PPG–ECG Validation of a 25 Hz Board-Level Photoplethysmography Sensor for Future Thermal Comfort Experiments Under Resting Low-Motion Conditions

Thermal comfort has emerged as a critical determinant of health, productivity, and overall well-being, particularly as modern indoor environments accommodate increasingly diverse patterns of work, study, and daily living [1, 2]. Contemporary research highlights that thermal comfort is closely linked to cognitive performance, physiological stability, and user satisfaction across buildings and urban microclimates [1, 3]. As indoor environmental quality becomes a central component of sustainable building design, recent studies emphasize the need for more adaptive and individualized approaches to thermal comfort research, reflecting variations in human physiology, behavior, and real-time interaction with the built environment [2, 4].

Recent studies in thermal comfort science also emphasize the importance of activity-specific evaluation, given the differences in human thermal sensation and thermoregulatory responses across sedentary, cognitive, and low-intensity physical states. Research indicates that thermal discomfort can lead to a measurable decline in cognitive performance, particularly affecting attention and task execution [3, 5]. This evidence shows the limitations of static thermal comfort models when applied to dynamic settings characterized by varied activity levels [3, 5]. To address these limitations, researchers have increasingly incorporated physiological monitoring to complement environmental measurements. Cardiovascular markers, such as heart rate (HR) and heart rate variability (HRV), have demonstrated potential for quantifying autonomic adjustments triggered by thermal exposure under sedentary and low-motion conditions [6, 7]. HRV features have also been reported to exhibit modulation as individuals transition between cooler, neutral, and warmer environments, highlighting the potential relationship between autonomic nervous system activity and thermal perception [8]. These developments have encouraged the integration of wearable physiological sensing with environmental data in future personalized thermal comfort studies [9-11]. Accurate physiological sensing is an important prerequisite for developing future physiology-driven thermal comfort experiments. Recent thermal comfort research has shown increasing interest in integrating cardiovascular markers, such as HR and HRV, with environmental measurements to better understand individual physiological responses in indoor environments [1, 2]. However, before these physiological features can be used in experimental thermal comfort protocols, the sensing modality used to acquire them must first be validated against a reliable reference method. This requirement is particularly important when low-cost or board-level sensors are used, because measurement errors in cardiovascular signals can directly affect the interpretation of physiological responses. Photoplethysmography (PPG) offers a practical and non-invasive approach for monitoring cardiovascular activity in laboratory and wearable-sensing applications. Compared with electrocardiography (ECG), PPG sensors are easier to deploy, lower in cost, and more suitable for embedded or portable experimental systems [10, 12]. Nevertheless, PPG does not directly measure cardiac electrical activity. Instead, it detects peripheral blood-volume changes, which are influenced by pulse transit time (PTT), vascular tone, sensor contact pressure, skin characteristics, ambient conditions, and motion artifacts [12]. These factors can introduce timing deviations in inter-beat interval estimation and may reduce the reliability of HRV features, especially those requiring high temporal precision. Although many studies have evaluated PPG performance in exercise, stress, and free-living conditions, validation evidence for board-level PPG sensors under thermally stable and low-motion laboratory conditions remains limited. This gap is relevant to future thermal comfort experiments, where participants are commonly seated, exposed to controlled indoor environments, and expected to minimize movement during physiological recording. Board-level PPG modules, such as the HRM-2511E, differ from consumer wearable devices in optical configuration, signal conditioning, contact interface, and sampling characteristics. Therefore, their performance cannot be assumed to be equivalent to that of commercial wearable systems without dedicated PPG-ECG validation.

The present study addresses this gap by conducting a preliminary laboratory validation of a 25 Hz board-level PPG sensor against a clinical-grade ECG reference. The validation was performed under one controlled indoor condition, namely 27 ℃, 60% relative humidity, and resting low-motion seated posture. No thermal exposure variation, thermal sensation vote, thermal comfort prediction model, emotional-state scale, and stress-score measurement were included in this experiment. Accordingly, the study does not directly assess thermal comfort and does not verify emotional relaxation. Instead, it evaluates whether HR and selected HRV features derived from the HRM-2511E PPG sensor can provide reliable cardiovascular measurements under conditions representative of a controlled laboratory setup for future thermal comfort experiments. By comparing PPG-derived and ECG-derived cardiovascular features, this study establishes baseline evidence regarding which parameters can be reliably extracted from a low-cost 25 Hz board-level PPG sensor under thermally stable and low-motion conditions. The findings are intended to support sensor selection, preprocessing design, and feature-selection decisions in future physiology-driven thermal comfort studies. The contribution of this work is therefore limited to sensor feasibility and validation, with particular attention to the sampling-resolution constraints affecting pNN20, pNN50, and frequency-domain HRV metrics.

Figure 1 illustrates the overall experimental and data-analysis workflow used for synchronized PPG–ECG validation under controlled thermal and low-motion conditions.

1.png

Figure 1. Methodology

3.1 Physiological signal acquisition using HRM-2511e and Elitech 1200G

The PPG analog output was connected to a microcontroller-based acquisition unit and sampled at 25 Hz using a 12-bit analog-to-digital converter. ECG reference signals were recorded concurrently using an Elitech 1200G clinical-grade electrocardiograph configured in a three-lead arrangement. The ECG signal was sampled at 500 Hz, exported as raw waveform data for offline analysis, and used as the reference modality for R-peak timing. Device-level filtering and internal preprocessing were retained as part of the acquisition chain before the same offline preprocessing workflow was applied consistently during analysis. Let the discrete PPG and ECG signals be represented as $x_{P P G}[n]$ with $n=0,1, \ldots, N-1$ and $x_{E C G}[m]$ with $m=0,1, \ldots, M-1$ with corresponding time indices, $t_n^{P P G}=\frac{n}{f_s^{P P G}}$, and $t_m^{E C G}=\frac{m}{f_s^{E G G}}$. The 12-bit ADC resolution supported digitization of the PPG waveform for subsequent systolic-peak detection. In the validation framework, ECG R-peaks were treated as reference cardiac electrical events, whereas PPG systolic peaks were treated as peripheral pulse-arrival events.

To address potential temporal mismatches, post-acquisition temporal alignment was implemented instead of exact temporal synchronization. While both systems were initially started within a 3-second window, the bulk lag was estimated using normalized cross-correlation between the preprocessed ECG and PPG waveform envelopes over the common recording window. The optimal time lag was determined from the maximum cross-correlation coefficient and then applied as a coarse time shift. The average synchronization delay applied across all subjects was -113.87 ± 126.91 ms. After this coarse alignment, residual beat-level offsets were preserved because ECG and PPG represent different physiological events: cardiac electrical activation and peripheral pulse arrival. These residual offsets were therefore interpreted as physiological pulse-transit effects and measurement jitter rather than synchronization errors.

During the acquisition phase, participants were seated comfortably and instructed to minimize movement while viewing a neutral five-minute video. This procedure was used to maintain a resting low-motion seated posture; no emotional-state scale or stress-score measurement was collected. Ambient conditions were maintained at 27 ℃ and 60% relative humidity, as monitored using a calibrated digital hygrometer-thermometer. These conditions were selected to minimize peripheral vasoconstriction and maintain PPG signal stability, as environmental factors have been shown to significantly influence PPG waveform morphology [16]. The outputs were exported as Comma-Separated Values (CSV) files containing raw waveforms and timestamps for reproducibility.

3.2 Data collection

3.2.1 Ethics approvals

This study received an ethical exemption from the Health Research Ethics Committee of Poltekkes Kemenkes Yogyakarta (Approval No. DP.04.03/e-KEPK.1/115/2025). The research protocol was reviewed and deemed ethically appropriate in accordance with the seven WHO (2011) ethical standards, which encompass social values, scientific values, equitable assessment and benefits, risks, persuasion and exploitation, confidentiality and privacy, and informed consent, as outlined in the CIOMS 2016 Guidelines. The ethical exemption is valid from February 4, 2025, to February 4, 2026. All participants provided written informed consent prior to data collection and were informed about the study’s objectives, data handling procedures, and their right to withdraw at any time without consequence. All procedures were conducted in full compliance with WHO and CIOMS ethical principles and approved by the authorized ethics committee to ensure participant protection and data integrity throughout the research process.

3.2.2 Participants

We recruited participants via public announcements to the Universitas Gadjah Mada students. Inclusion criteria required that participants be right-handed, physically healthy adults aged 20–30 years with no self-reported cardiovascular, neurological, or metabolic disorders. Individuals with a history of smoking, use of vasoactive medication, or prior cardiovascular disease diagnosis were excluded, as such factors can alter peripheral vascular tone and confound PPG signal physiology [15]. A total of 15 participants (8 females) with a mean age of 24.1 ± 2.3 years participated in the study after providing written informed consent. The sample therefore represents a limited and relatively homogeneous group of healthy young adults. Skin tone, finger geometry, contact pressure, and sex-related physiological differences were not stratified in the analysis and are treated as limitations of this preliminary laboratory validation. The sample size (n = 15) is consistent with prior PPG–ECG validation studies conducted under controlled laboratory conditions, where participant numbers between 10 and 20 are commonly adopted to establish baseline agreement and sensor accuracy. Such sample sizes have been shown to be sufficient for detecting moderate-to-strong correlations in criterion-validity analyses when motion and environmental confounders are minimized.

2a.png

2b.png

Figure 2. Experiment setup: (a) the layout of the chamber and (b) sensor setup

3.2.3 Collection setup

All recordings were performed in a climate-controlled chamber (2 m × 4 m) maintained at 27 ℃ and 60% relative humidity, as environmental stability has been shown to improve peripheral blood flow and reduce PPG signal variability [18]. This single controlled indoor condition was used only for laboratory sensor validation and was not designed as a multi-condition thermal exposure protocol. The spatial arrangement of the chamber and sensor setup is presented in Figure 2. Participants were seated in an ergonomic chair, positioned 1.5 m from a display screen, and instructed to remain still while watching a five-minute neutral video stimulus. The stimulus was used to support a resting low-motion posture rather than to verify a specific emotional state. PPG was recorded using an HRM-2511E reflective optical sensor module (940 nm infrared LED, integrated photodiode) secured to the participant’s left index finger to ensure consistent contact pressure. Contact pressure was controlled procedurally by using the same placement protocol, although it was not quantitatively measured. The analog output was connected to a microcontroller-based acquisition unit, sampling at 25 Hz with onboard analog filtering. ECG signals were recorded concurrently using a clinical-grade Elitech 1200G three-lead system, sampled at 500 Hz, with leads placed in standard limb configuration for accurate R-wave detection [20]. Before each session, sensors were visually inspected, cables were arranged to avoid strain artifacts, and signal quality was confirmed in real-time. Participants underwent a 5-minute adaptation period before recording to acclimatize to the environment and prevent novelty-induced autonomic fluctuations [12]. Data from both devices were saved in raw waveform form along with synchronized timestamps, enabling subsequent alignment and cross-modal HR/HRV analysis.

3.3 Signal processing

3.3.1 Signal pre–processing

Raw PPG and ECG waveforms were visually inspected to identify segments affected by motion artifacts, saturation, or baseline drift. Signal-quality screening flagged abrupt spikes, clipped amplitudes, non-physiological baseline shifts, and visually apparent motion artifacts before feature extraction. No participant-level recording was discarded; all 15 five-minute paired recordings were retained after quality control for subsequent analysis. To remove slow-varying non-physiological trends, each signal was detrended using third-order polynomial as shown in Eq. (1).

$p(t)=a_0+a_1 t+a_2 t^2+a_3 t^3$                  (1)

with the detrended signal $x_d(t)=x(t)-p(t)$.

Specifically, a 4th-order band-pass Butterworth filter was then applied to isolate cardiac-related frequency components. Filtering was performed at the native acquisition rates of each modality, namely 25 Hz for PPG and 500 Hz for ECG. For PPG signals, the passband was set to 0.5–4 Hz, whereas ECG signals were filtered within 0.5–40 Hz [12]. The filtering operation can be expressed as Eq. (2).

$y[n]=\sum_{k=0}^K b_k x_d[n-k]-\sum_{k=1}^K a_k y[n-k]$             (2)

where, the numerator and denominator terms denote the Butterworth filter coefficients, and the filter order was set to four. Zero-phase forward–reverse filtering was applied to prevent phase distortion.

3.3.2 Peak detection

ECG R-Peak Detection. Figure 3 shows ECG R-peak detection. ECG R-peaks were detected using a modified Pan–Tompkin’s algorithm involving differentiation, squaring, adaptive thresholding, and search-back logic. The detection procedure was applied to the 500 Hz ECG reference signal. Detected R-peak times were refined within a ±50 ms local window to correct temporal offsets. Intervals outside the physiological range (300–1500 ms) were manually inspected. Intervals remaining physiologically implausible after inspection were excluded from the beat-level feature calculation.

3.png

Figure 3. Electrocardiography (ECG) R-peak detection

PPG R-Peak Detection. Figure 4 shows PPG systolic-peak detection. PPG systolic peaks were detected using an adaptive derivative-based algorithm. Because the PPG signal was sampled at 25 Hz, peak timing was limited to a 40 ms temporal resolution. Let the detected systolic peak times be denoted as peak events. Spurious detections were removed using an IBI consistency filter defined as $RejectI B I_i \,if \mid I B I_i- median\left(I B I_{i-w i+w}\right) \mid>0.2 \cdot median \left(I B I_{i-w i+w}\right)$, where the moving window length was used to evaluate local IBI consistency. PPG detections with implausible IBIs outside 300–1500 ms or inconsistent local beat intervals were visually rechecked before feature extraction. PTT correction was not applied, as intrinsic timing differences between modalities were explicitly evaluated.

4.png

Figure 4. Photoplethysmography (PPG) systolic peak detection

3.3.3 Feature extraction

Following peak detection, inter-beat intervals (IBIs) were derived from both ECG R-peaks and PPG systolic peaks were computed as shown in Eq. (3). ECG-derived IBIs were used as the reference intervals, whereas PPG-derived IBIs represented peripheral pulse intervals.

$\begin{aligned} I B I_i^{E C G} & =t_{i+1}^R-t_i^R \\ I B I_i^{P P G} & =t_{i+1}^S-t_i^S\end{aligned}$                (3)

These IBIs were subsequently cleaned using artifact-rejection and interpolation procedures to obtain time series appropriate for HRV analysis. Time-domain HRV metrics were computed from cleaned normal-to-normal interval series, whereas frequency-domain metrics were computed after interpolation of the unevenly spaced IBI series. Recent studies emphasize the importance of preprocessing accuracy when estimating HRV from wearable sensors due to sensitivity to noise, sampling rate limitations, and beat-detection errors [17, 21]. HR was computed as the reciprocal of the mean IBI, expressed in beats per minute as follows in Eq. (4).

$H R=\frac{60}{I B I}$             (4)

Standard time-domain HRV features were extracted, including the standard deviation of normal-to-normal intervals (SDNN) and the root mean square of successive differences (RMSSD), both of which have been validated as robust indices of autonomic modulation in low-motion experimental protocols [22]. Specifically, the SDNN metric was computed to quantify the overall variability as shown in Eq. (5).

$S D N N=\sqrt{\frac{1}{N-1} \sum_{i=1}^N\left(I B I_i-\overline{I B I}\right)^2}$              (5)

Meanwhile, the RMSSD metric was calculated to estimate the short-term variance in heart rate using Eq. (6).

$R M S S D=\sqrt{\frac{1}{N-1} \sum_{i=1}^{N-1}\left(I B I_{i+1}-I B I_i\right)^2}$            (6)

Additional indices, pNN20 and pNN50, representing the proportion of successive IBI differences ($\Delta$) exceeding 20 ms and 50 ms, respectively, were computed to characterize rapid parasympathetic-driven fluctuations, which have been shown to be strongly relevant in short-duration laboratory recordings [18] which defined in Eq. (7). However, because the 25 Hz PPG signal has a 40 ms temporal resolution, pNN20 is fundamentally constrained by sampling resolution and cannot be interpreted as a reliable PPG-derived metric in this sensor configuration. The same timing limitation may also affect pNN50 and RMSSD, although these metrics are less directly constrained than pNN20.

$p N N_{\Delta}=\frac{1}{N-1} \sum_{i=1}^{N-1} \|\left(\left|I B I_{i+1}-I B I_i\right|>\Delta\right) \times 100 \%$                  (7)

Eq. (8) shows frequency-domain HRV estimation that was performed using Welch’s power spectral density (PSD) approach, which remains the recommended method for short-term HRV assessment in wearable-device validation studies due to its robustness to non-stationarity and noise [14]. Before PSD estimation, the cleaned IBI series was interpolated and resampled at 4 Hz to obtain an evenly sampled tachogram for spectral analysis.

$\hat{P}(f)=\frac{1}{L} \sum_{l=1}^L\left|\sum_{k=0}^{M-1} w[k] s_l[k] e^{-j 2 \pi f k}\right|^2$               (8)

where, $w[k]$ is a Hann window with a segment length of 256 data points and a 50% overlap. Spectral power was reported in raw ms² values. Spectral power was integrated over the standard frequency bands shown in Eq. (9), with the low-frequency band defined as 0.04 to 0.15 Hz and the high-frequency band defined as 0.15 to 0.40 Hz.

$\begin{aligned} & L F=\int_{0.04}^{0.15} \hat{P}(f) d f \\ & H F=\int_{0.15}^{0.4} \hat{P}(f) d f\end{aligned}$               (9)

and the autonomic balance index was computed as Eq. (10).

$L F / H F=\frac{L F}{H F}$             (10)

Only clean and stationary IBI segments were included in spectral computations, following best-practice guidelines for short-duration HRV recordings in controlled environments [19]. Because spectral estimates amplify beat-timing errors, frequency-domain HRV results from the 25 Hz PPG configuration were interpreted cautiously and were not treated as primary evidence of autonomic agreement.

3.3.4 Feature correlation and statistical analysis

To evaluate linear association between ECG-derived and PPG-derived cardiovascular features, Pearson correlation coefficients were computed for each HR and HRV metric as shown in Eq. (11). Correlation was not interpreted as measurement agreement.

$r=\frac{\sum\left(F_i^{P P G}-\bar{F}^{P P G}\right)\left(F_i^{E C G}-\bar{F}^{E C G}\right)}{\sqrt{\sum\left(F_i^{P P G}-\bar{F}^{P P G}\right)^2 \sum\left(F_i^{E C G}-\bar{F}^{E C G}\right)^2}}$                 (11)

Linear correlation analysis remains a widely adopted method for validating HRV features obtained from emerging or low-cost sensing technologies against clinical-grade references [19]. In this study, Pearson correlation and R² were used only as supporting measures of linear association. Agreement and reliability were evaluated primarily using error metrics, Bland–Altman analysis, and intraclass correlation coefficients (ICC). This statistical structure enabled the characterization of both feature association and measurement discrepancy between modalities under stable thermal and minimal-motion conditions.

3.3.5 Error analysis

Error-based statistical analyses were performed to complement correlation findings and quantify absolute discrepancies between modalities. Mean absolute difference (MAD), root mean square error (RMSE), and mean squared error (MSE) were calculated to assess the magnitude and variability of estimation error across features. The units of each error metric followed the corresponding feature units: bpm for HR, ms for SDNN and RMSSD, percentage points for pNN20 and pNN50, ms² for LF and HF power, and dimensionless values for LF/HF. Error-based metrics were computed using Eqs. (12)-(14).

$M A D=\frac{1}{N} \sum\left|\left(F_i^{P P G}-F_i^{E C G}\right)\right|$              (12)

$R M S E=\sqrt{\frac{1}{N} \sum\left(F_i^{P P G}-F_i^{E C G}\right)^2}$            (13)

$M S E=\frac{1}{N} \sum\left(F_i^{P P G}-F_i^{E C G}\right)^2$             (14)

These metrics are commonly employed in validation studies comparing wearable or optical sensors against ECG because they capture both systematic bias and random fluctuations inherent to pulse-derived intervals [22]. The coefficient of determination (R²) was additionally used to quantify the proportion of variance in ECG-derived features explained by PPG-derived estimates. Bland–Altman analysis was used to estimate mean bias and 95% limits of agreement, whereas ICC was used to summarize measurement reliability between the two modalities. Collectively, these statistical measures provided a comprehensive evaluation of the accuracy, reliability, and limitations of the HRM-2511E PPG sensor for HR and HRV estimation under thermally stable, low-artifact conditions.

Table 3. The comparative analysis of recent studies

To contextualize the present findings, Table 3 summarizes selected recent studies related to PPG-based cardiovascular monitoring. This comparison is presented as background context rather than a direct performance benchmark, because the listed studies differ in device type, sampling structure, experimental stimulus, participant context, and validation target. Therefore, the table is used to position the present work within the broader PPG-validation literature rather than to claim direct superiority over previous methods.

The comparative metrics in Table 3 are grouped into three broad dimensions of cardiovascular sensing validation: HR/RR accuracy, time-domain HRV, and frequency-domain HRV. HR/RR accuracy reflects the ability of the sensor to capture basic cardiac timing, whereas time-domain HRV metrics such as SDNN and RMSSD describe overall and short-term variability in inter-beat intervals. Frequency-domain HRV, represented by LF/HF, remains the most challenging dimension for low-resolution PPG sensors because it depends on spectral-power estimation and is highly sensitive to sampling-induced jitter and interpolation artifacts.

The HR estimation result in the present study (R² = 0.8826) is comparable with values reported in controlled and wearable PPG studies, particularly when the measurement condition is thermally stable and low motion. However, this comparison should be interpreted cautiously because some previous studies used higher sampling rates, commercial wearable algorithms, different physiological stimuli, and different validation endpoints [13, 16, 22]. Thus, the present result indicates feasibility under a controlled laboratory condition rather than general superiority over other PPG systems.

Regarding HRV metrics, the present findings are consistent with previous reports showing that basic time-domain HRV features are more stable than beat-difference and frequency-domain indices when derived from PPG signals [16, 22]. The lower reliability observed for LF/HF (R² = 0.4893; ICC = 0.534) supports the conclusion that frequency-domain autonomic interpretation is not sufficiently robust for the 25 Hz HRM-2511E configurations. By explicitly reporting this limitation, the present study provides a practical boundary for using board-level PPG sensors in future controlled thermal comfort experiments.

Several limitations should be acknowledged. First, the relatively small and homogeneous sample limits population-level generalizability, particularly because the participants were healthy adults aged 20-30 years and sex-related physiological differences were not examined. Second, the study was conducted under a single controlled indoor condition and resting low-motion posture; it did not include thermal exposure variation, thermal sensation votes, skin-temperature measurements, metabolic-response measurements, emotional-state validation, or stress-score assessment. Third, the low sampling rate of the PPG sensor constrains the estimation of high-resolution HRV metrics, particularly pNN20, pNN50, and frequency-domain indices. Fourth, reflective PPG measurements may be affected by skin tone, finger geometry, local vascular characteristics, and contact-pressure variability, none of which were explicitly controlled in this preliminary validation. Future research should therefore investigate multi-condition thermal exposures, higher-sampling-rate PPG systems, more diverse participants, and sensor-fusion approaches combining PPG with complementary physiological signals [23, 24].

The authors gratefully acknowledge the financial support provided by the Program Riset Konsorsium Unggulan Berdampak (RIKUB) 2026, Directorate of Research and Community Service, Ministry of Higher Education, Science, and Technology, Republic of Indonesia. This study was funded under Contract Numbers 013/C3/DT.05.00/RIKUB/2026 and 1270/UN1/DITLIT/Dit-Lit/PT.01.03/2026. The authors also extend their appreciation to all participants and laboratory personnel who contributed to the successful execution of this research.