How to Work with Breath Belts: Tips for Better Respiratory Data
During my PhD at University College London, I used respiratory inductance plethysmography, or breath belts, to study how we breathe to speak. For many of us, inhaling during speech happens automatically and with little conscious thought, unless we've been strenuously exercising or have a bad cold. But when you stop to think about it, it's amazing how well we fit our inhalations in while we talk—rapidly, flexibly, precisely, even gracefully—often without interrupting our speech flow.
Speech breathing has grown even more interesting to neuroscientists, as more and more research shows that our respiratory cycle interacts with cortical neural activity and may contribute to the dynamics of attention, memory, and learning. I cover a little bit of this exciting research, as well as some breathing scientific history, in this talk.
If you are interested in incorporating respiratory measures into your experiments, I would certainly recommend it. But breath belts are delicate and prone to noise, especially if participants are moving around to speak or respond in some way during your task. Here, I've compiled some lessons learned from recording more than 60 speakers over my doctoral work. And check out the Speech Breathing Toolbox, a set of MATLAB functions that you can use to automatically annotate your breath belt data with objectivity and good temporal precision. For reference, I worked with belts similar to this one by ADInstruments.
The Speech Breathing Toolbox is available on Github and MATLAB File Exchange.
Setting Up
- Pick a space with minimal electrical interference, and use battery-powered lighting if available. This wasn't consistently a problem, but occasionally, I found line noise in my breath belt recordings.
- It's important that you have reliable temporal calibration between your respiratory and other data acquisition devices. I first tried to pair up separate recordings based on synchronised events, but different devices have different internal clocks, and even if you are using the same sampling frequency, there may be tiny differences that amount to an annoying drift between your signals. Besides, speaking from experience, finding the exact moment of simultaneity (e.g., across the acoustic and plethysmographic traces) isn't easy. Instead, hook the breath belts up to the same piece of hardware you will be using to record other signals of interest, such as your EEG acquisition device.
- You may need to source an adapter, such as XLR to BNC or vice versa, to make this work.
- I recorded plethysmography and acoustic speech recordings to the same analog-to-digital converter. Even if there is no acoustic component to your experiment, record their breathing sounds anyways, so that you can make sure your breath belts are doing what you think they are.
- Allow for two breath belts per participant. I was astonished by how much people vary in their respiratory patterns, and you can't predict it just based on their reported meditation or wind instrument playing experience. Can you use one belt but move it around? I have fit someone with two belts, observed viable respiratory traces from each belt during set-up, and then watched that person acclimatise right back into using only their chest to breathe. If you have two, I suggest placing one belt at about navel height, and the other at the broadest part of their chest.
- Sometimes, you might get lucky and both belts clearly track their breathing cycle in a similar way and with matched timing. Great, take the average of the two belts for some free noise reduction. In my experience, most people do not, and whereas one breath belt will clearly correspond to respiration, the other one will consist of a whole lot of movement artifacts—this is specifically more of a problem for speech breathing, however.
Data Collection
- Ask your participants to wear layers, preferably with a tight-fitting base layer. Big, chunky sweaters are your enemy! Since the stretchy breath belt provides a measure of torso displacement, your data quality is utterly dependent on a snug and direct belt fit: bunches of fabric will cushion the belt, resulting in a mushy, temporally smeared signal. Hence, a tank top or t-shirt made from a thin, elastic material is your best bet.
- Metallic jewellery, wiring in bras, and metal buttons generate artifacts. Warn your participant to avoid wearing anything dangly or made of metal, or ask them to make sure that it can be removed prior to beginning the experiment.
- Anticipate movement artifacts, which are close to unavoidable in breath belt data. Gesture-related movements and postural sway are a natural and fascinating aspect of speech production, but they will swamp your respiratory signal and make it very difficult to tell what is happening when.
- Our solution was to give participants a strip of cardboard to hold with both hands, arms hanging loosely at their sides. This intervention was very effective at dampening gesture and gently restraining movement more generally. I don't think it interfered terribly with task naturalism—consider meeting a friend after doing your grocery shopping, and carrying on a conversation with bags in your hands.
- Don't worry about the breath belts drawing participants' attention to their own breathing. OK, so I haven't tested this formally, but one of the amazing things about breath control is how it slips in and out of the purview of our conscious attention. When the participant first puts the belts on, they will probably be hyper-conscious of their breathing at first. But as soon as a little time goes by and/or the task begins, this is unlikely to be an issue.
- Calibrate the belt to the participant's typical breathing movements. I would often have participants take as big of a breath as they could. Besides getting the scale right for online monitoring, you can also use this information if you would like to make a rough estimate of volume and/or standardise units in order to compare across speakers.
- Collect a lot of data. Breathing takes time, and depending on the task and dependant variable, a couple of minutes' recordings may only yield 10-20 observations (e.g., individual inhalations or cycles). Given artifacts, other noise, and inter-individual variability, make sure that you have what you need to detect your effect of interest.
Pre-Processing and Analysis
- Avoid filtering your data. If things go sideways with the line noise, this may not be feasible. But filtering your signal can introduce all sorts of temporal distortions, including sometimes subtle ones that are difficult to catch. This is a well-documented problem for EEG and MEG analysis, but similar problems apply to respiratory data. A simple moving mean will smooth the signal without transforming or degrading its time series too badly.
- Don't trust simple thresholding or statistical approaches to automatically annotate the respiratory cycle. As discussed by me and other respiration researchers, it's not a good idea to take the zero-crossings or peaks and troughs of the breath belt signal to detect inhalation events or demarcate cyclical units. This will result in low accuracy, whether estimated from the perspective of signal detection or temporal resolution.
- If you have access to MATLAB, try my Speech Breathing Tool Box. Its development was informed by a lot of trial and error, as well as careful calibration of the respiratory signal with concurrent acoustic recordings.
- You can also look at the BreathMetrics toolbox, which is aimed at nasal, non-speech-related respiration, but offers great functionality (e.g., a GUI).