Respiration shapes response speed and accuracy with a systematic time lag

Sensory-cognitive functions are intertwined with physiological processes such as the heartbeat or respiration. For example, we tend to align our respiratory cycle to expected events or actions. This happens during sports but also in computer-based tasks and systematically structures the respiratory phase around relevant events. However, studies also show that trial-by-trial variations in the respiratory phase shape brain activity and the speed or accuracy of individual responses. We show that both phenomena, the alignment of respiration to expected events and the explanatory power of the respiratory phase on behaviour co-exist. In fact, both the average respiratory phase of an individual relative to the experimental trials and trial-to-trial variations in the respiratory phase hold significant predictive power on behavioural performance, in particular for reaction times. This co-modulation of respiration and behaviour emerges regardless of whether an individual generally breathes faster or slower and is strongest for the respiratory phase about two seconds prior to the participant’s responses. The persistence of these effects across 12 datasets with 277 participants performing sensory-cognitive tasks confirms the robustness of these results and suggests a profound and time-lagged influence of structured respiration on sensory-motor responses.

Methods
Respiration was recorded using a temperature-sensitive resistor that was inserted into disposable clinical oxygen masks (Littelfuse Thermistor No. GT102B1K, Mouser Electronics). This effectively captures the continuous temperature changes resulting from the respiration-related airflow. The voltage drop across the thermistor was recorded via the analogue input of an ActiveTwo EEG system (BioSemi BV; Netherlands) at a sampling rate of 500 or 1000 Hz. We verified that the voltage drop of the temperature sensor follows the respiratory air flow without time lag. For this, we combined the temperature probe with two short-latency airflow sensors (F1031V, Mass Airflow Sensor, Winsen) and confirmed that the temperature change tightly aligns with the directional change in airflow.

Compared to our previous work we improved the processing pipeline for respiratory data. The respiratory signals were filtered using 3-rd order Butterworth filters (high pass at 0.03 Hz, low pass at 6 Hz) and subsequently resampled at 100 Hz using the FieldTrip toolbox. The signals were then converted to z-scores to facilitate comparison across participants (Fig. 1C, D). To detect individual respiratory cycles, we applied the Hilbert transform to determine local peaks based on the respective phase. Individual respiratory cycles were determined based on the data in windows of 7 seconds around each peak, whereby individual peaks were only retained for further analysis if the z-scored trace exceeded a value of z=0.5. Note that alternative algorithms to detect individual respiratory cycles exist and in a previous study, we found little difference between these. The inhalation period was defined as the continuous period with a positive slope prior to the local peak (whereby interruptions of the positive slope shorter than 500ms were interpolated). The exhalation period was defined as the continuous period with a negative slope subsequent to the local peak (again interruptions shorter than 500ms were interpolated). This definition effectively splits the respiratory cycle effectively into the two main periods of inhalation and exhalation; though for some cycles short exhale pauses were classified as the third state and not analysed. In particular, compared to previous work this procedure assigned a defined inhalation/exhalation phase to more time points than in the previous work. To characterize atypical respiratory cycles, we compared the overall time courses of individual respiratory cycles using their mean-squared distances. We calculated the participant-wise distributions and excluded cycles with a distance larger than 3 standard deviations from the centroid as atypical. These cycles were excluded as they do not reflect the prototypical respiration under investigation here. Further individual trials were excluded from the analysis as noted below. From the full datasets, we retained only participants for which these procedures excluded less than 30% of the available trials for the final statistical analysis.

感知认知功能（sensory-cognitive functions）与心跳、呼吸等生理过程紧密交织。例如，人们往往会将自身呼吸周期与预期事件或动作同步，这一现象不仅出现在运动场景中，也存在于计算机辅助任务中，并会围绕相关事件系统性地构建呼吸时相。不过已有研究表明，呼吸时相的逐试次变异会影响脑活动以及个体反应的速度与准确性。本研究证实，呼吸与预期事件的同步、呼吸时相对行为的解释力这两种现象可共存。事实上，个体相对于实验试次的平均呼吸时相，以及呼吸时相的逐试次变异，均对行为表现具有显著的预测力，尤其在反应时方面。无论个体整体呼吸频率快慢，呼吸与行为的共同调制效应均会显现，且在参与者反应前约2秒的呼吸时相阶段效应最强。该效应在涵盖277名参与者完成感知认知任务的12个数据集间均保持稳定，证实了这些结果的稳健性，同时表明结构化呼吸对感知运动反应存在深远且时滞性的影响。

研究方法
本研究采用植入一次性医用氧气面罩（disposable clinical oxygen masks）的热敏电阻（temperature-sensitive resistor，Littelfuse Thermistor No. GT102B1K，贸泽电子（Mouser Electronics））采集呼吸信号，可有效捕捉呼吸相关气流带来的连续温度变化。通过ActiveTwo脑电系统（ActiveTwo EEG system，BioSemi BV；荷兰）的模拟输入端口记录热敏电阻两端的电压降，采样率设置为500或1000 Hz。我们验证了温度传感器的电压变化与呼吸气流无明显时滞：将温度探头与两款短延迟气流传感器（F1031V，空气质量流量传感器，Winsen）结合，确认温度变化与气流方向变化高度一致。

相较于此前的研究工作，我们优化了呼吸数据的处理流程：首先使用三阶巴特沃斯滤波器（Butterworth filters，高通截止频率0.03 Hz，低通截止频率6 Hz）对呼吸信号进行滤波，随后通过FieldTrip工具箱（FieldTrip toolbox）将信号重采样至100 Hz，并转换为Z分数（z-scores）以方便跨参与者比较（图1C、D）。为检测个体呼吸周期，我们采用希尔伯特变换（Hilbert transform）基于对应相位确定局部峰值。以每个峰值为中心的7秒窗口内的数据来界定单个呼吸周期，仅当Z分数化的信号轨迹超过z=0.5时，该峰值才会保留用于后续分析。需说明的是，现有多种检测个体呼吸周期的算法，我们此前的研究表明不同算法间差异极小。吸气期定义为局部峰值前斜率为正的连续时段（斜率中断时长短于500 ms的部分将被插值补全）；呼气期定义为局部峰值后斜率为负的连续时段（同样将时长不足500 ms的斜率中断进行插值补全）。该定义将呼吸周期有效划分为吸气、呼气两个主要阶段，但部分周期中短暂的呼气暂停会被归类为第三种状态，不予分析。相较于既往研究，本流程可为更多时间点分配明确的吸气/呼气时相标签。

为表征非典型呼吸周期，我们基于均方距离（mean-squared distances）比较个体呼吸周期的整体时域特征，计算参与者层面的分布，并将与质心距离超过3倍标准差的周期判定为非典型周期并剔除，因其无法反映本研究关注的典型呼吸模式。此外，如后文所述，部分单独试次也会被剔除。在完整数据集中，我们仅保留最终统计分析中被剔除试次占比低于30%的参与者数据。