This paper presents the design and development of a fog-IoT/AI asthma exacerbation system with full functionality. The system was developed using open-source platforms that monitor real-time medical data. Users administer the Asthma Control Test (ACT) to determine the likelihood of asthma exacerbations. The asthma data set comprises panel data from 10 individuals, with 1010 ACT scores as the desired output. ACT scores 19 reflect uncontrolled asthma; >19 reflect well controlled asthma. This paper proposes a federated learning-based asthma exacerbation prediction system named FELAE. Specifically, the FELAE system protects data privacy through local learning, in which devices benefit from the knowledge of their peers by sharing only updates from their model with an aggregation fog layer that produces an enhanced prediction model. The results demonstrate that the FL approach outperforms the classic or centralized versions of machine learning (non-federated learning). Moreover, using the essential performance indicators, namely, accuracy, precision, f1score, and recall, the proposed model detects asthma exacerbations with the highest accuracy of 97.02%.
Healthcare Internet of Things (HIoT) refers to the
application of Internet of Things (IoT) devices and sensors in
the healthcare area. Connected to the internet and able to
collect, transmit, and analyze data in real time [
Temperature monitoring. Health IoT devices like wearable sensors and smart thermometers ofer real-time body temperature data, transmitted to healthcare systems. Analysis of this data detects temperature trends or abnormalities for early intervention, benefiting patients and caregivers. Additionally, these devices aid in monitoring the health of vulnerable individuals, notifying caregivers of potential health concerns.
BP monitoring. Blood pressure monitoring assesses arterial force during circulation, vital for cardiovascular health and detecting conditions like hypertension. A sphygmomanometer, with cuf, gauge, and stethoscope, measures systolic (during heartbeats) and diastolic (between beats) pressures. These two numbers reveal crucial information about an individual’s blood pressure and overall well-being. night waking up, activity limitation, lung function, and medication usage over time. It is a critical tool for asthma management, changing treatment programs, and lowering the risk of asthma exacerbations. Individuals who have efective asthma monitoring may attain optimum asthma management and enhance their quality of life. The symptoms of asthma can vary in severity and may include: • Wheezing: A whistling sound when breathing out; • Wheezing: A whistling sound when breathing out; • Chest tightness: A feeling of tightness or pressure in the chest; • Shortness of breath: Dificult breathing, especially during physical activity [4].
Figure 1: The classification of HIoT applications. tackled field by focusing on diferent asthma care approaches as well as describing some relevant healthcare applications to our system. The impact of meteorological conditions on people with asthma is characterized by Oxygen saturation monitoring. Achieved through distinct and unique patterns, which may be attributed non-invasive devices like pulse oximeters, assesses the to the intrinsic diversity in lung function found among oxygen levels carried by red blood cells, aiding in res- asthmatic patients. The extent of this diversity is depenpiratory function evaluation. This crucial process helps dent on demographic characteristics, such as age and detect issues like hypoxia or respiratory failure by mea- gender. In addition, the geographic location adds an addisuring SpO2, expressed as a percentage. Using light ab- tional level of complication since the connection between sorption, sensors on the fingertip or earlobe accurately meteorological conditions and the symptoms of asthma determine peripheral oxygen saturation, making it a valu- demonstrates inconsistency across various climatic zones. able health assessment tool. Moreover, asthma systems are data-hungry, and data is Medication management. Encompasses safe and ef- scattered over several hospitals with privacy limitations fective medication use, including prescribing, patient in place. Traditional ML solutions need centralized data education, and monitoring for side efects. It plays a collection and processing, which is becoming more imcrucial role in optimizing patient health outcomes and practical due to eficiency challenges and growing data minimizing adverse drug events. Efective medication privacy concerns [5]. As an outcome of these limitations, management is essential for patient well-being. in 2017, Google introduced the federated learning (FL) Glucose level monitoring. is vital for diabetes manage- approach, where the objective is to train a high-quality ment through various methods like blood glucose meters, centralized model using training data dispersed over a CGM systems, or lab tests. Frequency and targets depend large number of clients, each with unreliable and generon individual needs. Mood monitoring. Utilizing tools ally sluggish network connections. As a result of these like mood diaries or apps, aids in tracking and recording features and inspiration from the previous federated sysemotional states. It assists those with mental health con- tems, FL is a hot research topic in smart HIoT. For examditions by revealing patterns, identifying triggers, and ple, the data of several hospitals is segregated and forms guiding treatment decisions, commonly used for depres- "data islands." Due to each data island’s size and approxsion, anxiety, or bipolar disorder. imation constraints, a single hospital may not be able Wheelchair management. Encompasses assessing mo- to train a high-quality model with excellent prediction bility needs, choosing the right wheelchair, and ensuring accuracy for a particular application. In addition, reguits safe and efective utilization. It enhances indepen- lations cannot force hospitals to provide data in many dence and well-being while preventing complications cases. However, hospitals participating in FL can benefit and improving functional outcomes for those with mo- from it, e.g., with higher model accuracy. A challenging bility impairments. problem is designing a fair incentive mechanism to allow Asthma monitoring. Asthma control tests, peak flow the contributing entity to benefit from FL. The rest of the meters, spirometry, and electronic health records are paper is organized as follows: Section 2 briefly surveys used to track an individual’s asthma symptom diaries, the related works. Sections 3 and 4 provide the materials and methods used in this study, as well as the results of ity of life. The efectiveness of the proposed framework the comparative study and the performance evaluation. for monitoring and regulating asthma exacerbations is Finally, Section 5 concludes the paper and outlines the demonstrated through a case study. The previous disperspectives for future work. cussion made it evident that there are some gaps in the literature, including privacy concerns, computational issues, and accuracy limitations for centralized models, all 2. Related works of which must be successfully addressed to secure smart HIoT data. This paper presents the FELAE framework to In this section, we present and review recent research solve these challenges. Sharing patient electronic health works on IoT-based asthma exacerbation prediction sys- information across hospitals may not be possible due to tems that investigate the most recent developments and the sensitive nature of healthcare data. In such cases, challenges in this field. By analyzing these studies, we FELAE ofers a viable approach, enabling the creation hope to provide insights and recommendations for future of a collaborative learning model for asthma data. The IoT-based asthma prediction system research directions. main contributions of this paper are as follows: Raherison-Semjen et. al [6] addressed asthma management during pregnancy and the influence of environmen- • The design and development of a low-cost IoT/AI tal factors on asthma. It emphasizes the significance of asthma exacerbation system. taking risk factors and prospective comorbidities into • We investigate the implementation of three account in asthma management and individualized man- deep learning classifiers: deep neural networks agement plans for asthma patients. Oletic and Bilas [7] (DNNs), convolutional neural networks (CNNs), described a method for detecting asthmatic wheezing and long short-term memory recurrent networks noises using compressive sensing and machine learning (LSTMs) architectures. techniques to analyze respiratory sound spectra. Us- • In addition, we present a comprehensive perforing a digital stethoscope, the authors acquired respira- mance evaluation and comparison between the tory sound data from asthmatic and non-asthmatic sub- FL approach and centralized learning models. jects. The proposed method detects asthmatic wheezing sounds with high accuracy and has potential applications in portable and non-invasive asthma monitoring devices. 3. Materials and methods Using computer science (CS) and ML techniques, the study presents a promising strategy for the development of an automated and accurate asthmatic wheezing detection system. Anan et. al [8] described the creation of an IoT-based remote health monitoring system for asthmatic patients. The system consists of an Android application, a website, and multiple sensors to collect health-related data and facilitate communication between physicians and patients. The system was determined to be accurate and afordable for low-income asthma patients after being tested on actual human subjects. Tsang et. al [9] reviewed the use of machine learning algorithms in mobile health for asthma management. The review highlighted the potential of machine learning to improve asthma management but also noted the need for larger sample sizes and external validation of algorithms before they can be used in clinical practice. The article discussed various studies on the use of machine learning algorithms for asthma management, including activity detection and breathing monitoring. A fog-driven IoT e-Health surveillance and control framework for asthma exacerbations is proposed by Maach et al. [10]. The framework gathers physiological data from asthma patients using ubiquitous sensors, then processes the data using fog nodes and cloud computation. By providing personalized and timely interventions, the proposed framework is scalable and has the potential to enhance asthma patients’ qual
been proposed to collect, transmit, and process the physical parameters (temperature, humidity, O2, air flow) of patients along with the weather forecast information to manage the decision of asthma exacerbation. In our study, it is necessary to transform pressure readings into airflow. We convert the values obtained from the MPX5010 sensor to kilopascals (kPa), utilizing a scale ranging from 0 to 40 kPa. As depicted in Figure 2, the network component of our platform is supplied via multi-hop communication between box A (the data collection layer) and box B (the gateway node or fog layer) in order to deploy classification and prediction models. We have used the NRF24L01+ (i.e., 2.4Ghz radio) module for wireless communication and the Atmega328P (with Arduino bootloader) as microcontrollers, as well as Raspberry Pi 3 B+ as a fog layer. Data is stored closely in the fog layer, so doctors can rapidly access data during interventions. In addition, asthma data is accessible even if the internet connection is temporarily lost. Processing and validating data at the fog level reduces the data transmitted to the cloud and conserves the energy consumption and global network bandwidth. In a centralized learning configuration, as illustrated in Figure 3, every client uploads his data to a centralized deep learning server to train the prediction model. In contrast, with the FELAE framework, the entire process is adapted from the basic and widely used framework of Federated Averaging (FedAvg) [5] [11]. In particular, instead of training and assessing the model on a single machine, all the clients train their local models sharing the same structure, but with distinct and individual datasets. Subsequently, the trained local models are submitted to the aggregation server that combines all the models to produce a single global model with optimized parameters. This method allows the participants (typically the hospitals) to share knowledge while protecting the confidentiality of their sensitive information. Importantly, this collaborative approach eliminates the need for a high-authorization third party, a requirement frequently associated with high levels of trust and sturdiness. Such a requirement may impose financial restrictions that inhibit broader participation in FL-related initiatives. The current version of our system uses four lfoating values to transmit, as well as the user’s ID code. Each floating value is encoded in 4 bytes, while the user’s ID code is encoded in 8 bytes. Overall, we have 32 bytes to transmit from the emitter to the receiver. NRF24L01+ modules are configured to transmit and receive 32 bytes at a rate of 2 Mbps for real-time monitoring. However, due to the collisions and connections lost, we have implemented a mechanism for auto-restarting the module after ifnishing each transmission. This feature causes a minor delay, but it ensures the stability of the transmission over time.
tory [12], where Python 3 served as the primary programming language. The implementation of Convolutional Neural Networks (CNNs), Deep Neural Networks (DNNs), Long Short-Term Memory networks (LSTMs), and FL models leveraged widely recognized libraries. Specifically, we utilized NumPy for the manipulation of multidimensional arrays and matrices, as well as Pandas for the manipulation of data structures and the utilization of rich analytical tools.
processed input data is filled, digitized, and normalized. Fortunately, the chosen dataset has no missing NaN values, and the corresponding numerical data are all digitized. In this study, we used existing datasets in the context of the asthma dataset with the target variable ACT score.
ACT scores 19 reflect uncontrolled asthma; >19 reflect well controlled asthma [ 4]. The provided dataset [13] [14] is divided in half at an 80:20 ratio. In other words, 80% of the data is utilized for training, while the remaining 20% is used for testing. Additionally, 80% of the data from the training step is divided into K=4 clients, each representing a hospital’s data in our example. For the data distribution among the various clients, we employed independent and identically distributed (IID): Each FL client’s data distribution aligns with the distribution of all the dataset’s data.
Accuracy =
+ + + + Precision =
+ Recall =
+ F1-score = 2 · ( · )
+ • True Positive (TP): reports the number of ACT scores samples that are correctly classified as well controlled asthma. • False Positive (FP): reports the number of ACT scores samples that are wrongly classified as well controlled asthma. • True Negative (TN): reports the number of ACT scores samples that are correctly classified as uncontrolled asthma. • False Negative (FN): reports the number of ACT scores samples that are wrongly classified as uncontrolled asthma. • Accuracy: reports the proportion of properly categorized samples to all other samples in the testing set. • Precision: reports the percentage of samples properly categorized for all TP and FP in the testing set. • Recall: the ratio of TP samples to the total number of TP and FN samples is known as recall. • The F1-score reports the harmonic mean between precision and recall. (1) (2) (3)
Table 2 displays the precision, recall, accuracy, and F1(4) score for binary-class classification for the centralized model that used deep learning methods. It shows how to identify positive instances, both the CNN and DNN well the model performs in predicting the severity of architectures generally outperformed the LSTM. This inasthma for specific asthmatic patients (well controlled dicated their superior capability in identifying patients asthma, uncontrolled asthma). In this analysis, we em- with severe asthma, which is especially noteworthy given ployed three distinct neural network architectures, CNN, the absence of sequential data patterns in this context. LSTM, and DNN, to forecast the severity of asthma for Lastly, precision, which measures the model’s ability to specific patients in four diferent clients. We evaluated classify positive instances accurately, illustrated that the the outcomes using various performance metrics, includ- DNN architecture consistently maintained higher preciing accuracy, F1-score, recall, and precision. To begin, sion in most clients, implying a lower rate of false posiwe assessed accuracy, a measure of overall correctness. tives. Conversely, the LSTM architecture produced more
Table 2 shows that the CNN architecture consistently false positives, resulting in lower precision scores. In sumachieved the highest accuracy across the clients, rang- mary, the DNN architecture emerged as the most efective ing from 84.15% to 88.11%. The DNN architecture also choice for predicting asthma severity across the clients, demonstrated respectable accuracy, ranging from 78.71% consistently excelling in accuracy, F1 score, and precito 90.59%. In contrast, the LSTM architecture consistently sion. The CNN architecture also performed admirably, exhibited the lowest accuracy, ranging from 72.27% to particularly in identifying cases of severe asthma. The 88.61%. Next, we considered the F1-score, which balances DNN architecture emerged as the most efective choice precision and recall. The analysis revealed that the DNN for predicting asthma severity across hospitals, consisarchitecture consistently had the highest F1 score across tently excelling in accuracy, F1 score, and precision. The all clients, indicating its efectiveness in classification. CNN architecture also performed admirably, particularly The CNN architecture also displayed robust F1 scores, in identifying cases of severe asthma. while the LSTM architecture needed to catch up, suggesting dificulties in achieving high precision and high recall. Turning to recall, a measure of the model’s ability
LAE framework. This method requires the participation of multiple clients, specifically hospitals, to share knowledge while protecting the confidentiality of their sensitive information. Importantly, this collaborative approach eliminates the need for a high-authorization third party, a requirement frequently associated with high levels of trust and sturdiness. Such a requirement may impose ifnancial restrictions that inhibit broader participation in FL-related initiatives. To evaluate our proposed FELAE scheme, we have conducted a series of tests. These in- Figure 4: Learning performances using a FedAvg-based CNNvestigations involved the construction of a controlled FL model. environment in which deep learning models (i.e., CNN, DNN, and LSTM) were implemented on a Raspberry Pi 3 board. The figures 4, 5, and 6 show how well all four global models worked over 10 rounds, with three diferent deep-learning classifiers used for each client (hospitals).
It is worth mentioning that the FL training process is done over 10 rounds, where each model is saved after every round to avoid overfitting after a long period of training. The primary conclusion drawn from this study is the discernible improvement in precision as a function over iterative rounds across all FL global models. This improvement signifies the concurrent progress and mutual benefits realized by all the participants due to their participation in the global model. A notable corollary observation is that, in certain instances, global models have demonstrated the ability to approach or closely rival Figure 5: Learning performances using a FedAvg-based DNNthe performance levels attained by the centralized model. model.
In the evaluation, we ran the FL training process for 10 rounds. However, we save the global model at each round to avoid overfitting issues after a long training period. accurately discerning patients as either ’well controlled’ Table 3 shows a full analysis of how well three diferent or ’uncontrolled, boasting many true positives in these neural network architectures CNN, DNN, and LSTM can categories. predict the severity levels of asthma in a dataset of pa- Conversely, the DNN classifier accurately identifies tients. The CNN classifier exhibits notable proficiency in patients at the extremes of well-controlled and uncontrolled asthma, achieving high true positive rates. Nevertheless, it also needs help classifying ’partially controlled’ cases, as it tends to make false predictions of ’well controlled’ and ’uncontrolled’. The LSTM classifier accurately distinguishes between well-controlled and uncontrolled asthma cases, with notable true positives in these extreme categories. According to Table 3, the CNN model stands out as the top performer among the three evaluated deep learning models, with a substantial accuracy of 97.02%. It excels at accurately categorizing patients with varied levels of asthma severity and a substantial number of true positives and true negatives. Notably, the F1score, a metric balancing precision and recall, reaches an impressive score (i.e., 97.01%), highlighting the model’s efectiveness in minimizing false positives while capturing true positives. In contrast, the LSTM model, although reasonably accurate with an 89.10% accuracy rate, grapples more with false positives and negatives. As a result, its F1-score and precision are lower than those of CNN, indicating dificulties in striking the ideal balance between precision and recall. Finally, DNN reports good performances with 94.05% accuracy. Like CNN, it maintains an efective equilibrium between precision and recall, leading to a high F1 score (i.e., 94.01%). Furthermore, its precision slightly outperforms CNN, which lowers the number of false positives. As presented in Table 3, CNN can perform better with 97.02%, 94.05% for DNN, and 89.10% for LSTM, respectively. Overall, through these experiments, we can highlight that CNN is the more reliable and ranked first due to the strengths of extracting the category features.
IoT-based asthma exacerbation prediction system using the FL approach. Our proposed system aims to provide asthma patients a user-friendly solution for monitoring their symptoms and anticipating potential crises. The three phases of data collection, analysis, and treatment serve as a road map for the system’s ongoing development. It is essential to acknowledge that there are several perspectives for the future enhancement of this project. As we continue our work, we plan to explore additional features and functionalities to improve the system’s efectiveness. The following recommendations summarize the research challenges that could enhance the performance of the proposed asthma exacerbation system: • Using the proposed low-cost IoT/AI asthma exacerbation system in order to generate our dataset for the pulmonology department at the university hospital in Oran. • Use teacher and student networks and knowledge distillation (KD) techniques to make models smaller, faster, and more eficient. • Include the most specific asthmatic symptom in respiratory sounds, such as wheezing, in our dataset.
der "Respiratory pathologies via artificial intelligence," supports this work. The authors would like to acknowledge the medical team of the pulmonology department, Oran University Hospital, and the Thematic Research Agency in Health and Life Sciences (ATRSSV).