Energy Efficient Heartbeat-Based MAC Protocol for WBAN Employing Body Coupled Communication

Wireless Body Area Networks (WBANs) are a fast-growing field fueled by the number of wearable devices developed for countless applications appearing on the market. To enable communication between a variety of those devices, the IEEE 802.15.6 standard was established. However, this standard has some intrinsic limitations in addressing the heterogeneity of the network nodes in terms of activity, data rates (from less than bit/s to multiple Mbit/s), energy availability, form factor, and location on, around or inside the body. To address these concerns, an alternative model is proposed that could serve as an extension of the IEEE 802.15.6 Standard. At its core is an adaptive and low-overhead synchronization scheme based on heartbeat sensing. This forms the base for a TDMA-based (Time Division Multiple Access) Media Access Control (MAC) protocol dedicated to multi-tier networks. While this effort focuses specifically on Capacitive Body-Coupled Communication (C-BCC), other physical layers can be easily incorporated as well. Based on these premises, this paper compares various random-access slot allocation approaches to accommodate the multiple data rates matching the system requirements, while incorporating a duty-cycling strategy anchored by heartbeat detection. This work proposes a novel, flexible, and robust solution, making use of heartbeat synchronization and addressing the corresponding challenges. It efficiently interconnects multiple device types over a wide range of data rates and targets a mesh of stars topology. At the cost of an increased communication latency, the proposed protocol outperforms the IEEE 802.15.4 MAC standard in terms of energy efficiency by a factor of at least 12x in a realistic scenario.

inductive, capacitive, etc). In this paper, we pay special attention to sub-GHz Capacitive Body-Coupled Communication (C-BCC), a prime candidate for the physical layer of human-centered body communication, due to its efficiency and robustness. However, to reach high performance at a low energy cost, co-optimization between PHY and MAC layers is mandatory. A crucial challenge in reaching the best energy efficiency in managing communications in hybrid and versatile BANs supporting widely varying workloads such as the Human Intranet [1], is the establishment of a reliable synchronization scheme. While synchronization could be accomplished through the broadcast of beaconing signals, or rendezvous and wake-up schemes, the overhead of these is substantial. An alternative approach could be through the detection of global body-wide bio-signals, such as the heartbeat, an approach that is still in its infancy but holds great promises. Using the heartbeat as a synchronization signal offers obvious advantages: it frees the system from generating a synchronization signal and each node from embedding a high accuracy timing block, usually bulky and powerconsuming. Since the heart is the unique beacon generator, there are no synchronicity concerns other than its propagation delay. Moreover, state-of-the-art heartbeat detectors reach energy efficiencies that make this approach appealing in an ultra-low-power setting.
Heartbeat synchronization is at the core of the H-MAC protocol, introduced in [6], but this protocol limits its topology to a star network and trades off robustness and flexibility for device battery life extension. To the authors' knowledge, [6] and [7] were the only two attempts to use heartbeats as a synchronization signal until today. Reference [7] takes advantage of the heartbeat to synchronize and enable data exchange between different types of nodes. It uses the heartbeat as a central clock for all the nodes of the network. This method, if exploited properly, provides a reliable always present zero-energy (for the network) synchronization signal. If a node cannot efficiently sense the heartbeat, a local or WBAN-global lighthouse approach can be imagined with a node broadcasting the heartbeat signal. This topic is not covered in this paper.
In a related effort, [8] proposes a beacon-based synchronization MAC approach for the WBAN, which shares similarities with the present heartbeat synchronization capable context. It proposes a scheduled Time Division Multiple Access (TDMA) scheme, splitting frames into sensor-to-hub and hub-to-hub communication. However, the scheme targets two-tier star topologies, and the communication schedule is fixed. It assumes similar communication latency and volume for all sensor end-nodes around a given master node, which does not provide efficient support for various data rates and event-driven traffic.
The LDC-MAC protocol, introduced in [9], proposes a new beacon-based protocol that regroups essential features for a real WBAN: relatively low power consumption and effective cohabitation between low and high data rate traffics. To reduce energy consumption, LDC-MAC tends to use long superframes, in which regularly spaced slots are inserted to reduce the communication latency. However, a concurrent attempt to reduce power consumption in low-power communicating devices may lead to substantial clock drift [7], limiting practical superframe lengths. Heartbeat synchronization outperforms this trade-off.
Both IEEE 802.15.4 and 802.15.6 standards are not suitable for the targeted applications. Their non-beacon operation modes are mostly suited for low-traffic environments, not appropriate in our context [9]. On the other hand, their beacon operation mode suffers from the overhead addressed by LDC-MAC in [9]. Furthermore, both standards specify frequencies of operation and bandwidths which do not reach the Human Body Communication characteristics highlighted in [10].
Heartbeat synchronization presents some unique advantages as the synchronization beacon is globally available and not system generated. By nature, it addresses issues present in beacon-enabled protocols [11], such as the need for a central time coordinator or expensive distributed synchronization mechanisms. Indeed, heartbeat synchronization frees the network from a complex organization and selection process to define a node in charge of broadcasting the synchronization signal. This increases the system's robustness by not relying on a unique node in case of a failure. It also enables a completely distributed mesh operation, without the need for additional beacon schedule management. Furthermore, heartbeat-based synchronization does not suffer from any range limitation. Ultimately, no bandwidth allocation is necessary for broadcasting the synchronization signal, a desired feature while dealing with single-channel PHY such as C-BCC.
However, heartbeat-based synchronization comes with its specific challenges. The first is the heartbeat period's local and global variability. Local variability corresponds to the difference between two consecutive heartbeat interval durations. It is modeled in the present work as a normal distribution with a standard deviation of σ = 30 ms [12]. Global variability refers to the heart rate (HR) upper and lower bounds. It is assumed to fall in the range from HR min = 36 beats per minute (bpm) to HR max = 210 bpm [13], [14]. The second challenge is the non-negligible heartbeat signal propagation time. Its propagation speed is at least v HB >250 m s −1 [15]. It requires device calibration in addition to substantial guard intervals. Ultimately no MAC protocol is today suitable to address the requirements of modern and future WBANs while benefiting from a heartbeat-based synchronization.
To accommodate the cohabitation between distinct classes of data rates, a TDMA protocol supporting a puncturing feature is adopted in [7]. The main high data rate traffic (i.e. between primary nodes) is periodically interrupted allowing secondary nodes to upload their data. The principle is illustrated with a simple 2-node example in Fig. 1, and forms the base for the TDMA scheme adopted in this paper.
The proposed work presents HB-MAC, a flexible and realistic MAC protocol for the Human Intranet. It effectively supports a wide span of high and low data rates as well as event-driven sensor traffic and emergency signaling. HB-MAC has been designed to address the intrinsic challenges of heartbeat synchronization, taking advantage of this synchronization to lower the global energy consumption.
This paper is organized as follows: Section II presents the Human Intranet and physical layer requirements. Section III presents the proposed MAC protocol. Section IV presents metrics and simulation results. Section V discusses these results and compares them to prior art. Finally, Section VI concludes this work.

II. HUMAN INTRANET NETWORK REQUIREMENTS A. THE APPLICATION REQUIREMENTS
The Human Intranet is meant to interface a wide variety of wearables. From [16], [17] and [18], the amount of data generated for different kinds of sensors (e.g. ECG, Temperature, motion. . . ) is evaluated and synthesized in Table 1. From Table 1, the need for two types of nodes can be identified: those generating a limited amount of data, such as heart rate and temperature monitoring, and those generating a significant quantity such as motion sensing or neural recording which can stream around 25 Mbit/s for a 64-channel implant.
We define those two types of nodes as hubs and leaves [7]. The former offers greater computing capabilities and energy availability, whereas the latter is a lot more constrained and is energy-frugal.
In terms of behavior and role within the network, a leaf is connected to a unique hub, also called parent hub. A hub, however, can be connected to multiple leaves and hubs at the same time (see Fig. 2).

B. THE PHYSICAL LAYER
There are many options in developing a wireless network that spans the human body, each of which has its advantages and challenges. Often heterogeneous approaches combining multiple communication modalities are required. Considering the needs formulated in the paragraph above, one option that addresses both the upper and lower performance requirements is the use of Human-Body Coupled Communication in the 400 to 500 MHz band. While this choice is not essential, it helps in providing a unifying perspective.
In [10], it was established that a C-BCC solution operating in the 400-500 MHz band (with a center frequency of 450 MHz) offers some distinct advantages. At these frequencies, surface waves are the dominant propagation mechanism, making the channel less sensitive to changes in the environment and to coupling with external references. Its form factor is smaller and more convenient for the end-user than antenna-based RF technologies. Its attenuation per unit distance is also lower [19], at the cost of higher insertion loss. C-BCC matches the needs set by the applications and the developed scheme is tuned given C-BCC specifications. It is adaptable to other technologies, not covered in this paper.
From measurements [10], it was established that a bandwidth of at least 100 MHz is available with flat attenuation. The derived channel loss model is presented in (1), and its parameters as measured at 450 MHz are enumerated in Table 2. σ is the standard deviation of the measurement compared to the theoretical fit, explained by the non-homogeneity of the medium (i.e. the human body). L 0 expresses the loss for the shortest distance recorded at d 0 =5 cm. L 1 represents the loss at d 1 = 15 cm, transition distance between the quasi-static (QS) and surface wave regimes.
To take full advantage of such bandwidth, pulse-based communication is envisioned. It allows an aggressive transmitter duty-cycling leading to energy-efficient data transmission. Based on the analysis from [20], for a windowed-cosine pulse model, a 20 ns Gaussian pulse in the time domain corresponds to a 150 MHz-bandwidth signal at −10 dBc in the frequency domain.
With a simple modulation such as OOK or BPSK transmitting 20 ns pulses, the maximum theoretically achievable data rate is 50 Mbit/s. Given a similar architecture as presented in Fig. 2, the type and amount of data exchanged from Table 1, the channel characterization results from [10], and the pulse-based communication centered on 450 MHz, the hub-to-hub communication data rate is set to 50 Mbit/s with BPSK modulation scheme. The leaf-to-hub link is however set to 100 kbit/s with OOK modulation scheme. It offers a good trade-off between the transmission time of a given message and the achievable transmitter energy efficiency, relaxing the constraints on the leaf timing.
The receiver sensitivity S is calculated in (2) as a function of the integrated thermal noise N th over the bandwidth of consideration BW (see (4)) [21].
where NF is the receiver noise figure and SNR req is the required signal-to-noise ratio expressed in (3) as a function of the desired data rate DR.
N th = 10 · log 10 (K · T · 10 3 ) + 10 · log 10 (BW ) (4) For a usual noise figure of NF = 6 dB, a bandwidth of BW = 150 MHz, a ratio E b /N 0 = 14 dB and a data rate of DR = 100 kbit/s and DR = 50 Mbit/s, the sensitivity becomes respectively: In terms of geometry, the leaf-hub distance is limited to 15 cm, which means from equation 1, that the signal attenuation is not greater than 55 dB. Given a receiver sensitivity of −104 dBm and a 3 dB margin, the minimal output power the leaf transmitter should radiate is about −46 dBm. The inter-hub range is set to 50 cm (about an adult arm length) allowing a large coverage of a human body. Applying the same calculation, the hub transmitter output power should be around −10 dBm.
The above-listed network details regarding the communication mechanism, distance covered, channel attenuation, signal modulation, and transmitted power set the constraints affecting the HB-MAC protocol.

III. COMMUNICATION & MAC LAYER PROPOSAL A. MODEL & ARCHITECTURE
The proposed MAC scheme, HB-MAC, is a TDMA-based scheme adapted for two-stage networks. Each hub surrounded by its leaves forms a structure called a cluster, while the hubs are connected in a mesh. The considered network topology is thus, a mesh of stars. Given the substantial fading over distance that comes with Body Coupled Communication, it is assumed that intra-cluster communications (i.e., between a hub and a leaf) do not interfere with intra-cluster communications in other clusters. Thus, at any time, all leaves in the network have at most one parent hub.
While clusters are formed by positioning leaves around hubs, HB-MAC is designed to be adaptable. The link is dynamically adjusted to mitigate for body dynamics. This approach brings system flexibility and robustness against connection/disconnection. It also offers the ability to add, move or remove a leaf from the network during runtime without impacting the other devices.
Leaves can interact with their parent hub in two different ways: • In detached mode, a leaf communicates opportunistically upon the occurrence of an event. Detached communication can also be used to initiate an attachment procedure.
• In attached mode, a leaf is connected to its hub for a longer term. It reduces the overhead by avoiding random-access procedures. Moreover, an attached leaf has a well-defined location in the network and can consequently be addressed more effectively by distant hubs. It is further assumed that all devices embed heartbeat sensing capability and the relative detection time offset has been calibrated. This offset is caused by a non-negligible heartbeat signal propagation time [15].
HB-MAC uses superframes as its main structure. Superframes are defined as the time interval between two heartbeats. Since the signal skew is dependent on the device location on the body, guard intervals are introduced. The present protocol uses two different forms of superframes: • The regular superframe is the most represented and is shown in Fig. 3. This superframe type only allows attached and inter-hub traffics, using a protocol based on existing standards for the latter.
• The detached superframe allocates time to both detached leaves and hub-to-hub communication as shown in Fig. 4. Detached superframes are less frequent than regular superframes, as their purpose is to serve rare or event-driven traffic, including emergency communication and management procedures. A detached superframe may appear for two different reasons. It occurs at pre-defined rates corresponding to typical leaf latency requirements. The second possibility is in case of an emergency. When a leaf calls for an emergency (as described in Section III-B), a detached superframe is started both for the concerned hub and its direct neighbors.  In this second case, the detached superframe is referred to as an emergency superframe.
Except in the emergency case (either triggered locally by a leaf using a dedicated slot or initiated by a hub), detached superframes occur simultaneously over the whole body. The detached superframes' phase and period of occurrence are agreed by the hubs. This could be accomplished by any distributed algorithm, and in the most straightforward case simply decided by a mother hub. This scheme allows for dynamic reallocation in case of a workload change.

B. SUPERFRAME PREAMBLE
All superframes, regardless of their type, start with the same preamble scheme. The preamble's purpose is twofold. First, it provides an interface for low-latency emergency signaling. Second, it allows leaves to know when the next detached superframe will occur. Its structure is illustrated in Fig. 5.

1) LEAF ALARM SLOT
It is crucial for a WBAN MAC protocol to support low-latency emergency signaling, allowing data upload with minimal delay. This must have minimal impact on the regular operation of the network.
It is implemented by reserving an emergency slot immediately after the post-heartbeat guard interval. This is an analog slot: if a hub senses energy, then it enters the emergency mode. Precisely, the latter hub propagates a specific signal, sets its countdown value (described below) to the emergency value, and enters an emergency superframe.
This emergency scheme is robust against cardiac arrest, as any hub is capable of sensing the heartbeat. If the elapsed time from the last detected heartbeat exceeds a predefined limit, it reacts accordingly.

2) ALARM PROPAGATION SLOT
An emergency state cannot be strictly local to a given cluster. First, an emergency event is likely to be a global event happening within the human body, hence is likely to be triggered in different clusters simultaneously. Second, all neighboring hubs must be aware of the emergency mode to avoid interfering with the unpredicted communication between the hub in alert and its leaves.
An alarm propagation slot is thus implemented after the alarm trigger slot, which is intended for hub-to-hub emergency noticing. A hub sensing a signal during the alarm propagation slot enters an emergency superframe as well.

3) LEAF COUNTDOWN SLOT
Not requiring any attachment procedure is a beneficial feature when targeting ultra-low power consumption on the leaf side. This is enabled by the introduction of a countdown slot. It also gives leaves the ability to detect a parent hub change.
Periodically, hubs broadcast a countdown value to their leaves. The countdown message carries the hub address along with the number of heartbeats separating the current superframe from the next detached superframe. Transmitting the hub address ensures that a leaf communicates with the intended hub, especially in attached mode. The countdown value is typically encoded as an 8-bit unsigned integer, since the highest representable value corresponds to at least 73 seconds, for any heart rate in the considered range. According to the typical data generation rates given in Table 1, this upper-bound duration is sufficient. A specific countdown value indicates that the current superframe is in an emergency state. The countdown slot can be extended to contain additional information, such as the random-access strategy to adopt by detached leaves for instance.
The countdown system allows a leaf to hibernate for an indefinite amount of time, turning off communication and heartbeat sensor, and to opportunistically enter the network without other overhead. This is of particular interest not only for long hibernations but also to minimize the overhead of leaf re-attachments due to changes of the network topology, typical for Human Body Communications.
Finally, the countdown slot is also an acknowledgment informing a potential alarm-triggering leaf whether the alarm signal has been received and accepted by the hub.
The regular superframe is divided into two parts as depicted in Fig. 3. The first part is dedicated to communication with attached leaves during the Attached Leaf Guaranteed Time Slots (ALGTS), whose internal structure is application dependent.
The second part is dedicated to inter-hub communication. Its precise specification is not crucial to the present work. We propose to divide this part into multiple smaller frames similar to the 802.11 PSM protocol [22]. Each frame includes a reservation sub-window following a CSMA/CA scheme, materialized by the use of ATIM (Ad hoc Traffic Indication Messages) time slots, and then a HGTS (Hub Guaranteed Time Slots) sub-window. This structure has several advantages. First, it avoids the idle listening overhead from random-access schemes such as simple or slotted Aloha. Second, keeping the frames short improves both throughput and reliability. The smaller the frames, the faster the packets can flow across the entire network since a packet cannot perform more than one hop per frame. In addition, longer frames are naturally more prone to be affected by channel changes, common for body deployed networks [23]. Finally, as shown further below, collisions between concurrent communications are inevitable and may become frequent in some configurations. The choice of a division between reservation and guaranteed slots reduces the collision's energy cost [24].

2) HEARTBEAT VARIABILITY MANAGEMENT
One important parameter impacting the scheme efficiency is the length of the guard interval at the end of the superframe illustrated in Fig. 6. It is required to compensate for the heart rate local variability. The standard deviation between two successive heartbeat signals is 30 ms. The heartbeat occurrence can be predicted based on the previous superframe. Overestimating the available time can lead to packet failure and increased energy consumption.

3) RE-SCHEDULING PROCEDURE
Leaves should be reschedulable. During a hub's ALGTS window, neighboring hubs are not allowed to communicate with other hubs. Thus, it is beneficial to locally re-schedule leaves to minimize the hub idle duration due to neighboring hubs still communicating in their ALGTS windows.
The rescheduling algorithm requires acknowledgments from the reallocated leaves. The time spent waiting for all the required acknowledgments through successful ALGTS is referred to as the rescheduling latency and can cause a global leaf rescheduling procedure to last long. In the worst case, hubs need to wait for all the leaves' attachments to time out. Moreover, rescheduling requires the use of additional temporary ALGTS slots.

D. DETACHED SUPERFRAME 1) GENERAL STRUCTURE
Only detached superframes allow detached leaves (i.e. leaves which haven't completed a still valid attachment process) to communicate. They allow leaves to communicate with the hub in range and then let the hubs communicate with each other for the remaining duration of the superframe. The uplink traffic is privileged. To receive downlink data, a leaf has to initiate the communication with the hub, which is a common decision among energy-efficient protocols, such as LoRaWAN Class A [25]. The detached superframe structure shares similarities with the regular superframe structure. It substitutes the ALGTS window with a detached leaf communication window, illustrated in Fig. 4.

2) ATTACHMENT PROCEDURE
Only an overview of the attachment mechanisms is provided. It underlines the challenges faced by such procedures. The attachment procedure is performed during a detached superframe. Being attached provides two advantages: (i) facilitated communication with the hub, and (ii) a fixed address in the network, allowing data to be routed to the leaf through the WBAN.
To get attached, a leaf sends to the hub the preferred period of communication measured in superframes, its communication class which defines the ALGTS internal structure (e.g., actuator-type communication or sensor-type communication), data related to the communication class (e.g. its data generation rate for a sensor-type communication) and finally a list of identifiers dealing with the types of data to be addressed to the leaf by other devices in the WBAN. The hub immediately responds with information regarding the scheduled phase and period, as well as the dedicated ALGTS.

3) DETACHED LEAF COMMUNICATION WINDOW
A detached leaf communication window is illustrated in Fig. 4 as part of the detached superframe. It is structured with an initial CSMA/CA sub-window, followed by a sub-window of guaranteed time slots. The first sub-window contains Leaf Communication Request Time Slots (LCRTS), each followed by an acknowledgment slot for the hub. These Leaf Communication Requests (LCR) sole purpose is to reserve the following Detached Leaf Guaranteed Time Slots (DLGTS), which are followed by an acknowledgment slot.
The size of both LCRTS and DLGTS sub-windows is later addressed in this paper. It depends on the number of leaves in a cluster and on the target leaf energy consumption. The latter trades off with the time remaining for inter-hub communication. Typically, in an emergency superframe, the dimensions of the detached leaf communication window are larger than during regular operation to maximize the probability of communication success, as the leaf uplink traffic may brutally increase.
Locating leaf-to-hub communication at the beginning of the detached superframe offers several advantages. First, it increases the likelihood that the leaves can still communicate with the hub, once they have read the zero-countdown value (described in Section III-B) since WBAN does not guarantee a constant channel quality between two fixed devices on the body [23], [26]. It also serves as a detached superframe occurrence confirmation and Quality of Service (QoS) information. Second, leaves may be equipped with inaccurate clocks [7]. Hence, communicating early in the detached superframe avoids excessive clock drift accumulation since the clock has been reset by the heartbeat detection. Finally, this detached superframe arrangement allows information to be spread as deep as possible in the network, and thus offers relatively low end-to-end latency.

4) POST DETACHED LEAF COMMUNICATION WINDOW
Once the leaf communication window is complete, the scheme proceeds to inter-hub communication. Typically, no ALGTS is inserted behind the detached leaf communication window for two reasons. First, the length of the detached leaf communication window is not fixed. It is allowed to temporally vary to accommodate the workload, as long as it has the same global value spatially across the network. Second, scheduling a consecutive attached communication window without a substantial risk of interruption by the next heartbeat would be too constraining.
It is not always satisfying to forbid collisions between leaves granted ALGTS (in terms of phase and period of ALGTS occurrence) and detached superframes. The detached window period may vary, as long as it remains global to the WBAN. Leaves could be rescheduled accordingly, but rescheduling latency could substantially delay the modifications of the detached communication period. Ultimately, enforcing that no ALGTS ever coincides with a detached superframe, is equivalent to ensuring (5): where T A represents the scheduled ALGTS period, φ A the scheduled phase with reference to any occurrence of a detached superframe, and T D the detached superframe period measured in heartbeat intervals.
According to the value of T D , this condition may become restrictive, as it forbids, for instance, T D to be prime with T A . This may lead to choosing detached superframe periods accordingly.
If a collision occurs, either because the condition given in (5) has knowingly been infringed or because of an emergency event, leaves can either wait for their next ALGTS, or communicate with the hub through the detached communication window, at the cost of a random-access. Such collisions between ALGTS-containing superframes and detached superframes are dynamically detected by attached leaves, as they read the countdown value right before proceeding to attached guaranteed communication.

IV. METRICS AND SIMULATION RESULTS
Within this section, the HB-MAC protocol performance is evaluated. We focus on the detached mode as this is where the innovation is: the attached leaf operation and the inter-hub communication rely on existing protocols. First, the dimensions of the leaf reservation window will be analyzed and optimized in terms of energy cost and communication failures. The average detached leaves' ON-time is derived, and the impact on the hubs. Finally, a heartbeat detector duty cycling strategy, proper to this MAC protocol, is proposed to further improve the leaf energy efficiency.
The protocol parameters used in the analysis are given in Table 3 and Table 4. The overhead of the physical header consists of a preamble, a frame delimiter, and a CRC of sizes 7, 1, and 4 bytes, respectively. Leaf and hub addresses have sizes of 32 and 16 bits respectively. Guard intervals of 1ms are used for communications with leaves, sufficient to mitigate their clock inaccuracy over one heartbeat interval [7]. The remaining values are calculated based on the data rates and packet structures.

A. DETACHED COMMUNICATION STRUCTURE 1) CHALLENGES AND DEFINITIONS
The detached leaf communication window dimensions result from a trade-off between the leaf communication success probability, hub energy consumption and inter-hub communication time. The number of DLGTS per detached leaf communication window should match the predicted number of leaves whose LCR succeeds. The corresponding negotiation window size calculation, on the other hand, requires attention. The LCR communication success rate, from a detached leaf to its parent hub, increases with the number of LCRTS available per superframe. However, an oversized LCR sub-window substitutes hub sleep or inter-hub communication time.
The number of leaves competing for simultaneous access to a specific hub, as well as the LCR random allocation strategy, significantly impact the negotiation success rate and energy efficiency. Given the energy scarcity, a leaf-dedicated power management strategy is required.
We define the failure rate (and its success rate counterpart) as the probability of unsuccessful leaf-to-hub communication, within a given LCR sub-window. Besides the physical channel variations or interference with the environment, the major source of transmission failures is the successive collisions during the LCR slots. It is the only element considered in this study. As there are no more than 80 bits sent per leaf per LCR, excluding the physical-layer preamble, and assuming a bit error rate of 10 6 /bit [27], the failure probability of a LCR message due to physical deterioration is lower than 8 · 10 −5 .
In the following, several probabilistic LCRTS allocation and back-off strategies are studied in terms of success rate and energy consumption, from a detached leaf standpoint. The energy consumption dedicated to random-access increases proportionally with the number of LCR messages sent by the leaf in a detached superframe.
This study focuses on symmetric strategies: all detached leaves trying to concurrently access the same hub follow the same rules. Since the set of symmetric LCRTS allocation and back-off strategies is large and difficult to parametrize, we only consider a set of common slot allocation families.

1) Uniform without back-off (UBS):
Each leaf chooses uniformly and independently the LCRTS to send a communication request, within the entire LCR subwindow. If it fails, no back-off is performed. This strategy does not take any parameter. 2) Uniform with back-off (UB): Similar to UBS, except in case of a communication failure. In this case, the leaf individually re-applies the same strategy on the remaining LCRTS in the current LCR sub-window. This strategy does not take any parameter.

3) Fixed contention window size (FCS): the LCR
sub-window is divided into contention windows of equal size. In each contention window, the leaves that have not successfully sent an LCR message yet, select uniformly one LCR slot within the contention window. This family of strategies takes one parameter: the contention window size.

4) Binary exponential back-off (BEB):
Similar to FCS, except regarding the contention window size. It doubles at each back-off and is bounded by a given value. This family of strategies takes two parameters: the initial and the maximal contention window size. Both families, FCS and BEB, can be implemented following two behaviors, illustrated in Fig. 7:

1) Exceeding bound (EB): If the last contention win-
dow exceeds the LCR sub-window, then the remaining leaves still willing to communicate choose their slot within the entire contention window (instead of restricting to the remaining LCR slots). If a leaf chooses a slot beyond the last LCRTS, then it cancels its communication attempt and typically waits for the next detached superframe.

2) Containing bound (CB): If the last contention window
goes beyond the LCR sub-window, then the remaining leaves attempting to communicate choose their slot only in the remaining LCRTS. UBS is a particular FCS parametrization, which is itself a particular BEB parametrization. In the following they are considered distinct to underline their specificities and simplify the strategy when possible.

2) PERFORMANCE EVALUATION
The various options are mathematically evaluated here. In the following formulae, the integer F represents the number of concurrent detached leaves trying to access the considered hub. The integer S represents the number of slots in the LCR sub-window, and M is the expected number of LCR messages sent by an involved detached leaf during the LCR sub-window. Since all considered strategies are symmetric among leaves, we can consider a fixed leaf F1 without loss of generality.

UB:
The success probability is recursively calculated considering a partition of five cases happening on the first LCR slot, as indicated in Fig. 8. The evaluated number of messages sent uses the same event set partition and the observation that F1 sends a new message in the cases (i) and (ii) and not in the other cases. Using these partitions, we obtain the success probability P s,UB and expected number of messages M UB sent by a candidate leaf during a LCR window given by (7a) and (7b) respectively. else.
(7b)  (8). This serves as a tool to evaluate the number of LCR slots with unsuccessful LCR communication.
The second step counts the number of configurations Conf k (i.e., allocations φ of LCR slot indices to leaves) such that exactly k leaves achieve successful LCR communication with the hub. The number of possible configurations is calculated in (9).
From there comes the probability P k,f expressed in (10) that k concurrent leaves achieve successful LCR communication with the hub during the given LCR window and that F1's LCR communication fails.
At this point, a partition is conditioned on the result of the first contention window. It allows to recursively compute the success probabilities P s,BEB in (11) and expected message numbers M BEB in (12) for the four remaining schemes. The considered partition is the event set: {F1 succeeds during its first contention window} where: η := min(2cw, cw max )

3) RESULTS
HB-MAC aims at efficiently supporting detached communication. Focus is put on transmission reliability and energy cost. To pursue this objective, optimal strategies have been derived. Additionally, robustness also needs to be ensured, as the number of detached leaves attempting to access a given hub is difficult to anticipate. For readability purposes, we systematically consider the optimal parameters for each family of strategies in this section. The decreasing priority has been set as UBS > UB > FCS/CB > BEB/CB > FCS/EB > BEB/EB. If several strategies offer identical performance, only the highest-priority strategy is given.
Derived from the previous analyses, the optimal LCRTS allocation and back-off strategies are illustrated in Fig. 9 for a minimal failure rate. This optimization provides a lower bound for the achievable failure rate due to collisions, along with the minimal expected number of messages per leaf to achieve it, as illustrated in Fig. 10. In particular, we observed that for nine leaves trying to concurrently access the same hub, the expected number of LCR messages sent in the LCR sub-window should never exceed 3.3 per concurrent leaf.
While the previous consideration focused on the failure rate as the optimization objective, another approach consists  of optimizing the number of messages sent by a detached leaf under failure probability constraint. This approach is more realistic, since the HB-MAC aims at ensuring low energy consumption, whereas minimizing collision probability significantly below the physical transmission failure rate is practically pointless. Fig. 11 illustrates optimization results minimizing the expected message number under a failure rate constraint of 1%. Backed with more detail by the quantitative analysis, it shows the robustness of the optimization: optimal families of strategies remain stable for small variations in the number of LCRTS or concurrent detached leaves. The analysis proves that their parameters vary smoothly as well. Fig. 12 shows the actual minimized number of messages under the same failure rate constraint. Optimal spots appear, located at the slopes' inflection points. As their location is related to the number of concurrent leaves, they should be considered with the proper margin to propose efficient LCR window sizings.
As a numerical example, if a Human Intranet application around a given hub requires a failure rate below 1% and if 3 concurrent leaves try to access the hub during the same detached superframe, Fig. 12 shows that a LCR subwindow size of 30 LCRTS, reaches such an optimal point. Fig. 11 shows that the corresponding strategy is a uniform allocation with back-off.
The detached leaves' energy consumption is approximately proportional to the number of messages sent per  LCR sub-window. In many cases, the choice of a sufficient number of LCR slots brings the message number below 1.2 per concurrent leaf even for dense clusters containing as many as nine detached leaves.
Those results show not only efficiency in terms of failure rate and energy consumption, but also the robustness of the detached operation. First, we demonstrated the existence of optimal slot allocation and back-off strategies that enable efficient detached operation for various numbers of LCR slots and for large numbers of concurrent leaves. Second, we showed that the optimal strategies and their parameters vary smoothly. This smoothness is crucial to ensure that optimal strategies for a given predicted number of concurrent detached leaves remain efficient as the actual number varies.
Ultimately, one major objective of leaf attachment is to reduce the load on detached leaf communication windows. Only leaves with rare or irregular traffic, as well as leaves initiating attachment procedures and attached leaves requiring extra communication, should communicate in detached mode. This brings further down the actual expected number of LCR messages to send and the failure probability. superframe preambles and detached leaf communication windows. Extra hub energy is spent among four actions: (i) listening during the alarm propagation slot; (ii) sending periodic countdown information; (iii) listening during the entire LCR subwindow and, (iv) communicating during the DLGTS windows.
The average hub communication time overhead due to detached operation is highly dependent on the detached superframe period and is represented in Fig. 13, considering the numerical values listed in Table 4 and Table 5. The communication time overhead due to distributed operations (such as modifying the detached superframe period or advertising LCR subwindow sizes for modifications) is not covered in the study. It is worth noticing the large contribution of the LCR windows, estimated in the worst-case scenario (where the hub sends an acknowledgment at each LCR slot). The important contribution of LCR slots makes the hub time overhead sensitive to variations in size of the LCR window. Adapting the LCR window length to the expected detached traffic will have a substantial effect by allowing an efficient trade-off with hub communication time overhead dedicated to detached leaf management.

2) DETACHED LEAF INTERNAL OVERHEAD
A major strength of HB-MAC is the detached leaves' low energy consumption. It is brought to a very low value by minimizing the communication time. Detached leaves only communicate to (i) read countdown information; (ii) attempt LCR requests and listen to the corresponding acknowledgment, and (iii) during the corresponding DLGTS.
Even if the leaves can read the countdown slot only once per detached superframe while maintaining a functional structure, a second countdown read operation is considered. The second countdown read offers strong benefits such as channel quality assessment, ascertaining that the hub is currently reachable. The optimizations performed in Section IV-A resulted in a small expected number of LCR messages. Calculations were made for 1.1 LCR message sent, according to Fig. 12 for three concurrent detached leaves. The resulting average leaf ON-time is represented in Fig. 14. As most of the LCR communication overhead has been moved to the hub, it takes a minor part of the total detached communication time. Moreover, the countdown and LCR overhead due to countdown reading and LCR communication can be brought to arbitrarily low values by tolerating higher communication latencies for a given detached leaf, granting very low energy consumption for sparsely communicating leaves.
Additionally, Fig. 14 shows the overhead of a detached operation compared to an ideal attached operation for different detached traffic sparsities. For instance, a leaf communicating every 20 superframes and using guaranteed time slot payloads of 4 kbit, in concordance with realistic data rates summarized in Table 1, would reduce its communication time by 9.10% by running an attached mode. Attachment would additionally reduce the load on reservation slots. Different leaf communication requirements are addressed by attached and detached modes. While leaving the precise attached operation aside in this study, HB-MAC introduces an efficient detached operation dedicated to event-driven and rare traffic. The protocol aims at keeping the detached overhead low, focusing on leaves' energy consumption, and we showed that it can efficiently be adapted to accommodate various energy management strategies, crucial for realistic WBAN implementations.

C. FULL CLUSTER SIMULATION 1) MATERIALS AND METHODS
In this subsection, we quantitatively compare HB-MAC with IEEE 802.15.4 [5] in a well defined and realistic scenario. The performance evaluation is conducted by simulating both protocols with Castalia [29].
The simulation focuses on the intra-cluster operation, as this is HB-MAC's most innovative element, and this topology is simple enough to provide a fair performance comparison. Its aim is twofold. First, it demonstrates the performance of HB-MAC in terms of latency and energy consumption  Table 4 and Table 5.
in comparison to the widely-used IEEE 802.15.4 standard. Second, it validates the effective operation of HB-MAC in a realistic setting including diverse data generation rates.
We simulated a single cluster composed of a hub surrounded by 6 leaves, supporting applications with diverse purposes summarized in Table 6. The packet generation rate is an application layer parameter only, independent from the MAC layer.
A communication period is defined as the time interval between two leaf-hub data exchanges. An attached leaf selects its own communication period during the attachement procedure. However, detached leaves rely on the detached superframe period (DSP), a network-wide setting.
The chosen parameters for the HB-MAC protocol are listed in Table 7 (except otherwise stated).
The chosen parameters for the IEEE 802.15.4 standard are listed in Table 8: The physical layer configuration is the same for both protocols. The associated power consumption are listed in Table 9.    The simulation starts with all the leaves in a reset state: they have no data to transmit and are unaware of their surroundings. It runs for a duration of 6000 s. Figure 15 shows the energy efficiency per useful bit (i.e. payload bit) delivered, for each leaf node in the network. Both the IEEE 802.15.4 standard and HB-MAC are simulated with a heart rate varying between 40 and 160 bpm. Overall, compared to the IEEE 802.15.4 standard, HB-MAC lowers the per-bit consumed energy by a factor between 12x and 16x depending on the considered leaf node. In addition, both protocols become more energy-efficient as the sensor data generation rate increases. This is explained by the minor contribution of packet transmission to the total energy consumption. The latter is dominated by idle listening or sleeping. Consequently, increasing the data generation rate improves the energy efficiency. Heart rate variation has a negligible impact on the energy efficiency.

2) SIMULATION RESULTS
Comparing performance in pairs of nodes with identical data generation rates allows to estimate the influence of the other communication parameters. Observing nodes 1 and 2, the emission of more frequent but smaller data packets improves the energy efficiency of IEEE 802.15.4, and degrades that of HB-MAC: smaller packets imply more communication overhead. The radio also consumes less in transmit than in receive, which is the default mode for IEEE 802.15.4. Finally, transmitting more decreases the energy consumption. For HB-MAC, the radio is in sleep mode by default, so more energy will be consumed with smaller packets.
Contrasting nodes 3 and 4 tells us that attached operation incurs a higher cost to IEEE 802.15.4, but slightly improves energy efficiency of HB-MAC. The overhead of the former comes from idle listening in both detached and attached mode during the whole random-access period, even though the attached leaf will communicate in its GTS. The slight improvement of HB-MAC is due to not having to reserve a slot before every guaranteed communication. Comparing nodes 5 and 6 shows that increasing the communication frequency of an attached leaf increases HB-MAC's consumption, but the additional cost is relatively low (< 8%). Figure 16 presents HB-MAC's latency statistics as a function of the heart rate and the chosen detached superframe period. The latency of IEEE 802.15.4 is also given for comparison. This analysis is done for a hub surrounded by three identical detached leaves, generating 120 bps. The error bars correspond to the interval [x − σ,x + σ ], withx the average latency and σ the standard deviation.
The average latency with HB-MAC can range from 1 to 15 s, which is significantly longer than the sub-1s latency result obtained by 802.15.4. The obtained result also greatly varies with heart rate and communication frequency. As expected, the observed latency increases with the detached superframe period, since a longer period means a lower frequency of communication for the detached leaves. A higher heart rate implies shorter superframes and thus increased frequency of communication when considering a fixed detached superframe period: latency is decreased with physical activity of the user.
The simulations set up in this subsection clearly show that HB-MAC improves the leaf energy efficiency significantly when compared to a wide-spread standard such as IEEE 802. 15.4. This comes at the price of a generally degraded latency, although tighter latency requirements can be met by operating in attached mode or using the emergency signaling mechanisms.

D. HEARTBEAT DETECTOR MANAGEMENT
Integrated circuit heartbeat detector state-of-the-art implementation reaches a power consumption as low as 58 nW [31]. Regardless of this ultra-low consumption, it remains non-negligible for a leaf given the targeted application. This section examines a duty-cycling scheme that aims to decrease the leaves' heartbeat detection energy cost. The envisioned duty cycling strategy consists of turning ON the heartbeat detector a short period of time γ ahead of the next expected heartbeat as represented in Fig. 17. The expected inter-heartbeat duration is assumed identical to its previous occurrence. Two sources of uncertainty are evaluated: the leaf clock uncertainty and the heartbeat local variability. They are modeled following two random variables: X , representing the heartbeat detector start time; and Y , representing the actual heartbeat occurrence time. The time reference is set to the expected heartbeat time of occurrence.
X and Y are modeled with independent normal distributions N (µ X = −γ , σ 2 X = 4µs 2 ) and N (µ Y = 0, σ 2 Y = 900µs 2 ). The standard deviations are respectively chosen according to the maximal clock inaccuracy given in [7], and the heartbeat local variability described in [12]. Hence, the difference Y − X follows a normal distribution centered in γ of variance σ 2 X + σ 2 Y . Optimizing γ is essential. Chosen too big, the detector will start too early and energy will be wasted. Chosen too small (including negative), the detector will miss the heartbeat. The heartbeat detection failure cost is modeled with the heartbeat detector staying ON for an entire superframe.
In addition, the leaf will listen to the countdown slot following the next successful heartbeat detection. We calculate E cdwn = 154 nJ/read the transceiver energy cost for a single countdown read operation, taking into account half of the corresponding guard interval.
The expected ON-time ratio E[d on ] is detailed in (13). It assumes that the heartbeat detector can be switched ON and OFF without associated cost and that no heartbeat is missed when the heartbeat detector is ON.
For robustness purposes, a misdetection identification feature is required. If the current heartbeat duration exceeds an upper limit (set as multiple standard deviations σ Y ), the heartbeat is considered missed.
From a leaf standpoint, a missed heartbeat only has a limited impact, as long as it does not directly precede a superframe where communication was intended. A hub misdetection would however block the communication for the entire superframe. Therefore, hubs should not implement such an aggressive heartbeat detector duty cycling scheme and should prefer more robust and reliable schemes, relying on a lower bound of heartbeat interval durations. Fig. 18 shows the leaf's average heartbeat detector power consumption with the duty cycling scheme activated. It achieves an energy saving from 74.7% at 210 bpm up to 94.7% at 36 bpm in comparison to an always-ON heartbeat detector.

A. RELEVANCE TO THE HUMAN BODY
Based on the heartbeat as a system-wide clock and on C-BCC as a physical layer, HB-MAC has been designed for VOLUME 8, 2020 energy-efficient communication on the human body, interfacing leaves and hubs.
Physical channel interruptions between hubs are mitigated by the use of pre-existing MAC layer standards, out of the scope of this work. Intra-cluster channel interruptions are mitigated by using a countdown slot and scheduling the intra-cluster communication right after it in the superframe structure. Countdown slots have a triple interest. They detect a channel deterioration and permit leaf re-synchronization in case of heartbeat misdetection or long channel interruptions. Ultimately, they allow leaves to notice a parent hub change, which is particularly important in attached mode.
Section IV-A shows the protocol's resilience to unusual traffic loads, which may follow long channel interruptions, treated with reasonable failure rate and energy cost. To further minimize the failure rate in random-access procedures after channel interruption, leaves could adapt their random-access strategy to the estimated number of concurrent leaves. Those adaptation mechanisms are beyond the scope of this work.
The Attached operation is beneficial to upper-layer protocols for regular traffics, as it provides leaves with a specific address within the network. This is also applicable to detached leaves, but the address may be unstable. Downlink traffic addressed to leaves is a crucial feature for human body networks. It provides flexibility by allowing, for instance, sensor sampling rate and behavior update, providing better measurements and higher energy efficiency. On the other hand, efficient detached operation is a strong feature which efficiently supports event-driven and irregular traffic.
A WBAN supporting diverse applications inherently contains a wide range of devices and data rates as listed in Table 1. The HB-MAC protocol effectively supports the networking of devices with data generation rates ranging from some bits to multiple megabits per second. It integrates generic protocols into a framework tolerating detached leaf operation and supporting imbalanced device capabilities by design.
Finally, a WBAN must manage emergency events efficiently. Emergency mechanisms must have minimal impact on the regular network operation, but need a low response latency. HB-MAC implements an emergency signaling mechanism with a latency of not more than one heartbeat period and can detect heart failures.

B. COMPARISON
Multiple existing protocols address media access control in WBANs and WPANs. This section compares HB-MAC with the IEEE 802.15.4 [5], IEEE 802.15.6 [4] and Bluetooth Low Energy (BLE) standards [32], with H-MAC presented in [6], the low duty cycle protocol presented in [8] and the LDC-MAC protocol presented in [9]. The comparison is summarized in Table 10.
H-MAC is another MAC protocol based on the heartbeat as a universal clock, whereas [8] presents a protocol whose topology and device hierarchy relate to HB-MAC. LDC-MAC relates to HB-MAC as it supports cohabitation between devices featuring various data rates.
Most protocols targeting low-power operation rely on time division, often including random-access. Among the comparison references, only the protocols presented in [6] and [8] do not offer random-access, while LDC-MAC only offers a random-access period optionally, which is activated in case of a frame collision. The gain in terms of energy efficiency is effective when not offering random-access, however this design choice impairs flexibility. In particular, it does not feature efficient support for event-driven applications. Therefore, HB-MAC only uses random access in detached mode. In the present work, we characterized costs of random-access and sized it for different situations to offer flexibility at a reasonable cost. BLE, on the other hand, relies on frequency diversity, which is not offered by UWB-based physical layers, which attract a lot of attention [33].
Mesh topologies offer robustness and become necessary for propagating a message across the whole body when relying on lossy physical layers, unable to reach the human body endpoints in one or two hops [10]. The protocol presented in [8] takes advantage of the limited range to isolate intra-cluster communications. Despite the similarities with this work, [8] limits its topology to a star of master nodes (equivalent to HB-MAC's hubs). Each of those masters is simultaneously the center of a local star network, made of sensor nodes (equivalent to HB-MAC's leaves). HB-MAC extends this topology to a mesh of hubs, each surrounded by a star of leaves, although the mesh management is delegated to an upper network layer.
Beacon synchronization is usually used to lower the sensor nodes' energy consumption. Several protocols, such as the IEEE 802.15.4 and IEEE 802.15.6 standards, make these optional. Similarly to H-MAC, the HB-MAC protocol uses heartbeat signals as out-of-band beacons. In a heartbeat-synchronized network, devices do not need to listen to an in-band beacon, potentially bringing down the communication system idle listening. This heartbeat-based beacon outperforms the LDC-MAC trade-off between coordinator power consumption, which decreases with the superframe length, and communication latency.
As a major requirement of modern WBAN applications, HB-MAC supports a wide span of data rates, up to multiple Mbps between hubs, depending on the inter-hub protocol implementation and the trade-offs made between hub-to-hub dedicated time and leaf-dedicated time. Such high data rates are not typical for low-power WBANs, therefore the only comparison references featuring similar maximal data rates are the IEEE 802.15.6 standard [4] and to some extent [8] and [9]. HB-MAC is consequently compatible with more demanding applications such as neural recording or smart prosthetic control. It is capable of flawlessly supporting high peak demands.
Body dynamics can modify the network geometry. While strong path loss permits cluster isolation in reference [8] and HB-MAC, it poses a specific challenge. In the case of C-BCC, for instance, movements can substantially alter the network topology [23]. The former shows poor resilience to topology changes, as the TDMA schedule is fixed. When it comes to HB-MAC, the resilience of the inter-hub mesh has not been precisely addressed as it relies on the inter-hub MAC protocol. The intra-cluster protocol has however been designed to be resilient, as detached superframes are globally scheduled and sized. In addition detached access is the same in all clusters. The resilience of the attached operation is granted by the hub address field in the countdown message. A leaf detecting a change in the hub address goes into detached mode to initiate a new attachment procedure. It avoids interference with the new cluster. As opposed to [8] and HB-MAC, protocols with a simpler topology naturally endure device movements with more ease.

VI. CONCLUSION
In this paper, we presented HB-MAC, a MAC protocol dedicated to WBANs relying on a heartbeat-based synchronization and compatible with Body Coupled Communication. It addresses important challenges of body area networks by supporting devices with imbalanced capabilities, a wide span of data rates, including regular as well as event-driven traffic.
We first presented a candidate physical layer, C-BCC on ultra-wide-band, adapted to low-power operation on the human body. Then, we described the MAC protocol architecture. As many low-power protocols, HB-MAC relies on a time division between random-access and guaranteed time slots. It has been designed for efficient operation under a heartbeat-based universal clock and supports an efficient integration of lightweight devices with irregular traffic.
We quantitatively addressed the energy cost of random-access for the case of detached operation including attachment procedures and collision probability during random-access. We deduced the optimal random-access strategy depending on the number of available time slots and concurrent leaves denoting the total number of leaves in a cluster along with their traffic. We deduced that the integration of efficient random-access algorithms allows HB-MAC to efficiently support event-driven and irregular traffic.
We also evaluated HB-MAC's performance in comparison with the IEEE 802.15.4 standard, by performing a network simulation of a realistic Human Intranet scenario. The performance analysis shows that, at the cost of increased latency, HB-MAC outperforms IEEE 802.15.4 by a factor between 12x and 16x in terms of energy efficiency.
To keep reducing the power consumption, we introduced a heartbeat detector duty-cycling scheme for leaves, proper to HB-MAC, and estimated the energy cost of heartbeat detection for leaves.
As this paper focused on the most innovative part of HB-MAC, which is the integration of heartbeat as a universal clock and joint support for regular and event-driven traffic, the integration of precise protocols for attached leaf communication and inter-hub communication remains as future work. The development of the upper layers of the network, as well as efficient distributed algorithms providing additional flexibility in the HB-MAC protocol with low overhead, remain as future work as well. VOLUME 8, 2020