FedRadar: Federated Multi-Task Transfer Learning for Radar-Based Internet of Medical Things

Owing to increasing connected medical devices and rapid development of deep learning technologies, the Internet of Medical Things (IoMT) has attracted great attention for healthcare monitoring with intelligent sensors. Radar serves as a non-contact healthcare device, continuously measuring human’s vital signs and behavior to provide daily and comprehensive long-term record on health status. However, radar sensor data collected from various users and families usually involve sensitive personal information, while traditional deep learning technologies using IoMT radar data may present high privacy leakage risks. This paper proposes FedRadar, a novel federated multi-task transfer learning framework for radar-based heartbeat rate and activity monitoring in IoMT to solve these challenges. It deals with the decentralized structure, training personalized multi-task models collaboratively, combining the shared relevance knowledge of human’s physiological information while keeping personal radar data locally for privacy protection. First, the multi-task neural network is built on the spatial-temporal radar data to capture the potential relationship and shared representations between human’s vital sign and activity. Furthermore, the federated learning with knowledge transfer scheme is designed to achieve personalized local models by transferring coarse relevant features and keeping fine-grained individual information. Extensive experiments demonstrate the effectiveness and robustness of FedRadar with 2.8% and 2.5% superior than local training model respectively on the accuracy of heartbeat rate estimation and activity classification in realistic constructed radar datasets. In addition, FedRadar is extensible and suitable to continuously monitor multiple health indicators with privacy protection in IoMT. The FedRadar codes and constructed radar datasets are available on https://github.com/bupt-uwb/FedRadar.


I. INTRODUCTION
H UMAN'S vital signs and activities are highly associated with the health status. Heartbeat serves as one of the most significant indicator for wellbeing, and the changes on motion pattern are proved to be useful symbols of cognitive disorders, such as Parkinson's and cerebral small vessel diseases [1]. Daily monitoring on physiological and behavioral activities has a potential to provide a comprehensive assessment on health status and early warnings for several diseases. The emerging Internet of Medical Things (IoMT) transforms healthcare from hospital-centric to home-centric, making the daily monitoring realizable and generating massive data by heterogeneous sensors [2]. With the increasing aging people, indoor personal healthcare is becoming significant especially for daily vital signs monitoring. Supposing that a living-alone elderly suddenly suffers from abnormality such as heart attack, he may not get timely rescue without caregivers or other people in company with him. The IoMT-enabled healthcare system can monitor his vital signs and When abnormality is detected, warning can be delivered to medical workers or hospitals with the patient's posture because he may fall on the ground and other required information in the IoMT database to conduct timely rescue.
A variety of sensors have been applied for IoMT-enabled healthcare. Contact devices such as Electrocardiograph (ECG) and Polysomnography (PSG) measure vital signs by attaching electrodes to human's skin, which bring discomfort to people and limit daily activities. Non-contact sensors provide pervasive and non-intrusive monitoring, and have attracted wide attention for remote healthcare monitoring [3]. Compared with the radio frequency identification (RFID) tags, Wi-Fi and camera, radar systems do not require the participant to equip with any device, and are not affected by illumination and facial privacy issues. In addition, compared with the single indicator as the physiological parameter or daily behavior, radars provide both information simultaneously. Impulse Radio Ultra-Wideband (IR-UWB) radar [4] transmits and receives the narrow impulse signal with a wide bandwidth, analyzing signals reflected from the environment for vital signs and activity monitoring. It has fine delay resolution and strong penetration, and achieves good performance in sleep stage classification [5], fall detection [6], and people counting [7].
Benefiting from the rapid development of deep learning, enormous health data are trained with machine learning models to perform effective monitoring. Human's vital signs and behaviour data collected in each family present as isolated islands, which obstruct effective training with lacking all of the valuable data. Traditional cloud-based method uploads a large quantity of data from the distributed sensors to a centralized cloud server for further analysis [8]. However, sending sensitive data with personal cardiopulmonary and motion information raises high security and privacy leakage risks. In addition, several policies and regularizations have been produced in China, the United States and the European Union [9] to protect data privacy and restrict the data sharing and data access. Furthermore, the model trained on the central server fails on personalization. Therefore, it is infeasible to integrate massive personal data from different families and train an effective machine learning model for healthcare monitoring.
This paper proposes a novel FedRadar framework for privacy-preserving vital sign and activity monitoring based on IR-UWB radar in IoMT-enabled healthcare to tackle these challenges. FedRadar collaborates the data from isolated participants to build secure distributed local models without compromising the privacy and security of personal data. In order to investigate the potential relevance of physiological radar data, a multi-task neural network is constructed to obtain the relationship between vital sign and activity information. Moreover, the federated learning with knowledge transfer scheme is designed to achieve personalized and tailored local models by transferring shared relevant features. The structure of the proposed FedRadar framework is shown in Fig. 1. In summary, the main contributions of this paper are summarized as follows.
• A novel framework FedRadar is proposed with IR-UWB radar for accurate end-to-end health monitoring, which collaborates all the radar data from different users to train distributed models without leaking any sensitive information. To the best of our knowledge, this is the first attempt on radar-based federated learning framework to achieve vital sign and activity monitoring simultaneously in IoMT-based healthcare. • The multi-task neural network is designed as local model to capture intrinsical relationships between cardiopulmonary and motion information, which shares common spatial-temporal radar data reflected from human body. The shared relevance feature extraction network catches potential connections between human's physiological data, while vital sign monitoring and activity monitoring sub-networks mutually learn extra information through joint training. • In order to achieve personalized local models, the idea of transfer learning is exploited in FedRadar, and the federated learning with knowledge transfer scheme is designed. The shared relevance features are transferred from collaboratively optimized weights, while the vital sign monitoring and activity monitoring sub-networks train and update locally with personal radar data to ensure tailored model for each specific user. The remainder of this paper is organized as follows. Section II introduces the related works. The proposed FedRadar framework is described in detail in Section III. Experimental results and analysis are presented in Sections IV and V concludes this paper.

II. RELATED WORKS
In this section, the related works are introduced in the perspective of radar-based vital sign and activity monitoring, federated learning, transfer learning, and multi-task learning respectively.

A. Radar-Based Vital Sign and Activity Monitoring
IR-UWB radar with penetrating pulses effectively captures the minor movements from heart and lungs, and is proved to Authorized licensed use limited to the terms of the applicable license agreement with IEEE. Restrictions apply. provide a high SNR for respiration and heartbeat rate monitoring compared with other radars [10]. Several researches have been conducted on heartbeat monitoring with the IR-UWB radar. Reference [11] proposes a vital sign monitoring method with random body movement cancellation using a 2 × 2 IR-UWB distributed MIMO radar system. In [12], the method based on autocorrelation and variational mode decomposition (VMD) is proposed to detect the people and measure the respiration rate (RR) and heartbeat rate (HR) with an UWB radar. A framework based on 7.29 GHz IR-UWB radar for reliable heartbeat monitoring is designed in [13] with four steps by selection and fuzzy logic rule-based method.
Besides monitoring vital signs, detecting the activities of human bodies based on the radar sensor has attracted various researches. The dynamic range-Doppler trajectory features are extracted for human motion recognition in [14], continuously classifying the motion of falling, stepping, jumping, squatting, walking and jogging. A comprehensive set of features are extracted and the Bhattacharyya distance is utilized for feature selection in [15] to recognize different predefined events in a room. In [16], the low-level deep features from CNN, empirical features and statistical features are extracted and fused to classify seven different human activities with unsupervised domain adaptation. Researches in [17], [18] detect the falling motion with Doppler features and frequency distribution respectively for in-home elderly monitoring.
In FedRadar, the deep neural network is designed with IR-UWB radar for accurate and robust end-to-end vital sign and activity monitoring simultaneously.

B. Federated Learning
Federated learning is firstly proposed by Google in [19], which learns from decentralized data by a shared global model on the central server without data sharing. It achieves performance improvements on the machine learning model by aggregating locally-computed updates from all of the clients, while keeps the personal data locally for privacy protection. Several studies have been conducted with federated learning for IoMT-enabled healthcare. Reference [20] designs a generative convolutional autoencoder to deal with the imbalanced and non-IID distribution problems in the cloud-edge federated learning framework for personalized in-home health monitoring. In [21], the FedHealth is proposed by aggregating wearable device data through federated learning, and building relatively personalized models by transfer learning. Reference [22] combines federated learning and emotion analysis for human emotion monitoring. In [23], a deep federated learning framework is proposed for healthcare data analysis with IoT devices. Reference [24] proposes the Double Deep Q-Network (DDQN) based on a Fully Decentralized Federated Framework (FDFF) for clinical decision system.
FedRadar exploits federated learning as the basic scheme in building and collaboratively training distributed models to prevent privacy leakage.

C. Transfer Learning
Transfer learning aims at improving the performance of the target task by transferring the knowledge from source tasks without learning from scratch with massive data [25]. It addresses the problem of insufficient well-labeled training data [26], and has been successfully applied to various smart health applications.
Reference [27] investigates the feature fusion of handcrafted method and deep learning method based on knowledge transfer from a ResNet-50 network for melanoma classification in dermoscopy images. In [28], a Hierarchical Attention Transfer Network (HATN) is proposed to learn attention from a source task, combining unsupervised learning, knowledge transfer and hierarchical attention for automatic assessment of depression from speech. Reference [29] explores the effects of training samples size on multi-stage transfer learning on breast cancer diagnosis in digital breast tomosynthesis. Reference [30] proposes a principal component analysis (RPCA)-embedded transfer learning (TL) to realize a personalized cross-day EEGbased emotion classification model with less labeled data. Reference [31] demonstrates that transfer learning brings benefits for deep learning methods on extracting primary sites from breast and lung cancer pathology reports, but the general trends are contingent on the cancer site and training methods. In [32], the Distant Domain Transfer Learning (DDTL) model is developed for COVID-19 diagnose. The two binary classifications based on deep transfer learning are proposed in [33] for automatic preterm newborn presence detection. Reference [34] discusses the two-step progressive transfer learning by finetuning the deep convolutional neural network on two skin disease datasets.
FedRadar introduces the idea of transfer learning and designs the federated learning with knowledge transfer for personalized local model.

D. Multi-Task Learning
Multi-Task Learning (MTL) is proposed to leverage valid information in multiple associated tasks to improve the generalization performance on all the tasks [35].
Reference [36] proposes a 3-D multi-attention guided multitask learning network for automatic gastric tumor segmentation and lymph node (LN) classification, fully combining complementary information between tumors and LNs. In [37], a Semi-Supervised Multi-Task Learning (SS-MTL) approach is developed for predicting short-term Kidney Disease (KD) evolution on general practitioners' Electronic Health Records (EHRs). An end-to-end multi-task deep learning framework is proposed in [38] for skin lesion analysis on dermoscopy images, achieving skin lesion detection, classification and segmentation tasks simultaneously. Reference [39] investigates a multi-task learning approach for simultaneously modelling multiple Type 2 diabetes mellitus (T2DM) complications, capturing the relationships between complication risks. A deep multi-task regression learning model is conducted in [40] for full quantification on left ventricle in magnetic resonance (MR) images. Reference [41] proposes the deep multi-task multi-channel learning (DM 2 L) framework for brain disease classification and clinical score regression from MR images, while [42] develops a Multi-task Multi-slice Deep Learning System (M 3 Lung − Sys) for multi-class lung pneumonia screening on CT images. In [43], a multi-task joint learning (MTL) method is proposed for tongue images segmentation and classification.
FedRadar constructs a multi-task neural network to investigate the shared representations between vital sign and activity monitoring tasks.

III. PROPOSED FEDRADAR FRAMEWORK
In this section, FedRadar, the proposed IoMT-enabled noncontact heartbeat rate and activity monitoring framework is introduced, as shown in Fig. 1. There are three innovative modules in FedRadar, including 1) the IR-UWB radar-based user's data collection module; 2) the personalized vital sign and activity monitoring module; and 3) the federated learning with knowledge transfer module.
In FedRadar, the IR-UWB radar serves as a household monitoring device, capturing the radar signal with user's behavior and heartbeat rate information in a non-contact and noninvasive way. Deep learning has shown a strong ability for IoMT-enabled healthcare monitoring, which requires to integrate a large amount of user's data. For radar nodes distributed at m users' homes, the local multi-task neural network is designed for heartbeat rate and activity monitoring simultaneously, capturing the potential relevance and underlying relationships between them and improving the generalization performance. In order to preserve radar data privacy, construct personalized model, and engage with the decentralized structure for health monitoring, FedRadar is constructed with a federated learning scheme with knowledge transfer. In following subsections, the framework design of FedRadar is described in detail.

A. IR-UWB Radar Signal Model
In FedRadar, the IR-UWB radar is utilized as a household device node for non-contact personal data collection. It periodically propagates the impulse signal with wide bandwidth, and receives reflections from the environment for analysis. The received IR-UWB radar signal r (t, τ) is represented in follows: where τ denotes the fast time along signal propagation, indicating the time of arrival (ToA) of the impulse signal, and t is the slow time for observing and accumulating radar signals. s(t, τ) and n(t, τ) represent the original transmitted impulse signal and the noise of the channel, respectively. l indicates the l-th travelling path and N path is the number of received paths. a lt and τ lt represent the scaling factor and time delay respectively of the l-th received path at the t-th slow time. IR-UWB radar stores several received signals as a two-dimensional (2-D) matrix R[o, k] for further processing, expressed as: where o and k are the bin numbers in slow time and fast time respectively. T s denotes the pulse repetition interval of slow time, and T f is the interval between two adjacent fast time samples. The distance d between the IR-UWB radar and the detected target can be inferred from the fast time τ , as: where c represents the light speed of 3 × 10 8 m/s. Thus in the 2-D radar matrix, each row represents the received signal along different distances at a slow time, while each column indicates the reflections at a distance during a period of observation time. This 2-D matrix contains both spatial distribution and temporal variation information of radar signals. Three preprocessing steps are firstly performed on received radar signals: 1) Direct Current (DC) component removal; 2) bandpass filtering; and 3) clutter removal. The DC component is caused by the direct wave from the transmitter to the receiver, bringing an offset to received radar signals. The average of each received signal is calculated and subtracted to remove the DC component. Then a Hamming window is designed to filter received signals to the radar operating frequency range. In the household environment, clutters reflected from the stationary objects badly affect the human information extraction. To obtain the refined radar signal with valid personal information, the running average method [44] is utilized for clutter removal, expressed as: where s(t) represents the current radar signal at slow time t. c(t) is the estimated clutter signal, which is the weighted average between c(t − 1) and s(t). The weighted factor α ranges from 0 to 1, determining the ratio of the current signal to the previously estimated clutter signal. x(t) denotes the refined signal after clutter removal. The refined radar data includes reflections from human body, which imply the backscattered signals from human's limbs and chests. Therefore, the spatial-temporal radar signal backscattered from human body involves both human's activity and vital sign information, and accumulates long-term personal recordings. Activity Monitoring: Multiple reflected impulse signals from the human body form several clusters, which consist of reflections from human's limbs and trunk, indicating the activity information, such as the posture and activity events. Several machine learning algorithms have been conducted with IR-UWB radar signals for activity monitoring by extracting valid features from these clusters, but hand-crafted features rely on manual processing and expert knowledge. Thus FedRadar designs the deep neural network for an end-to-end monitoring.
Vital Sign Monitoring: IR-UWB radar measures the movement of human chest as follows: where d p is the antenna-to-people distance and d c (t) represents the displacement caused by cardiopulmonary movement. a r and a h denote the displacement amplitudes caused by respiration and heartbeat motions, while f r and f h are the frequencies of them respectively. In conventional heartbeat monitoring methods, the signal with the maximal energy in a fixed distance is selected and regarded as the most vital sign information, and is processed with signal decomposition algorithms and FFT to estimate the heartbeat rate. However, human motions in daily life bring a changing d p and distortions on extracted vital sign signals, making the conventional methods unable to accurately monitor the heartbeat rate. In order to address these problems, a multi-task neural network is designed in FedRadar as the local model to achieve accurate and robust monitoring on human's activity and vital sign simultaneously.

B. Multi-Task Neural Network for Vital Sign and Activity Monitoring
Vital sign and activity monitoring tasks are intrinsically related and both based on human's physiological characteristics, which share common spatial-temporal information on radar signals reflected from human body. Considering the diversity and relevance of human's physiological information, a novel multi-task neural network is designed as the local model for end-to-end monitoring on both tasks simultaneously. The model takes the refined 2-D IR-UWB radar data as the input, and outputs the heartbeat rate estimation and activity classification results jointly. The proposed model architecture consists of a shared relevance feature extraction network, and two sub-networks for vital sign monitoring and activity monitoring tasks respectively, as shown in Fig. 2. The shared relevance feature extraction network captures coarse shared representations and general relationships between two tasks. Then the vital sign monitoring and activity monitoring sub-networks learn more fine-grained individual heartbeat and motion information.
1) The Shared Relevance Feature Extraction Network: The shared relevance feature extraction network is designed to capture the potentially relevant features and transferable knowledge between human's heartbeat and activity information. Considering that bottom layers produce coarse features while higher layers generate more complex and differentiating taskspecific features, shallow convolutional layers are applied to extract general and common features. The shared relevance feature extraction network consists of two convolutional layers with the size of 4 × 8, and each of which is followed by a batch normalization layer, a rectified linear unit (ReLU) activation function, a dropout layer and a 4 × 4 max-pooling operation with a stride of 2 × 2. This network generates feature maps with general physiological information relevant to both vital sign and activity.
2) The Vital Sign Monitoring Sub-Network: In the vital sign monitoring sub-network, four convolutional layers each with a ReLU activation function are firstly designed to remove body motion interference and extract valid vital sign information. The long short-term memory (LSTM) serves as a variant of recurrent neural network (RNN), which captures long-term temporal variations and overcomes the vanishing and exploding gradients problems of RNN. The LSTM unit consists of the memory cell and three gates, including the forget gate f t to discard useless previous information, the input gate i t to determine the storage of new information, and the output gate o t for deciding the output. This structure enables LSTM to memorize long-term dependencies and model temporal dynamics for the sequential data. Considering the temporal dependency and continuity of time-varying heartbeat signal, the LSTM structure is followed to analyze and capture temporal features, exploring the underlying periodic cardiac motion in radar signals. The structure and computation of the LSTM unit are described in the following equations: where t represents the time step. x t and h t are the input and output respectively. c t indicates the cell status. W f , W i , W c and W o are the weight matrices, while b f , b i , b c and b o are the bias vectors. σ is the logistic sigmoid function, and tanh denotes the hyperbolic tangent function. represents the scalar product. Three layers of the LSTM network are stacked for extracting more valid information, which imply the temporal variations of heartbeat signals. Finally, a dense layer is applied, and a single neuron is used for heartbeat rate estimation. The mean square error (MSE) loss is computed for heartbeat rate regression in the vital sign monitoring sub-network, defined as: where y i and y i denote the estimated heartbeat rate and the ground truth measured by the oximeter for the i-th radar sample, respectively. N represents the total number of radar samples.
3) The Activity Monitoring Sub-Network: In the activity monitoring sub-network, four convolutional layers with ReLU activation function are adopted for human posture and body motion features extraction, and each of them is followed by a dropout layer to prevent overfitting. The dense layer is utilized subsequently, and 3 neurons are used with SoftMax transfer function to generate the classification probability for three different activities. The cross entropy loss is utilized for activity classification, calculated as: where N indicates the total number of activity classes. y i and y i represent the ground-truth label and the classifying output probability respectively. For each input IR-UWB radar data x i , the uniform loss function of the multi-task neural network is composed by the regression loss l v (·) of the vital sign monitoring sub-network and the classification loss l q (·) of the activity monitoring subnetwork, as shown in the follow equation: where θ 1 , θ 2 and θ 3 represent the parameters of the shared relevance feature extraction network, the vital sign monitoring sub-network and the activity monitoring sub-network respectively. α is the factor controlling the relative importance of two tasks. The parameters in both tasks are jointly optimized, which makes two sub-networks mutually learn extra information, and improves the generalization ability and monitoring efficiency of the multi-task neural network.

C. Federated Learning With Knowledge Transfer
FedRadar exploits federated learning as the main scheme in building and collaboratively training models without sharing personal radar data in order to prevent privacy leakage. Each client node for different user generates radar data in diverse distributions, which brings Non-Independent Identically Distribution (Non-IID) problem due to the complexity of physiological information and individual differences. Moreover, conventional federated learning algorithm which trains a shared global model on the central server and distributes it to clients lacks personalization. Therefore, FedRadar introduces MOCHA [45] as the federated learning algorithm to build separate local models on distributed personal data, and achieves personalized local models by knowledge transfer on the shared relevance feature extraction network.
MOCHA trains separate but structurally related local models for each participant client without forging a global model, presenting a superiority dealing with Non-IID data, stragglers and fault tolerances [46], and has achieved accurate and robust performance in various applications [47]. Furthermore, considering that humans may have similar heartbeat rate changes when performing the same activity, which presents structure relevance between local models, MOCHA captures homogeneous and heterogeneous relationships between them according to distributed multi-task learning to improve performance. Given radar data X t ∈ R d×nt from m client nodes, the problem formulation of MOCHA is described as follows: where w t ∈ R d is the weight vector for the t-th client node through arbitrary convex loss function l t . x i t represents the i-th radar data in X t , while y i t indicates the label of x i t . W = [w 1 , . . . , w m ] ∈ R d×m is a matrix with the t-th column corresponding to the weight vector for the t-th client node. Ω ∈ R m×m represents a matrix modelling relationships amongst local models in client nodes. The regulation term R(W , Ω) is expressed as the bi-convex formulation: where constants λ 1 , λ 2 > 0 and tr (·) is the trace of a matrix. · 2 F indicates the L 2 regulation. MOCHA adopts the alternating optimization, generalizing the distributed optimization method CoCoA to approximately solve the minimization problem. Joint training for multiple client nodes also introduces significant distribution difference among users' personal radar data, which makes local models fail in personalization. For example, different users may perform specific heartbeat rate level and activity regularity due to individual differences, personal habits and health status. Therefore, the idea of transfer learning is introduced in FedRadar, where only the shared relevance feature extraction network in each client participates in the federated learning process, and the vital sign monitoring and activity monitoring sub-networks only train and update locally. Each local multi-task neural network holds the knowledge transferred from the collaboratively optimized weights in the shared relevance feature extraction network, and is trained with local radar data to achieve more tailored and personalized models for each specific user.
The flowchart of the federated learning with knowledge transfer in FedRadar is illustrated in Fig. 3, including four stages described below.

1) Stage 1(Central Server Initialization):
First of all, the central server communicates with each client node to acquire the participant client number m, and initializes the weight matrix W and the relationship matrix Ω.

2) Stage 2(Transfer Learning on Local Model):
Then the central server sends the model parameter w t as the t-th column of matrix W to each client node t. Subsequently, the client node assigns weights w t to the shared relevance feature extraction network in the local model, and trains the whole multi-task neural network in client t. The shared relevance feature extraction network, the vital sign monitoring sub-network and the activity monitoring sub-network are trained jointly with local radar data X t . The update weights Δw t in the shared relevance feature extraction network are then computed and uploaded to the central server.

3) Stage 3(Weight Matrix and Relationship Matrix Update):
After that, the central server updates the relationship matrix Ω and the weight matrix W alternatively according to each uploaded weights {Δw 1 , Δw 2 , . . . , Δw m }.

4) Stage 4(Personalized Local Model for Vital Sign and Activity Monitoring)
: Stage 2 and stage 3 are repeated for multiple iterations, and ultimately the personalized local model in each client node is realized by utilizing local radar data and the transferred knowledge according to federated learning in the shared relevance feature extraction network. Each tailored local model is adopted with forward propagation for individual heartbeat rate estimation and activity classification in each client.

IV. EXPERIMENTAL RESULTS AND ANALYSIS
In this section, extensive experiments are conducted and results are analyzed to validate the effectiveness and superiority of the FedRadar framework.

A. Experimental Setting and Dataset Construction
In this paper, an IR-UWB radar dataset for activity and heartbeat rate monitoring is constructed in Beijing University of Posts and Telecommunications for validation. The IR-UWB radar is based on the Novelda X4M03 chip, which operates at the center frequency of 7.29 GHz and the bandwidth of 1.5 GHz. The radar is deployed at a height of 1.45 m with the central angle of 65 • , and the detection range of 3 m. Participants are required to perform 3 activities including sitting, standing and lying with various body movements in the detecting area. The oximeter with FDA certification is equipped on each participant as the ground truth for heartbeat rate estimation, and the activities are recorded as the label for activity classification. This dataset is collected in three different indoor environments, including a cotton tent, a small room and an empty lobby, and 4 volunteers (3 men and 1 woman) participate in the experiments. A total of 640 minutes radar data with personal activity and vital sign information is utilized in this paper, while each radar sample is selected with 10-seconds duration and 5-seconds overlapping, which generates 7,680 samples.
To construct the problem situation in FedRadar, four participants are regarded as isolated users in different client nodes, which train distributed local models collaboratively without sharing their data for privacy preserving. The cloud AI platform BitaHub equipped with Intel Xeon Gold 5118 CPU and NVIDIA Tesla V100S graphic card is deployed for carrying out deep learning applications. 80% of the radar data (with 6,144 samples) is randomly selected for training local models, while the rest 20% (with 1,536 samples) is utilized for testing the heartbeat rate estimation and activity classification performance. The Adaptive Moment Estimation (ADAM) algorithm is utilized to optimize two tasks, with 10 local training epochs for local models on each client and 5 global training epochs.

B. Performance Evaluation on Heartbeat Rate Estimation and Activity Classification
The heartbeat rate estimation and activity classification accuracies Acc h and Acc a are utilized to evaluate the monitoring performance, and are defined as: where hr e is the estimated heartbeat rate from FedRadar, and hr o represents the heartbeat rate measured by the reference oximeter. X true and X all denote truly predicted samples and all of testing samples of activity classification respectively. Fig. 4 and Fig. 5 illustrate the heartbeat rate estimation and activity classification accuracies with different global epochs respectively. The loss for these two tasks over global epochs are shown in Fig. 6 and Fig. 7. The monitoring accuracy generally increases with the increasing global epochs, and tends to be stable when the number of global epochs reaches 4. The loss decreases over the epochs, revealing the satisfactory model performance. The heartbeat rate estimation achieves 95.8%, 95.1%, 95.4% and 95.9% accuracies on client 1-4   respectively on global epoch 5, while the accuracies on that for activity classification reach 96.3%, 95.2%, 94.2% and 94.8% respectively. Considering that the number of global epochs implies the cost of transmission and computation, this paper trains 5 global epochs to obtain a high monitoring accuracy while saves the costs.
To verify the generalization of FedRadar, data from different clients are tested on local and federated models of client 1 respectively. Data from client 1-4 are tested on the local model of client 1 with and without FedRadar training, respectively. The model is trained for 50 epochs. Accuracy of heartbeat rate estimation and activity classification are shown in Fig. 8 and Fig. 9. The heartbeat rate estimation achieves 13.3%, 13.7%, 21.1% and 21.9% increase on accuracy while  the activity classification achieves 2.9%, 5.2%, 0.8% and 1.3% increase respectively with FedRadar. The results imply that FedRadar has a stable generalization to learn features from different clients and improve the inference ability of model.

C. Performance Comparison With Other Methods
To validate the effectiveness and superiority of FedRadar, comparisons with other methods on IR-UWB radar data are carried out in this paper for vital sign and activity monitoring. Two heartbeat rate estimation methods are employed, including the selected signal with Variational Mode Decomposition (VMD) [48] and the Heartbeat Estimation And Recovery (HEAR) [49]. For activity classification, two deep neural networks LeNet and AlexNet are introduced as end-to-end classification methods for comparison. Results on the local model proposed in Section III-B without federated learning, and that with the other federated learning algorithm -FederatedAverage (FedAvg) [20] are also compared.
Comparisons on heartbeat rate and activity monitoring accuracies for each client are shown in Fig. 10 and Fig. 11 respectively. FedRadar achieves the highest accuracies on both tasks in average, reaching 95.6% for heartbeat rate estimation and 95.1% for activity classification. The averaging accuracy of FedRadar for heartbeat rate monitoring is 9.6% and 4.8% higher than those with VMD and HEAR methods. For activity classification, FedRadar also presents the effectiveness and superiority, improving the averaging accuracies by 12.3% and 11% than LeNet and AlexNet respectively. Table I shows the detailed accuracy, precision and recall comparison on each client. In effect, the proposed local model  without federated learning still exhibits better performance for both two tasks than other existing methods on clients 1 to 4 and in average. It is observed that VMD presents the lowest accuracy while HEAR performs a slight improvement in heartbeat rate estimation. This is because that HEAR introduces a motion compensation scheme to reduce signal distortions caused by movement interference. Results demonstrate the efficiency and robustness of the designed multi-task neural network for vital sign monitoring against random body movements. Compared with these two classic deep neural networks LeNet and AlexNet, the proposed model benefits from the proper structure and multi-task learning scheme to extract more intrinsically related spatial-temporal information for two tasks. Results show that FedRadar and the proposed model with FedAvg algorithm achieve better performance for both heartbeat rate estimation and activity classification than the local model. This superiority relies on the federated learning with knowledge transfer scheme, which combines massive knowledge from distributed data to train a better model, and keeps local model personalized according to the shared relevance feature extraction network with transfer learning. It is observed that both federated learning algorithms obtain higher accuracies than the local model on almost all clients and in average for two tasks, except for that on client 4 for activity classification. This can be explained that postures of volunteer 4 differ much from the other 3 volunteers and personal samples may overfit on the individual local model, while federated learning with knowledge transfer scheme mitigates this defect and strikes an balance over all clients.    server and the brown line is the averaging amount of transmitted data from each client. It is observed that the averaging transmission amount from client is almost constant to the number of clients, while the transferred data from the central serval increases linearly with more clients. This is caused by the MOCHA federated learning scheme, of which the client uploads its' own weights, and the central serval only sends each column of the weight matrix to each client. Thus increasing clients lead to linearly growing transmission amount from the central server. The averaging training time for each client with different number of clients in a global epoch is illustrated in Fig. 13, each with 10 local epochs. It is also approximately the same with increasing clients.

V. CONCLUSION
This paper proposes a novel federated multi-task transfer learning framework FedRadar for IR-UWB radar based vital sign and activity monitoring in IoMT. FedRadar collaboratively trains secure distributed models with massive knowledge based on federated learning scheme without leaking privacy data, and achieves personalized local models by designing the multi-task neural network with transfer learning. Extensive experiments on realistic constructed IR-UWB radar dataset demonstrate that FedRadar achieves excellent performance in heartbeat rate estimation and activity classification, and performs a superiority compared to other methods. This distributed framework is extensible and suitable to personalized healthcare applications for continuously monitoring multiple health indicators with privacy and security protection. To the best of our knowledge, this is the first attempt on radar-based federated learning framework to achieve health monitoring. In the future work, more radar data will be collected to further validate and improve the feasibility of FedRadar.