Source-linked AI summary
Heart Rate Variability as a Predictive Biomarker of Thermal Comfort
Kizito Nkurikiyeyezu, Yuta Suzuki, Guillaume Lopez
TL;DR
Thermal-comfort models often omit personal physiological, psychological, and psychophysical differences. This paper tests HRV-based classification of thermal state and reports up to 93.7% accuracy, suggesting a possible basis for automated thermal controllers.
Problem
Many thermal-comfort mechanisms do not account for personal physiological, psychological, and psychophysical factors relevant to comfort.
Source: §1 · pp. 1–2Method
The paper proposes using HRV to classify people’s thermal-comfort state, drawing on its link to homeostasis.
Source: §1 · p. 2Results
93.7% accuracy was achieved by the machine-learning-based classification model for predicting thermal state.
Source: §4 · p. 11Takeaways & Limitations
The findings suggest that automated thermal controllers based on people’s HRV may be possible.
Source: §4 · p. 11
Abstract
from arXiv · showhide
Thermal comfort is an assessment of one's satisfaction with the surroundings; yet, most mechanisms that are used to provide thermal comfort are based on approaches that preclude physiological, psychological, and personal psychophysics that are precursors to thermal comfort. This leads to many people feeling either cold or hot in an environment that was supposed to be thermally comfortable to most users. To address this problem, this paper proposes to use heart rate variability (HRV) as an alternative indicator of thermal comfort status. Since HRV is linked to homeostasis, we conjectured that people's thermal comfort could be more accurately estimated based on their heart rate variability (HRV). To test our hypothesis, we analyzed statistical, spectral, and nonlinear HRV indices of 17 human subjects doing light office work in a cold, neutral, and hot environment. The resulting HRV indices were used as inputs to machine learning classification algorithms. We observed that HRV is distinctively different depending on the thermal environment and that it is possible to reliably predict each subject's thermal state (cold, neutral, and hot) with up to 93.7% accuracy. The result of this study suggests that it could be possible to design automatic real-time thermal comfort controllers based on people's HRV.
1 Introduction
Existing thermal-comfort systems often rely on static or average-occupant models that omit important personal differences. The paper proposes HRV as a homeostasis-linked signal for classifying individual thermal comfort states.
1 Introduction
- Static models assume steady-state conditions, prior knowledge of clothing and activity, and narrow comfort zones, while adaptive models account for outdoor conditions and broader comfort ranges.Static models prescribe a fixed year-round set point; adaptive models are based on field studies.Source: §1 · pp. 1–2
- Existing thermal regulation remains unsatisfactory: a survey of 215 buildings found only 11% achieved 80% occupant satisfaction, with an average of 59 ± 16%.Source: §1 · p. 2
- Current models often target an average occupant, although thermal expectations vary widely and occupants may prefer non-neutral conditions.The passage also describes frequent complaints of feeling too hot or too cold.Source: §1 · p. 2
- The paper proposes classifying individuals’ thermal comfort with HRV, which varies with homeostasis as the body responds to external stimuli.The authors envision personalized, real-time thermal regulation based on HRV.Source: §1 · p. 2
- Prior thermal-comfort HRV studies had limited scopes, including reliance on LF/HF, whose origin and interpretation are described as disputed.Source: §1 · p. 2
2 Material and Method
The study recorded ECG and thermal-comfort ratings from 17 male students working in cold, neutral, and hot climate chambers. It computed statistical, spectral, and nonlinear HRV indices and used statistical tests and machine-learning classifiers to assess their relationship to thermal conditions.
2.1 Experiment
- ECGs from 17 male students doing simulated office work in three climate chambers were recorded, alongside post-experiment thermal-comfort ratings.The chambers represented cold, neutral, and hot thermal sensations; each experiment lasted about 30 minutes.Source: §2.1 · p. 3
2.2 HRV Calculation
- The analysis computed statistical, spectral, and nonlinear HRV indices, including measures of beat-to-beat variation, frequency components, and nonlinear dynamics.For each subject, five-minute ECG windows were shifted by about 15 seconds to recalculate indices throughout the recording.Source: §2.2 · pp. 3–5, §2.5 · p. 6
2.3 Statistical Analysis
- Because most HRV indices were not normally distributed, the authors used a Mann-Whitney-Wilcoxon test to compare independent groups across environments.Normality was assessed with Levene’s test.Source: §2.3 · p. 6
2.4 Machine Learning Classification
- Ten supervised classifiers predicted cold, neutral, or hot conditions from HRV indices, with randomized logistic regression selecting indices and accuracy estimated by 10-fold stratified cross-validation.The authors describe feature selection as reducing model complexity and computation requirements while increasing accuracy.Source: §2.4 · p. 6
2.5 HRV Indices Computation
- HRV indices were recalculated from five-minute ECG windows shifted by approximately 15 seconds, and reported on a normalized 0–1 scale.The windows were advanced until the end of each ECG recording.Source: §2.5 · p. 6
3 Results
HRV indices varied across cold, neutral, and hot environments, though some indices showed clearer separation than others. An SVM classified thermal environment with 93.7% accuracy.
3 Results
- 93.7% accuracy was achieved by the Support Vector Machine classifier predicting subjects’ thermal environments from selected HRV indices.Randomized Logistic Regression selected Mean RR, RMSSD, SDSD, pNN25, VLF, and sample entropy.Source: §3 · p. 6
- Mean RR, VLF, sample entropy, and pNNx indices were generally highest in cold and lowest in hot conditions, while some subjects deviated.The pNNx difference was most pronounced for R-R increments of roughly 20–40 milliseconds.Source: §3 · p. 6
- Spectral, Poincare-based, and several statistical indices showed no clear separation among environments over time.These included TP, LF, LF/HF, HF, RMSSD, SDRR, and SDSD.Source: §3 · p. 6
- Subjects’ reported thermal sensations differed even when they experienced the same thermal environment.Ratings were collected on a visual analogue scale.Source: §3 · p. 6, Figure 2 · p. 7
- DFA(α1) was generally highest in the hot environment and mostly lowest in the cold environment, with exceptions among some subjects.Figure 3 presents the mean DFA(α1) pattern across environments.Source: §3 · p. 6, Figure 3 · p. 7
4 Discussion
Across cold, neutral, and hot environments, subjects’ HRV differed by setting, while reported thermal sensation varied between subjects. The discussion reports classification accuracy up to 93.7% and suggests HRV may support automated thermal-comfort control.
4 Discussion
- Although reported thermal sensation varied between subjects, no subject reported feeling hot in cold conditions or cold in hot conditions.Most subjects reported cool or cold sensations in the cold chamber, comfort in the neutral chamber, and warm or hot sensations in the hot chamber.Source: §4 · pp. 7, 9
- Mean RR and VLF were highest in cold conditions, while sample entropy was highest in cold and hot conditions and lowest in neutral conditions.The authors relate VLF to thermal regulation and interpret higher sample entropy in hot and cold settings as greater heartbeat unpredictability.Source: §4 · p. 10
- HRV differed significantly across thermal environments (P < 0.01), with DFA(α1) generally highest in hot and lowest in cold conditions.The hot–cold separation for DFA(α1) was clearer than the cold–neutral separation; Figure 4 marks subjects whose DFA(α1) was higher in cold than hot.Source: Figure 4 · p. 8, §4 · p. 10, Figure 9 · p. 11
- HF peaked in neutral conditions, whereas pNNx was highest in cold and lowest in hot conditions, with its greatest hot–cold separation at R–R increments of 20–40 milliseconds.The authors note pNNx is less computation-intensive and could be used on less powerful embedded microcontrollers.Source: Figure 7 · p. 10, Figure 8 · p. 10, §4 · p. 11
- 93.7% accuracy was achieved by the HRV-based classifier; Table 4 reports cross-validation accuracy for each subject.The authors also report that simpler Naïve Bayes models provided adequate accuracy, although the highest accuracies came from more complex models.Source: Table 4 · p. 8, §4 · p. 11
5 Conclusion
HRV varied with thermal environment, and machine-learning algorithms predicted subjects’ cold, neutral, or hot state with up to 93.7% accuracy. The study also identified HRV indices that distinguished conditions and were most useful for prediction.
5 Conclusion
- DFA(α1) was highest in hot conditions, while VLF, mean RR, sample entropy, and pNNx were highest in cold conditions, with hot–cold differences steady during the experiment.Source: §5 · p. 11
- Sample entropy was highest in cold and hot conditions; the authors interpret this as greater heartbeat unpredictability and complexity outside neutral conditions.They also report that cardiac sympatho-vagal balance increased in cold and hot conditions and decreased in neutral conditions.Source: §5 · pp. 11–12
- Up to 93.7% accuracy was achieved in predicting each subject’s cold, neutral, or hot state; six HRV indices were identified as most significant for prediction.The selected indices were Mean RR, RMSSD, SDSD, pNN25, VLF, and sample entropy.Source: §5 · p. 12
6 Practical Implication
The authors propose automated, real-time thermal comfort control based on HRV. Their approach could allow indoor temperatures to drift from thermal neutrality until occupants feel cold or hot, potentially reducing energy use.
6 Practical Implication
- HRV-based automated, real-time control could adjust indoor temperature when occupants feel cold or hot, rather than maintaining thermal neutrality.Source: §6 · p. 12
- Simulations cited project 29% cooling-energy and 27% total HVAC-energy savings from expanding the cooling setpoint by 2.8°C.Lowering the heating setpoint by 1.1°C was projected to save an average of 34% of heating energy.Source: §6 · p. 12
- Non-contact ECG and wearable PPG sensors could support non-invasive, low-cost monitoring; PPG is described as an unobtrusive HRV surrogate.Source: §6 · p. 12