1

An implantable instrument for studying the long-term flight biology of migratory birds.

Robin J. Spiveya & Charles M. Bishopb

a r.spivey@bangor.ac.uk; b c.bishop@bangor.ac.uk

a,b Department of Biological Sciences, Bangor University, Gwynedd, UK. LL57 2UW

Abstract: The design of an instrument deployed in a project studying the high altitude Himalayan migrations of

bar-headed geese (Anser indicus) is described. The electronics of this archival datalogger measured 22x14x6.5

mm, weighed 3g, was powered by a ½AA-sized battery weighing 10 g and housed in a transparent

biocompatible tube sealed with titanium electrodes for electrocardiography (ECG). The combined weight of 32

g represented less than 2% of the typical bodyweight of the geese. The primary tasks of the instrument were to

continuously record a digitised ECG signal for heart-rate determination and store 12-bit triaxial accelerations

sampled at 100Hz with 15% coverage over each two minute period. Measurement of atmospheric pressure

provided an indication of altitude and rate of ascent or descent during flight. Geomagnetic field readings

allowed for latitude estimation. These parameters were logged twice per minute along with body temperature.

Data were stored to a memory card of 8GB capacity. Instruments were implanted in geese captured on

Mongolian lakes during the breeding season when the birds are temporarily flightless due to moulting. The goal

was to collect data over a ten month period, covering both southward and northward migrations. This imposed

extreme constraints on the design’s power consumption. Raw ECG can be post-processed to obtain heart-rate,

allowing improved rejection of signal interference due to strenuous activity of locomotory muscles during

flight. Accelerometry can be used to monitor wing-beat frequency and body kinematics, and since the geese

continued to flap their wings continuously even during rather steep descents, act as a proxy for biomechanical

power. The instrument enables detailed investigation of the challenges faced by the geese during these

arduous migrations which typically involve flying at extreme altitudes through cold, low density air where

oxygen availability is significantly reduced compared to sea level.

Keywords: flapping flight, accelerometry, electrocardiography, magnetometry, heart-rate, altitude

Introduction

Obtaining detailed physiological and biomechanical measurements from free-ranging birds has become

increasingly feasible in recent years due to advances in electronics and transducer technology [1-3].

Miniaturisation and power reductions have allowed instruments to be implanted into the abdominal cavity of

birds, recording heart-rate for many months [4-5]. Real-time QRS detection from the electrocardiogram (ECG)

signal is an error-prone process and notoriously challenging when weight, size and power consumption are

prime considerations [6]. The present design successfully implemented an alternative approach which has only

very recently become viable for year-long deployments: the storage of raw ECG and subsequent signal post-

processing of arbitrary sophistication to obtain robust measures of heart-rate following recapture. Since

immense amounts of memory are required, the possibility arises of simultaneously storing additional

information from other high data-rate transducers such as accelerometers. For animals whose recapture is not

impractical, this archival approach to data collection can collect vast biologically informative datasets,

overcoming the data restrictions associated with remote download systems.

2

Bar-headed geese (Anser indicus) are evolutionarily adapted to flight at high altitudes owing to the geologically

recent formation of the Himalayas cutting directly across their annual migration routes [7]. Although evolution

could have tackled this problem by favouring longer migratory routes to the east of the Qinghai-Tibetan

plateau, natural selection has instead conferred on this species an extraordinary ability to tolerate hypoxia.

While Brent geese (Branta bernicla) apparently struggle to cross the ice-caps of Greenland at 2500m [8], bar-

headed geese are able to fly comfortably at twice (and possibly even three times) that altitude. Anecdotal

eyewitness accounts by mountaineers climbing the treacherous Annapurna massif, which forms a natural 23km

wide barrier to goose migration, have reported bar-headed geese circling overhead at altitudes of 7000-8000m

[9]. However, such events may be relatively infrequent and dependent upon favourable weather conditions.

General results and two record altitudes (6540m and 7290m) from GPS satellite tags suggest that Himalayan

crossings more typically involve maximum altitudes of between 5000m and 6000m, with flight tending to

follow the valleys [10].

The detailed study of the complex migrations of birds such as bar-headed geese demands sophisticated, high

performance instrumentation. This work describes the design of a compact, ultra-low power miniature archival

datalogger for abdominal implantation. The instrument was first deployed during the summer of 2010, the

design having been undertaken over the course of the previous twelve months. The system was capable of

monitoring body temperature, atmospheric pressure, the Earth’s geomagnetic field, recording ECG

continuously and triaxial accelerometry periodically. Since the flight corridor of the bar-headed geese between

Mongolia and the Himalayan crossing into India happens to run parallel to the horizontal projection of the

geomagnetic field, magnetometry could permit estimates of latitude, even for an implanted system. In

addition, several systems were mounted on neck collars and deployed on Whooper Swans (Cygnus cygnus) in a

pilot study intended to investigate the practicality of intelligently managing acceleration data collection during

flight.

Mechanical Considerations

GPS tracking devices communicating results in real-time via satellite telemetry had established by 2009 that

bar-headed geese do not all spend the winter in India with a minority content to settle into valleys in

southeastern Tibet [11-12]. However, the weight of the GPS backpacks may have contributed to the reluctance

of those birds to follow the main flocks into India. Therefore, at the outset of this project, there was a need to

attempt to keep the weight and volume of the instrument to a minimum. In order to reliably collect ECG so as

to permit heart-rate measurements, the logger had to be implantable. This imposed the additional need for a

biocompatible housing. An initial design request for a flashing LED was intended to confirm the continuing

functioning of the logger at the time of surgical implantation. To achieve both goals, transparent

perfluoroalkoxy (PFA) tubing was chosen (Polyflon, Stone, Staffordshire, UK). Cut to 56mm lengths, the ends of

the tubes were sealed by titanium electrodes providing ECG pick-up. Titanium is a low density metal and PFA is

a fluoropolymer similar to polytetrafluoroethylene (PTFE). These materials are highly inert and widely used in

medical implants. The design of a sealing mechanism between the PFA tubing and the electrodes was

conducted in collaboration with a company that machined the mechanical components for this project

(Beaumaris Instruments, Anglesey, UK). A small board incorporating a combined pressure and temperature

sensor was mounted on an external recess of one of the electrodes. This allowed the communication of

ambient atmospheric density which would not have been possible from within a rigid, sealed housing.

3

Design of Electronic Circuitry

Instrument design centred around a micro SD card for storage and a microcontroller (μC) programmed in C

making use of inline coding techniques and assembly language macros to optimise for speed and power

consumption. The first step during development was to assess whether to perform ECG beat detection in real-

time, as has been the traditional approach in the field study of animals [6], or whether the continuous logging

of raw ECG was feasible. Assembler code was used to program an 8-bit μC (PIC16LF872, Microchip, Chandler,

Arizona, USA) to control an SD card over a serial peripheral interface (SPI). This allowed its power requirements

to be investigated under a variety of operating conditions. Data could be stored with good energetic efficiency

but the standby current of the SD card during inactivity was sufficient to fully drain the preferred battery over

the course of a year. Specifications state that around half a second is required for initialisation of an SD card

following power-up. Since the initialisation current could be as high as 50mA, this posed another difficulty. As

the internal RAM of most μCs is rather limited, it was decided that an intermediate buffer would be required to

hold data, and power would be completely withdrawn from the micro SD card under the control of a P-channel

MOSFET when it was not in use. Dynamic RAM offered ample memory capacity but the complexity of the

interfacing and the need for refresh cycles was unnecessarily problematic. Serial memory devices were

considered but neither static RAM (SRAM) nor ferroelectric RAM was available with sufficient density, and

serial Flash and EEPROM drew relatively large currents. A parallel SRAM of 1MB capacity offered a good overall

compromise in terms of performance and power consumption, with a standby current of just 2μA and

operating current of less than 2mA at 1MHz (CY62158, Cypress, San Jose, California, USA). However, its control

required an 8-bit data bus and a 20-bit address bus, which was not easily driven using a low-end 8-bit μC. Thus,

Microchip’s 64 pin PIC24FJ128GA106 was selected for its enhanced I/O, superior processing throughput, low

power consumption, flexible sleep and standby modes, multiple SPI interfaces, 10-bit A/D conversion

capability, 128KB program memory and 16KB of internal RAM. This processor was capable of directly driving

the data bus and twelve of the address lines to the SRAM, along with handshaking lines. A commonly available

serial-in, parallel-out shift register (74HC164) was used to generate the remaining eight address lines. These

were assigned to the uppermost addresses so that they only needed to be changed as 4KB page boundaries

were crossed. ECG and accelerometry data had a high throughput and were asynchronously collected into 4KB

blocks of the internal RAM (one block for ECG and two for accelerometry) and flushed to pages in the external

RAM once filled. Due to the responsiveness of the parallel SRAM and the speed of the processor, this could be

accomplished without any pause or delay in either the ECG or the accelerometry sampling.

ECG amplifier

Due to the limited bandwidth of the ECG signal (2‒64Hz passband), it was possible to use ultra-low power

operational amplifiers drawing only 1μA per op-amp (AD8504, Analog Devices, Norwood, Massachusetts, USA).

Given the limited 7kHz gain-bandwidth product of these devices and the need for a high order filtering cut-off

so that the digitisation sample rate could be minimised, a two-stage approach was used for amplification with

gain set to 22 on the first stage and 43 on the second stage, resulting in an overall gain approaching 1V/mV. A

single pole high pass filter was implemented in the first stage and a third order low pass filter with a Q-factor of

2.7 was used on the second stage. The input impedance of the ECG amplifier was several MΩ and the output

signal was sampled at 32768/182≈180s/sec, just above the Nyquist rate required for a 64Hz signal and to

faithfully capture QRS complexes. The ECG amplifier circuit required only about 7μA when operating which was

easily driven by a port of the μC, an RC filter being added as a precaution against system noise.

4

Accelerometer

A “nano” triaxial accelerometer (LIS331DLH, ST Microelectronics, Geneva, Switzerland) with on-board 12 bit

digitiser supporting SPI output update rates as high as 1kHz and with a range of ±8g was chosen. Using

capacitive bridges as sensing elements, its current consumption was 250μA when active and 1μA in power-

down mode. This compared favourably with commercially available alternatives at the time but was expensive

to run compared to the ECG circuit, even including the burden on the processor due to regular interrupt driven

ECG sampling and data buffering. Continuous activation of the the accelerometer would have completely

discharged the battery in 3½ months. Though flight activity could have been intelligently detected in firmware,

it was felt that data on background activity of the geese was useful to collect. Therefore, the decision was

taken to collect data from the accelerometer every two minutes. Given the buffering arrangement with

internal RAM, approximately 18 seconds of data would fill 8KB or two pages in external RAM. This resulted in a

duty cycle of 15% which would only drain about half the battery.

Pressure and Temperature

A barometer was crucial to the study of the trans-Himalayan flights, allowing altimetry and rate of ascent

estimation. The transducer selected (MS5540-C, Intersema, Bevaix, Switzerland) was capable of measuring

absolute pressure to a resolution of 0.1mbar and an absolute accuracy of ±1.6mbar throughout the range 10 to

1100 mbar, encompassing all possible experimental changes in atmospheric pressure with altitude. Equipped

with a 16 bit ADC, it featured a built-in temperature sensor and calibration registers programmed at the factory

prior to shipping. The temperature sensor was needed to refine the calculation of atmospheric pressure and

also record the internal body temperature of the goose, providing insights into episodes of intense flight and

environmental changes such as sudden changes in wind conditions, as for instance might occur due to ridge

soaring. The sensing element was supplied coated in a layer of gel claimed to be waterproof to a depth of

100m which was covered in a layer of protective medical grade silicone. The standby current of the device was

less than 0.1μA. Its serial output was multiplexed with the SPI bus used by the accelerometer since pressure

and temperature measurements were only required every 30 seconds so simultaneous collection of

accelerometry with barometry could be avoided. The processor was required to actively generate a 32kHz

square wave for clocking the barometer. This was possible using the built-in output compare module but

meant the processor could not enter sleep mode during the 35msec needed for each temperature or pressure

reading. To conserve power, the processor clock was reduced from 8MHz to 1MHz at these times.

Magnetometer

The intensity of the Earth’s geomagnetic field ranges between 0.25 and 0.65 Gauss. Along the migratory route

of the bar-headed geese it varies from 0.43 Gauss in central India to 0.6 Gauss in Mongolia. It is quite possible

that the birds are more sensitive to changes in field inclination than intensity. Inclination varies considerably

over the same route, between about 20 and 70 degrees. The datalogger was designed to accept a triaxial

digital magnetometer (HMC5843, Honeywell, Morristown, New Jersey, USA), however, its performance proved

somewhat erratic on some loggers during trials and so only a minority of instruments were fitted with it during

manufacture. A ferromagnetic spring within the latching mechanism of the micro SD connector or magnetic

fields generated by the logger itself could also have interfered with measurements. Being cylindrical in shape,

the dataloggers were free to rotate within the abdominal cavity which may have resulted in some internal

5

movement of the battery wires in the vicinity of the sensor. The combined influence of these factors may have

caused the readings obtained to vary according to changes in orientation of the instrument.

Power

A 3.7 Volt, ½ AA sized Lithium Thionyl Chloride battery (ER14250, Omnicel, Phoenixville, Philadelphia, USA) was

used as an energy source. Tests determined its capacity to be somewhere between 800 and 1000 mAh at a

typical operating currents. Linear regulators (TPS78228 and TPS78223, Texas Instruments, Dallas, Texas, USA)

rated at 150mA output current and 1μA quiescent current were used to generate voltage rails of 2.8V and 2.3V.

The latter provided core power to the μC, avoiding the need to enable the built-in voltage regulator which

would either have drawn a high current during sleep or incurred an unacceptably lengthy start-up delay as it

was re-enabled on waking up. Provision was made for the processor to monitor the battery voltage. The output

current capability of the battery varied according to pulse duration but was capable of sustaining 20mA until

deeply discharged. Mean currents were kept below this level by episodically pausing in sleep mode as contents

of external SRAM were being transferred to the micro SD card.

Figure 1. Mechanical assembly of the logger showing the three PCBs (4 layer boards, FR4 substrate, thickness

0.8mm), their interconnections, the locations of the sensors and circuit elements and their physical positions

relative to one another. Programming of firmware was achieved via five prominent pins of the primary

interconnect using a PicKit3. Internal boards were protected by dipping in wax. Putty was used to prevent the

battery from sliding back and forth within the PFA tubing which might otherwise cause spurious impacts to be

transmitted to the accelerometer.

Circuit Layout and Mechanical Assembly

Electronic circuitry was laid out on three printed circuit boards (PCBs). One was a small circular board of

diameter 10mm mounting the pressure/temperature transducer and a decoupling capacitor. This was

externally mounted at a right angle to the other boards within a recess inside one of the titanium electrodes

and connected via a 6-way, 0.05” pitch connector passing through a slot in the electrode to one of the other

6

two boards, see Fig.1. Once soldered into position, the slot was sealed against ingress using epoxy resin

adhesive. The other two boards were of the same rectangular shape and size (22mmx14mm) and were stacked

above one another with the micro SD card connector sandwiched in between. A 22 position 0.05” pitch

connector in a dual-row 2x11 arrangement was used to join the two boards electrically, running width-wise

along the edge of the board opposite the end to which the electrode board attached. The width of the

assembly was intended to match the diameter of the battery so that all components could fit tightly within a

tube to reduce space and weight. The uppermost board was used to mount the 64 pin μC, the 48-pin external

SRAM and the 14-pin address register, components sharing many interconnections. Oscillator-related

components were positioned underneath this board. The high density of this PCB required manual tracking on

four layers taking several days. An initial solution was found using a slightly enlarged PCB area with the design

progressively reduced in size until further space savings became very difficult. This took several weeks and the

outcome was used to determine the dimensions of the second board which, like the first board, was populated

with components on both sides. On the innermost surface was placed the micro SD card connector, its high-

side switching MOSFET, glue logic for the barometer and some passive components. The ECG circuit,

accelerometer, magnetometer, linear regulators and the majority of the passive components were placed on

the underside. This second board was not quite as dense as the first but also required manual tracking on four

layers. It was to this board that the barometer was soldered. The titanium electrodes could not be soldered to

directly so metal pegs were tapped into drilled holes. This provided a reliable interference fit which required no

additional adhesive. Photographs of the physical assembly are provided in Fig.2.

Figure 2. Top left: titanium electrodes and view of the sensor board. Top right: with pressure/temperature

sensor attached through the slot in the titanium electrode (magnetometer also fitted). Bottom left: Top view

showing the recess of the titanium electrode sealed with silicone. Bottom right: Extraction of the logger from

the housing following recapture. Micro SD cards (15x11mm) are intended to provide a sense of scale.

7

Firmware Design

The PIC24FJ128GA106 μC contains a wide range of peripheral modules which can be individually enabled and

disabled to save power. An in-built RC oscillator runs at approximately 8MHz and allows the execution of code

to be resumed rapidly after waking from sleep without needing to wait for a phase-locked-loop (PLL) to settle.

In order to achieve extremely low power consumption during ECG sampling, low latency sleep recovery was

required. This particular design consideration featured prominently in the selection of this μC. An external

32kHz crystal oscillator was used to provide accurate timing, generating interrupts at the ECG sampling rate.

This oscillator consumed very little power but sometimes required several seconds to stabilise after application

of power. A flowchart for the firmware is provided in Fig.3. Accelerometry was collected when a hardware line

signalled the availability of a new sample. Other measurements were synchronised to the ECG sampling with

various flags and status registers being used to coordinate background operations.

Once the processor was brought out of reset, pin functions and mappings were configured, peripherals were

initialised and time was allowed for the 32kHz oscillator to stabilise. The firmware then proceeded to establish

communications with the micro SD card by first activating the MOSFET powering the card, allowing time for it

to power-up then engaging in preamble. The SD card was then interrogated via an SPI link to ensure that it was

functioning correctly and of the expected type etc. Initial transactions were carried out at a μC clock rate of

1MHz but once the SD card was properly initialised, the μC clock rate was raised to 16MHz using the PLL

allowing the SPI data rate to be increased from 256kb/s to 8Mb/s. The boot records were then read and the

contents of their registers used to determine such things as the number of bytes per sector, sectors per cluster

and number of entries in the root directory. This provided sufficient information to access the file allocation

table (FAT), two copies of which are usually present. Errors can easily arise when one copy is being modified

and power is interrupted, as might happen due to the card being ejected. This sensible precaution, however, is

not true of the root directory, only one copy of which exists. In order to streamline the logging of data,

minimise battery drain and reduce the risk of programming errors, the simplest possible storage arrangement

was used involving writing fresh data to consecutive sectors on the SD card starting from the first available data

sector. This resulted in a particularly simple FAT structure involving a consecutive sequence of incrementing

cluster references until the last cluster of each file was reached. The file organisation in the root directory

reflected this simplicity with a distinct file name being assigned to each consecutive gigabyte of logged data

and no other files, folders or directories being present. For a typical 8GB card (which actually contains 8 billion

bytes rather than 8GB, not all of which is available for data storage) this resulted in seven files of exactly one

GB and an eighth file of size 388 MB. This reduction in capacity was acceptable as the memory requirement for

a full year’s data storage was slightly less than 7GB. In this way, data recovery was extremely straightforward,

involving only removal of the SD card from its holder and insertion into a PC which could directly access the

data in each of the seven relevant files. Disk preparation typically took 60 seconds, including deliberately

inserted delays to reduce the mean current draw from the battery to an acceptable level. This would only

consume approximately 0.04% of the total battery charge.

After the preparation of the SD card, a sleep mode lasting 25 days was entered where the battery drain was

typically reduced to about 6μA. This allowed time for the instruments to be assembled, shipped to Mongolia

and transported 400km to a remote lake. Geese were ushered ashore using a flotilla of kayaks, and corralled

into a net. After ethylene oxide sterilisation and travelling over rough terrain to the catch site 400km remote

8

from Ulaanbaatar, instruments whose LEDS were still flashing correctly were implanted into captured geese by

a veterinary surgeon. Once the effects of anaesthesia had abated, geese were released back to the lake.

Implantation reduces adverse effects on animals over the long-term compared to external attachments [13].

Software was written for the PC to strip out the binary format data logged to the micro SD cards into distinct

files, a second program was used to display the data and a third to extract heart-rate from the raw ECG.

Various scripts were then written in R to further process and analyse the data.

Figure 3. Flowchart showing the sequence of primary tasks performed by system firmware.

9

A speed-optimised subroutine was used to transfer 4KB pages of data quickly to external SRAM. This involved

briefly increasing the μC clock rate from 8MHz to 32MHz and utilising the Parallel Master Port of the μC which

automatically performed the data transfer and the strobing of control signals. Whenever the external RAM

became full, data was written to the SD card in sector bursts by specifying the address of the first sector and

using the write_multiple_sector command. Firmware also kept track of the current sector number for the next

cycle of writing. In this way there was no need to call on library routines for opening, accessing and closing

files, which have additional overheads, and no need to access the FAT or the root directory again after the

initial preparation of the SD card.

Three seconds were typically required to write 1026 KB of data to the SD card, including inserted delays to

reduce mean current drawn from the battery. A μC clock rate of 16MHz was used during data transfer as it

offered superior energy efficiency to the alternative of 32MHz. If any error condition was detected, data

transfer was immediately aborted and logging resumed. This provided a safeguard against momentary loss of

electrical contact due to the mechanical bounce of the fingers within the SD card and also allowed some data

collection to continue when batteries were so discharged as to be incapable of sustaining an output current of

20mA for very long. The external SRAM buffer would refill in 4231.5 seconds (70.525 minutes), collecting

744KB of ECG digitised at 8 bits/sample along with 280KB of 12-bit triaxial accelerometry collected at a nominal

rate of 100Hz in 35 blocks of 8KB spanning approximately 18 seconds each. To conserve memory and reduce

power, pairs of 12 bit triaxial accelerometry samples were packed into packets of 9 bytes prior to being

buffered within internal RAM. Since the sample rate of the accelerometer was not accurately set by the

transducer itself, and drifted a little with time and temperature, the duration of each block of accelerometry in

ECG samples was appended to the accelerometry in the two spare bytes left over after 910 packets of data had

been placed in the 8KB accelerometry buffer. In this way, the true accelerometer sample rate could be

determined to an accuracy of 0.016%. In addition to the data in external SRAM, a 2KB internal RAM allocation

was used to hold raw barometer readings (pressure and temperature) along with magnetometer

measurements. These results were collected every 30 seconds, data being manually clocked out under

firmware control. The μC clock rate was dropped to 1MHz at this time, and fresh measurements took less than

0.1 seconds to collect. This internal buffer also stored calibration registers for the barometer, a battery reading

and an absolute time stamp in units of ECG sampling interruptions since logging began (these interrupts were

always enabled and were assigned a high priority so that they were never skipped, even though the ECG signal

was not sampled during disk writes). This internal RAM buffer was written to the SD card immediately after the

data from external SRAM had been written. A write_multiple_end command then instructed the SD card to

transfer its buffer contents to non-volatile memory. Once this was accomplished, power was withdrawn from

the card whereupon datalogging was resumed.

Discussion

The earliest trans-Himalayan flight recorded by this instrument occurred in late 2010. This system may well

have been the first implantable datalogger design capable of collecting year-round continuous ECG from a

flying animal. To the best of our knowledge, the archival recording of ECG for a full year has never been

attempted before on any animal, although a system relaying beat intervals telemetrically was capable of

operating for several years [14]. The data collected from the present instrument design will allow the long-term

study of avian migration in unprecedented detail. Comparisons will be possible between power estimated from

10

heart-rate combined with knowledge of body and/or heart mass [15] and power estimates obtained purely

from measurements of bodyweight and RMS vibrations collected from a body-mounted accelerometer [16].

Results and findings are currently being prepared for publication elsewhere.

As oxygen consumption is generally proportional to the square of the heart-rate [15], and bar-headed geese

exhibit a large variation in this physiological variable [17], see also Fig.4, their metabolic rate could vary

considerably. The pronghorn antelope (Antilocarpa americana) is adapted to outrun feline predators such as

the now extinct sabre-tooth cat and, to date, has the largest experimentally measured aerobic scope of any

animal [18]. Though a cheetah might be capable of catching this species, it is a specialist sprinter and no match

for the pronghorn on stamina. However, the bar-headed goose may rival the pronghorn both in terms of its

range of metabolic rate and its endurance.

During the first 2009 field trip when GPS systems were first deployed, a preliminary study involved the

implantation of a different instrument recording only body temperature and averages of fifty heart-rate

intervals. Several of those previous generation units were recovered in 2010. Although their ECG electrodes

were of a protruding, silicone-shafted flying-lead design (see [1]), allowing insertion closer to the heart, the

real-time beat detection algorithm was insufficiently robust to reject occasional interference from motion

artefacts and strong electromyographic production during bouts of intense flapping. Performing beat detection

`on the fly’ required the μC to draw a mean current of 100μA or less, severely restricting the sophistication of

real-time signal processing algorithms. Power limitations also prohibited alternative approaches, such as hybrid

processing via some combination of analog and digital circuitry. The recording of raw ECG for extended periods

overcomes any uncertainty regarding the reliability of heart-rate estimates, whether beat-by-beat or time-

averaged [19]. It can therefore enhance the practicality of using heart-rate as the basis for metabolic rate

estimation [6] by allowing heart-rate to be confidently determined even during intense and biologically

interesting episodes of flight, see Fig.4 & Fig.5. It also improves opportunities for beat-by-beat heart-rate

variability assessment [20] and identification of pathological heart conditions/cardiac arrhythmia.

Figure 4. Examples of ECG traces captured during the migratory flight of a single goose. Signal quality was

adequate despite the fact that flying lead electrodes adjacent to the heart were not used. Each trace covers 4

seconds and ECG has been upsampled by a factor of 8. Top: a clean signal, HR=420 bpm. Middle: a noisier signal

during ascending flight, HR=440bpm. Bottom: signal contaminated by motion artefacts close to take-off.

11

Figure 5. Heart-rate determined from post-processing of ECG covering the first few minutes of a lengthy flight.

The original impetus for the design approach described here was the desire to robustly measure heart-rate

during migratory flight. Prototyping demonstrated that storage of a digitised ECG signal to a micro SD card

using an appropriately selected μC and carefully written firmware would halve the battery drain compared to

real-time beat detection and averaging, despite the greatly increased memory requirements. With energy to

spare, it was then possible to also incorporate additional sensors. Over the course of a year, the collection of

accelerometry at 15% duty cycle and continuous ECG consumed a very similar fraction of the total energy.

Accelerometry required 1.8GB of memory storage compared to 4.9GB for ECG. The cost of collecting data from

other sensors was negligible both in terms of energy and memory. For an implantable system, geolocation via

either GPS satellite signals or daylight monitoring is impractical. However, it was possible to confirm that at

least one bird had crossed the Himalayas after median filtering of data from a unit equipped with a

magnetometer. This was apparent from changes in magnetic field intensity with time during the migration

period, see Fig.6.

Figure 6. The geomagnetic field after median filtering. Step changes were confirmed by accelerometry to be

associated with episodes of flight. The prominent final flight was a southward trans-Himalayan crossing

involving a 0.06 Gauss change in geomagnetic field intensity, a 12 degree change of latitude and an estimated

distance of 1300 km. The heart-rate plot of Fig.4 covered the start of this 26 hour flight.

12

A conscious decision was taken early on to monitor the year-round activity of bar-headed geese as opposed to

focusing only on episodes of flight. However, the acceleration spectrum of flight and non-flight exhibits a

strong dichotomy. In principle, this could have allowed for the intelligent collection of data during flight only.

Indeed, the firmware of this instrument has already been modified to do so and successfully deployed on

several neck collars for Mongolian Swans in the 2012/2013 season. For strength and UV resistance, a durable

opaque thermoplastic was used for the housing, Fig.7. Power at frequencies associated with the flapping of the

wings was estimated using a computationally efficient time-domain technique operating on short bursts of

accelerometry spanning several wing-beats. This proved very effective, enabling accelerometry to be collected

during flights with an acceptably small latency at the commencement of flight activity. The potential

advantages are apparent since, when flight can be efficiently discerned, both memory storage and battery

requirements can be reduced by an order of magnitude or more, dependent on the species under study. This

offers a very promising approach towards the reduction of size, weight and power requirements whilst

collecting high resolution flight data with improved coverage, thus minimising the burden and potential impact

to free-ranging birds during long-term experimental work.

Figure 7. Loggers within rugged housings were strapped and bonded to neck collars, internally lined here with

foam strips in an attempt to improve coupling of flight movements from the neck to the accelerometer. These

units were deployed on swans to collect accelerometry, temperature and altitude measurements.

Acknowledgements

We would like to thank our Mongolian colleagues who participated in the fieldwork and in particular

Nyambayar Batbayar and Tseveenmyaday Natsagdorj for securing the necessary authorisation. The support of

various colleagues from the University of British Columbia, the United Nations Food and Agriculture Program,

the US Geological Survey, the University of Tasmania and the University of Birmingham is gratefully

appreciated. Particular thanks are due to Bev Chua DVM, who led the surgical team.

Funding.

This work was supported by Bangor University and BBSRC grant BB/F015615/1 awarded to Dr C. M. Bishop and

Prof P. J. Butler.

13

References

[1] Woakes, A. J., Butler, P. J., & Bevan, R. M. (1995). Implantable data logging system for heart rate and body temperature: its application to the estimation of field metabolic rates in Antarctic predators. Medical and Biological Engineering and Computing, 33(2), 145-151. [2] Tucker, V. A., A 4 g, battery-powered, data logger accelerometer for free-flying birds, Review of Scientific Instruments, 70(11), 4394-4397. [3] Wilson, R. P., Shepard, E. L. C., & Liebsch, N. (2008). Prying into the intimate details of animal lives: use of a daily diary

on animals. Endangered Species Research, 4(1-2), 123-137.

[4] Butler, P.J., & Woakes, A.J. (2001). Seasonal hypothermia in a large migrating bird: saving energy for fat deposition?

Journal of experimental biology, 204, 1361-1367.

[5] Butler, P. J., Woakes, A. J., & Bishop, C. M. (1998). Behaviour and physiology of Svalbard Barnacle Geese Branta leucopsis during their autumn migration. Journal of Avian Biology, 536-545. [6] Butler, P. J., Green, J. A., Boyd, I. L., & Speakman, J. R. (2004). Measuring metabolic rate in the field: the pros and cons of the doubly labelled water and heart rate methods. Functional Ecology, 18(2), 168-183. [7] Yin, A., & Harrison, T. M. (2000). Geologic evolution of the Himalayan-Tibetan orogen. Annual Review of Earth and Planetary Sciences, 28(1), 211-280. [8] Gudmundsson, G. A., Benvenuti, S., Alerstam, T., Papi, F., & Lilliendahl, K. (1995). Examining the limits of flight and

orientation performance: satellite tracking of brent geese migrating across the Greenland ice-cap. Proceedings of the

Royal Society of London. Series B: Biological Sciences, 261(1360), 73-79.

[9] Blum, Arlene (1981). Annapurna: A woman’ s place. Sierra Club Books, ISBN 1578050222

[10] Hawkes, L. A., Balachandran, S., Batbayar, N., Butler, P. J., Chua, B., Douglas, D. C., Frappell P. B., Hou Y., Milsom W. K., Newman S.H., Prosser D. J., Sathiyaselvam P., Scott G. R., Takekawa J. Y., Natsagorj T., Wikelski M., Witt M. J., Yan B. & Bishop, C. M. (2013). The paradox of extreme high-altitude migration in bar-headed geese Anser indicus. Proceedings of the Royal Society B: Biological Sciences, 280(1750). [11] Takekawa, J. Y., Heath, S. R., Douglas, D. C., Perry, W. M., Javed, S., Newman, S. H., Suwal R. N., Rahmani A. R., Choudhury B. C., Prosser D. J., Yan B., Hou Y., Batbayar N., Natsagorj T., Bishop C. M., Butler P. J., Frappell P. B., Milsom W. K., Scott G. R., Hawkes L. A. & Wikelski, M. (2009). Geographic variation in bar-headed geese Anser indicus: connectivity of wintering areas and breeding grounds across a broad front. Wildfowl 59, 100-123. [12] Hawkes, L. A., Balachandran, S., Batbayar, N., Butler, P. J., Frappell, P. B., Milsom, W. K., Natsagorj N., Newman S. H., Scott G. R., Sathiyaselvam P., Takekawa J. Y., Wikelski M. & Bishop, C. M. (2011). The trans-Himalayan flights of bar-headed geese (Anser indicus). Proceedings of the National Academy of Sciences, 108(23), 9516-9519. [13] White, C. R., Cassey, P., Schimpf, N. G., Halsey, L. G., Green, J. A., & Portugal, S. J. (2013). Implantation reduces the negative effects of bio-logging devices on birds. The Journal of Experimental Biology, 216(4), 537-542. [14] Arnold, W., Ruf, T., Reimoser, S., Tataruch, F., Onderscheka, K., & Schober, F. (2004). Nocturnal hypometabolism as an overwintering strategy of red deer (Cervus elaphus). American Journal of Physiology-Regulatory, Integrative and Comparative Physiology, 286(1), R174-R181. [15] Bishop, C. M., & Spivey, R. J. (2013). Integration of exercise response and allometric scaling in endotherms. Journal of Theoretical Biology, 323, 11-19. [16] Spivey, R. J., & Bishop, C. M. (2013). Interpretation of body-mounted accelerometry in flying animals and estimation of biomechanical power. Journal of The Royal Society Interface, 10(87). [17] Ward, S., Bishop, C. M., Woakes, A. J., & Butler, P. J. (2002). Heart rate and the rate of oxygen consumption of flying and walking barnacle geese (Branta leucopsis) and bar-headed geese (Anser indicus). Journal of Experimental Biology, 205(21), 3347-3356. [18] Lindstedt, S. L., Hokanson, J. F., Wells, D. J., Swain, S. D., Hoppeler, H., & Navarro, V. (1991). Running energetics in the pronghorn antelope. Nature 353, 748-750. [19] Groscolas, R., Viera, V., Guerin, N., Handrich, Y., & Côté, S. D. (2010). Heart rate as a predictor of energy expenditure in undisturbed fasting and incubating penguins. The Journal of Experimental Biology, 213(1), 153-160. [20] Cyr, N. E., Dickens, M. J., & Romero, L. M. (2009). Heart Rate and Heart‐Rate Variability Responses to Acute and Chronic Stress in a Wild‐Caught Passerine Bird. Physiological and Biochemical Zoology, 82(4), 332-344.