based separation could help achieve credible performance in multi-person environments.

In addition to system level constraints, it is important to note the methodological limitations revealed in the literature. Inconsistencies in statistical reporting, evaluation measures, and validation procedures, combined with insufficient sharing of codes and datasets discussed in the previous section, compromise reliable cross-study comparison and replication. Addressing both technical and methodological challenges in parallel to advance Wi-Fi-based health monitoring toward standardized, clinically deployable solutions.

By explicitly linking these shortcomings to actionable recommendations, this review not only highlights what current Wi-Fi-based health monitoring systems lack but also provides a viable guide for how future research and development should proceed. Solving these problems will be central to the transition from exciting prototypes to dependable, secure, and clinically integrated systems that can function productively across diverse environments and patient populations.

Future directions and implications

Based on the outlined limitations and recommendations, there are several promising directions for the continued development of Wi-Fi-based health monitoring systems. These directions are not only oriented toward technical refinement but also toward integrating the technology into healthcare requirements, policy systems, and user acceptance. From a research perspective, investigation into the following areas is worth consideration:

• Edge-optimized architectures: Train highly efficient AI models that avoid major accuracy trade-offs while being deployable on low-power and embedded devices. Techniques such as model pruning, quantization, and on-device inference pipelines can support real-time monitoring without reliance on high-performance servers.

• Energy-aware sensing strategies: Develop adaptive sensing methods where data collection and processing rates vary according to patient activity, health status, or environmental stability. This minimizes power consumption without compromising the necessary measurements.

• Multi-person tracking: Seek efficient algorithms to differentiate concurring CSI designs in a common area. Hybrid signal separation, beamforming, and spatial filtering with AI are potential solutions for enabling multi-user monitoring in hospitals and home care.

• Increased physiological coverage: Extend monitoring beyond heart rate and respiration to include blood pressure, oxygen saturation and early detection of arrhythmia. An integration of multi-parameter monitoring and predictive analytics can be used to deliver proactive interventions in healthcare.

• Diverse and well-labeled datasets: Design standardized, diverse, and well-labeled datasets that capture different environments, demographics, and medical conditions, and make them publicly available. This would accelerate innovation, reproducibility, and fair benchmarking across research groups.

The implications of these directions are equally important from a practical and policy standpoint:

• Clinical integration and workflow adaptation: Collaboration between technologists and healthcare providers is required to ensure interoperability with EHRs, clinical dashboards, and telemedicine platforms.

• Regulatory and ethical regulation: Policymakers should establish clear deployment guidelines on deployment addressing data ownership, informed consent, and monitoring ethics, while also creating reimbursement models for remote patient monitoring.

• Public awareness and acceptance: Communication plans should ensure that patients and caregivers are informed about the benefits, limitations, and security of Wi-Fi-based monitoring, thereby leading to user acceptance and participation.

Following these directions, the field can move from isolated experimental successes toward sustainable real-world healthcare solutions that are technically dependable, ethically responsible, and socially acceptable. This two-pronged approach, innovation paired with readiness, will be crucial in making Wi-Fi-based health monitoring reach its full potential in improving patient outcomes and streamlining healthcare delivery. To conclude our synthesis, we have compiled all the keywords and search terms used during our research and created a network map showing the most frequently utilized terms in the literature. Figure 6 presents this keyword network map.

Search strategy overview

The search strategy was developed to identify peer-reviewed articles focusing on Wi-Fi CSI) and AI-based non-invasive vital signs monitoring. The search was conducted across six databases: IEEE Xplore, ScienceDirect, SpringerLink, ACM Digital Library, Google Scholar, and ResearchGate.

The overall search strategy could be outlined below:

• 1.

  Research questions formulation.

• 2.

  Develop search phrases.

• 3.

  Choose digital libraries and conducting searches.

• 4.

  Apply specific inclusion/exclusion criteria to select articles relevant to our topics of interest.

• 5.

  Gather essential details from the articles we chose.

• 6.

  Analyze the research questions in depth, allowing us to conduct a comprehensive study on the existing methods, their advantages, and the potential for future research, as well as their limitations.

Figure S1 illustrates the visual breakdown of these steps. A detailed list of the keyword combinations, Boolean query strings, and individual database inputs used during the search process is provided below.

Boolean query and keywords

To retrieve relevant articles that address the research questions from various search resources, a query string was necessary. Initially, we determined the primary search terms derived from the research questions. Subsequently, we listed synonyms and alternative expressions for these primary terms. These were then linked to each corresponding primary term using the OR operator. Finally, all primary search terms were connected using the AND operator. The culmination of these processes produced the query string that was utilized to search in different databases. The Boolean logic used is shown below:

(“Vital signs” OR “Heart Rate” OR “Respiration Rate” OR “Body Temperature” OR “Breathing” OR “Cardiac”) AND (“Wi-Fi Signals” OR “Channel State Information” OR “CSI” OR “Wireless security” OR “Router” OR “Wireless” OR “CSI-based healthcare monitoring”) AND (“Health monitoring” OR “Remote monitoring” OR “non-invasive” OR “Healthcare” OR “Tracking” OR “ECG” OR “Health Monitoring Application” OR “Patient-monitoring” OR “Sleep monitoring” OR “Home monitoring” OR “Secure Remote Monitoring” OR “Smart Health Monitoring” OR “Signal Processing in Healthcare” OR “Monitoring” OR “Detection” OR “Abnormality” OR “AI-driven Healthcare” OR “COTS device" OR “COTS router” OR “Wireless sensing” OR “Contactless Monitoring” OR “Multi-Person Detection” OR “Wi-Fi device”)

After experimenting with several combinations and permutations of key terms, we finalized our search terms, which are presented in Table S1. Searches were conducted using six major academic databases listed in Table S2.

Screening and selection process