Feasibility of Extracting Reliable Canine Heart Rate Variability from Ambulatory Electrocardiogram Recordings During Animal-Assisted Psychotherapy

2. Methods

2.1. Experimental Design

Ambulatory ECG recordings were collected from two certified therapy dogs during 38 routine CAP sessions conducted in a clinical setting. The dogs were two male Large Munsterlander aged 5 years (24 sessions) and 11 years (14 sessions). Sessions involved one therapist and varying patients and were performed without experimental manipulation or behavioral constraints, allowing spontaneous variation in movement, posture, and interaction behavior representative of routine therapeutic practice. Each session lasted 45 to 60 minutes. Raw ECG data is available at FigShare (https://doi.org/10.6084/m9.figshare.33080354).

Both therapy dogs underwent clinical veterinary assessment prior to the study and showed no evidence of cardiovascular abnormalities. All sessions were conducted by a licensed psychological psychotherapist with formal training and experience in dog-assisted psychotherapy. The therapeutic procedures followed the clinic’s routine practice and were not modified for the purposes of physiological recording. Prior to participation, all patients provided written informed consent for participation in the study and for audiovisual and physiological recordings. The study protocol was approved by the Ethics Committee of the Medical Faculty of Friedrich-Schiller University Jena (2025-3882_1-BO-A).

All sessions were video recorded using two synchronized digital cameras (HDR-CX240, Sony Corporation, Tokyo, Japan) positioned on opposite sides of the therapy room to ensure continuous visual coverage of both dogs throughout each session. Synchronization between video streams and physiological recordings was achieved using a light flash visible in both camera views and the ECG acquisition timeline.

2.2. Behavioral States

Behavioral state was coded retrospectively from the synchronized video recordings. One investigator (SS) performed the primary annotation, classifying each HRV analysis window as either active (walking, interacting with patients or therapist, or other overt movement) or inactive (lying, resting, or displaying minimal movement). All annotations were independently reviewed by a second investigator (AS), and disagreements (deviation by more than 30s) were resolved by consensus. Behavioral labels were subsequently aligned with the corresponding ECG analysis windows.

2.3. ECG Acquisition

ECG signals were acquired using a wireless recording system MP160 and RSPEC-R amplifier (BioNomadix, BIOPAC Systems Inc., Goleta, CA, USA). A three-channel ECG configuration was used to enhance signal robustness under ambulatory conditions. Data were sampled at a frequency of 1000 Hz.

Electrodes were attached to the dogs using a standardized placement adapted for ambulatory recordings (Figure 1) [20]. The wireless setup allowed unrestricted movement during therapy sessions while maintaining continuous signal acquisition.

2.4. ECG Processing and Beat Detection

ECG signals were preprocessed using the PhysioZoo toolbox implemented in MATLAB 2025b (The MathWorks Inc., Natick, MA, USA) which provides mammal-specific parameter sets for heart rate variability analysis and signal quality assessment [21]. To reduce baseline wander, muscle activity, and electrical interference, recordings were band-pass filtered between 3 and 45 Hz, thereby attenuating low-frequency drift and suppressing 50 Hz power-line noise. R-peaks were subsequently detected using two complementary algorithms with different sensitivities to noise [22,23]: ‘jqrs.m’, which has been shown to perform robustly in noisy recordings [24], and ‘gqrs.c’, which performs well under moderate-noise conditions [25]. Agreement between these independent detectors formed the basis for subsequent ECG signal quality assessment.

2.5. Signal Quality Assessment

ECG signal quality was quantified using the beat signal quality index, which estimates signal reliability based on the agreement between two independent R-peak detection algorithms with different sensitivities to noise [26]. The resulting signal quality index (SQI) was computed in rolling time window (10-s duration with a 1-s shift) throughout each recording, providing a quasi-continuous estimate of ECG quality over time. Based on these SQI values, ECG segments exceeding a predefined quality threshold can be identified and retained for subsequent HRV analysis, whereas segments below the threshold were excluded.

2.6. Sliding-Window HRV Estimation

Because standardized preprocessing criteria for ambulatory canine HRV recordings are currently lacking, physiologically plausible thresholds were defined a priori and applied consistently across all recordings. In addition to the minimal preprocessing parameter set for canine ECG [21], we excluded RR intervals outside the physiologically plausible range of 0.2 s – 2.0 s and successive RR intervals differing by more than 0.9 s. These thresholds were selected a priori based on the expected physiological heart-rate range of healthy dogs and visual inspection of representative recordings

The root mean square of successive RR interval differences (RMSSD), a time-domain measure of vagally mediated heart rate variability, was calculated in sliding windows of 60 s with a 10-s increment. RMSSD was computed only for windows meeting the predefined signal quality and preprocessing criteria described below.

2.7. Quality Regimes and Sensitivity Analysis

To evaluate the robustness of HRV estimates to preprocessing decisions, two quality regimes with different inclusion criteria were defined. The conservative regime prioritized signal fidelity by applying stringent quality thresholds (minimum SQI = 0.95) and excluding analysis windows containing more than 10 % missing or corrected R peaks. The permissive regime prioritized data retention by relaxing these criteria (minimum SQI = 0.80; maximum missing or corrected peak detections = 25 %), thereby allowing the inclusion of lower-quality ECG segments. Comparing results across both preprocessing regimes enabled assessment of the sensitivity of HRV estimates and statistical inferences to signal quality and artifact handling.

2.8. Statistical Analysis

Statistical analyses were performed in Python 3.15.3 (Python Software Foundation, Wilmington, DE, USA) using regression models with cluster-robust standard errors to account for repeated observations within therapy sessions. Data feasibility was evaluated using a generalized linear model with a binomial distribution and logit link, modeling the probability of obtaining a valid HRV estimate as a function of behavioral activity and therapy dog. Signal quality as measured by SQI was analyzed using linear regression with activity and therapy dog as fixed effects. The primary analysis modeled standardized RMSSD using linear regression including activity, standardized SQI, their interaction, and therapy dog as fixed effects. The interaction term was included to determine whether the association between SQI and HRV differed between activity states. All analyses were performed separately for the conservative and permissive preprocessing regimes to evaluate the robustness of the findings under different signal quality criteria.

3. Results

ECG recordings were obtained from 38 CAP sessions involving two certified therapy dogs (see one example in Figure 2). Application of the permissive preprocessing strategy yielded 10,075 valid one-minute HRV windows (71.0% of all candidate windows), whereas the strict preprocessing retained 4,529 valid windows (31.9%), illustrating the substantial impact of preprocessing criteria on data availability (Figure 3, A-B).

3.1. Data Feasibility

The probability of obtaining a valid HRV window was significantly influenced by behavioral state and dog identity. Under the permissive preprocessing strategy, active periods were associated with a lower likelihood of producing analyzable HRV data compared with active periods (β = - 0.789, 95% CI -1.175 to -0.402, p < 0.001). Dog identity also significantly affected recording feasibility (β = 1.107, 95% CI 0.536–1.678, p < 0.001).

The same pattern was observed using the strict preprocessing strategy. Active periods again reduced the probability of obtaining valid HRV estimates (β = -0.680, 95% CI -0.956 to -0.405, p < 0.001), while dog identity remained a significant predictor (β = 1.162, 95% CI 0.712–1.612, p < 0.001). These findings indicate that behavioral context systematically influences the feasibility of ambulatory canine ECG analysis regardless of preprocessing approach.

3.2. Signal Quality

Signal quality was evaluated using the average signal quality index (SQI). Under the permissive preprocessing strategy, SQI differed significantly between behavioral states (R² = 0.071). Inactive windows exhibited slightly higher SQI values than active windows (β = 0.271 standardized units, 95% CI 0.101–0.441, p = 0.002). A modest but significant difference between dogs was also observed (β = 0.503, 95% CI 0.120–0.886, p = 0.010).

In contrast, no significant effects of activity (β = 0.054, p = 0.412) or dog identity (β = 0.146, p = 0.173) were detected using the strict preprocessing strategy (R² = 0.004), indicating that nearly all retained windows already exhibited uniformly high signal quality.

3.3. Heart Rate Variability

The primary regression model examined standardized RMSSD as a function of behavioral state, SQI, their interaction, and dog identity. For the permissive dataset, the model explained 39.2% of the variance in RMSSD (R² = 0.392, p < 0.001). Inactive behavior was associated with substantially higher RMSSD than active behavior (β = 0.628 SD, 95% CI 0.519–0.736, p < 0.001). Dog identity exerted a similarly strong effect (β = 1.089 SD, 95% CI 0.838–1.339, p < 0.001).

SQI was positively associated with RMSSD during active behavior (β = 0.238 SD per one standard deviation increase in SQI, 95% CI 0.155–0.322, p < 0.001). However, this relationship was significantly moderated by behavioral state (interaction β = -0.218 SD, 95% CI -0.323 to -0.114, p < 0.001). Consequently, the association between SQI and RMSSD was largely absent during inactive periods, whereas increasing signal quality during active behavior was associated with progressively higher estimated HRV (Figure 3, C-F).

The strict preprocessing strategy produced highly consistent results despite retaining fewer observations. The model explained 42.1% of RMSSD variance (R² = 0.421, p < 0.001). Inactive behavior remained associated with higher HRV (β = 0.595 SD, 95% CI 0.444–0.745, p < 0.001), and dog identity continued to exert a strong effect (β = 1.377 SD, 95% CI 1.121–1.634, p < 0.001). Although the magnitude of the SQI effect was reduced compared with the permissive analysis, SQI remained a significant predictor (β = 0.111 SD, 95% CI 0.035–0.187, p = 0.004), and the interaction between SQI and behavioral state also remained significant (β = -0.107 SD, 95% CI -0.199 to -0.014, p = 0.024). Thus, the state-dependent influence of signal quality on HRV estimation proved robust across substantially different preprocessing strategies.

Overall, both analyses demonstrate that behavioral state is the dominant determinant of canine RMSSD during psychotherapy sessions, while signal quality primarily affects HRV estimates during periods of activity.

4. Discussion

The present study evaluated the feasibility of continuous ambulatory ECG monitoring for heart rate variability assessment during routine CAP sessions under naturalistic clinical conditions. Recordings were acquired while therapy dogs moved freely and interacted spontaneously with therapists and patients, introducing substantial motion-related challenges for ECG acquisition. Nevertheless, the wireless recording system combined with standardized signal processing yielded a large number of analyzable HRV segments across 38 therapy sessions. These findings demonstrate that reliable ECG-based HRV monitoring is feasible during routine clinical practice without restricting animal movement or altering therapeutic procedures.

A central objective of this study was to investigate how ECG preprocessing strategy influences both data availability and subsequent HRV analyses. The permissive preprocessing strategy retained approximately 71% of candidate analysis windows, whereas the strict strategy retained approximately 32%. Despite this more than twofold reduction in usable data, both approaches produced remarkably similar conclusions regarding the effects of behavioral state and the interaction between signal quality and HRV. This finding suggests that increasingly restrictive preprocessing primarily reduces statistical power rather than substantially changing the physiological interpretation of the recordings. From a practical perspective, this result is encouraging because ambulatory recordings inevitably contain movement-related artifacts that cannot be completely avoided in real-world applications. Rather than maximizing data rejection, an optimized balance between data retention and signal quality may therefore represent the most efficient approach for wearable physiological monitoring in freely moving dogs.

Behavioral state consistently emerged as the strongest predictor of HRV across both preprocessing strategies. Inactive periods were associated with higher RMSSD than active periods, reflecting increased parasympathetic modulation during resting behavior. This observation agrees with previous canine HRV studies reporting increased vagal activity during quiet resting conditions compared with periods of locomotion or behavioral engagement [4,5,6,7,8,9]. Importantly, the consistency of this effect across both preprocessing strategies supports the physiological validity of the proposed acquisition and processing pipeline and indicates that the observed autonomic differences are unlikely to be artifacts introduced by signal processing.

An important methodological finding concerns the role of the signal quality index (SQI). Under the permissive preprocessing strategy, higher SQI values were associated with higher RMSSD during active behavior, whereas this relationship was largely absent during inactivity. The same interaction remained statistically significant after application of the strict preprocessing criteria, although the effect size was reduced. This pattern suggests that signal quality is not merely a criterion for excluding poor-quality recordings but represents a context-dependent determinant of HRV estimation. During inactive periods, ECG quality is generally sufficient to permit stable R-peak detection and reliable RR interval estimation, resulting in only minimal influence of SQI on HRV. In contrast, during active behavior, movement-related disturbances increase the likelihood of detection errors, making HRV estimates more dependent on signal quality. Explicitly modelling SQI therefore provides additional information that would be lost if recordings were filtered solely using fixed quality thresholds [27].

These findings have practical implications for the development of wearable physiological monitoring systems for companion and working dogs. Long-term ECG monitoring outside laboratory environments requires acquisition systems that tolerate motion artifacts while maintaining reliable HRV estimation. The present results suggest that integrating objective signal quality measures into the statistical analysis may allow substantially greater data retention without compromising the validity of physiological conclusions. Such an approach may improve continuous welfare monitoring of therapy dogs, assistance dogs, service animals, and other working dogs operating under natural conditions [27,28].

The comparison between preprocessing strategies further illustrates the importance of evaluating signal processing choices explicitly. Although the strict preprocessing approach produced uniformly high-quality recordings, it substantially reduced the number of analyzable windows while providing only modest improvements in statistical model performance. Conversely, the permissive strategy preserved a much larger proportion of the recordings and yielded essentially identical conclusions regarding behavioral effects and the interaction between SQI and HRV. These observations indicate that moderate-quality ECG segments may still contain physiologically meaningful information when appropriate quality metrics are incorporated into the analysis. Consequently, future ambulatory HRV studies should consider combining objective SQI measures with statistical adjustment rather than relying exclusively on increasingly conservative exclusion criteria.

Several limitations should be acknowledged. First, recordings were obtained from only two certified therapy dogs, limiting the generalizability of the physiological findings to other breeds, ages, and working populations. Nevertheless, the study included 38 independent therapy sessions conducted under authentic clinical conditions, providing a large number of repeated observations across diverse therapeutic situations. Second, behavioral state was classified retrospectively into broad active and inactive categories. Although this approach was sufficient to demonstrate robust differences in HRV and signal quality, future studies could benefit from more detailed behavioral ethograms or automated video-based activity recognition to characterize the influence of specific behaviors on ECG quality and autonomic regulation. Third, cluster-robust regression accounted for repeated observations within therapy sessions but does not explicitly model the hierarchical structure of repeated measurements nested within individual dogs. Future investigations including larger numbers of animals should employ mixed-effects models with random effects for both dog and session to improve estimation of between-subject variability. Finally, only one time-domain HRV measure was examined in the present work. Additional evaluation of frequency-domain and nonlinear HRV indices under ambulatory conditions would provide a more comprehensive assessment of the robustness of wearable ECG monitoring.

In conclusion, continuous wireless ECG monitoring enables reliable HRV assessment in therapy dogs during routine CAP conducted under unrestricted clinical conditions. Behavioral state was the primary determinant of HRV, while signal quality exerted a behavior-dependent influence that was most pronounced during periods of activity. Importantly, similar physiological conclusions were obtained using markedly different preprocessing strategies, indicating that robust ambulatory HRV assessment is achievable without excessively restrictive data rejection. These findings support the use of objective signal quality metrics as an integral component of ambulatory ECG analysis and provide methodological guidance for future wearable sensing applications aimed at monitoring canine physiology and welfare in real-world environments.