Pain detection using biometric information acquired by a wristwatch wearable device: a pilot study of spontaneous menstrual pain in healthy females

Abstract

Objective

Pain is subjective, and self-reporting pain might be challenging. Studies conducted to detect pain using biological signals and real-time self-reports pain are limited. We evaluated the feasibility of collecting pain data on healthy females’ menstrual pain and conducted preliminary analysis.

Results

Five healthy adult females participated. They wore two wristwatch devices (Silmee and Fitbit) and a Holter ECG (electrocardiogram) during menstruation to record the pain intensity and timing. Subsequently, we analyzed the correlation between heart and pulse rates and assessed pre- and post-pain biometric differences. We collected sixty pain records from five participants. The correlation coefficients between heart rate and pulse rate ranged from 0.79 to 0.95 with Holter ECG vs. Fitbit and 0.32 to 0.74 with Holter ECG vs. Silmee. Analysis revealed significant changes in motion frequency post-pain (p = 0.04). For abdominal pain with a numerical rating scale score of ≥ 4 (n = 13), motion frequency (p < 0.001) and pulse rate (p = 0.02) showed significant differences post-pain compared to baseline values. Healthy females could wear the wristwatch device in daily life and report pain in real time. Wristwatch devices can effectively collect biological data to detect moderate pain by focusing on acceleration and pulse rate.

Introduction

Pain can lead to depression and reduce quality of life, so early management is crucial [1, 2]. Pain assessment commonly relies on the patient’s subjective report [3]. However, some individuals, such as terminally or critically ill patients, have difficulty in self-reporting [4]. In such cases, proxies like medical professionals or caregivers must evaluate pain, but it is often under- or overestimated, hindering accuracy [5, 6].

Recent studies have investigated the use of biological data to detect pain [7, 8]. Wristwatch devices, incorporating multiple sensors for non-invasive collection, can enable long-term patient monitoring [9,10,11,12]. Several reports indicate that photoplethysmography, electrodermal activity, skin temperature, and accelerometer data can detect pain signals. However, data collection remains a challenge. Some studies focused on real-time analysis, collecting data at 5-s or 2-min intervals [9, 10], capturing short-term physiological changes. However, biological signals change due to various factors, necessitating analysis of data collected over extended periods in daily life. Conversely, other studies recorded longer-term data [11, 12], but employed daily averages or maximum pain ratings, which are unsuited for developing a real-time automatic detection system.

We aimed to develop an automatic pain detection system using wristwatch devices in real-world settings. Because patient recruitment can be difficult, we conducted a pilot study involving menstrual pain in healthy females. We assessed the feasibility of real-time self-reported pain logging and performed preliminary data analyses.

Methods

Participants

Inclusion criteria

Females ≥ 18 years old with menstrual pain rated ≥ 4 on the numerical rating scale (NRS), ranging from 0 (no pain) to 10 (worst pain). Participants provided responses based on their typical pain experience.

Participants with irregular menstrual cycles and those who were unlikely to have their next period within a few days were excluded. Participants who reported that they used analgesics to proactively prevent menstrual pain were also excluded.

Five healthy females were enrolled. The sample size was similar to that of previous studies on biological signals using wristwatch devices: 1 [13], 5 [14], 8 [15], and 11 [16]. We estimated a sample size of five based on study feasibility and the researchers’ experience.

This study was approved by the Tohoku University Graduate School of Medicine Ethics Review Committee (No. 2022-1-890) and conducted in compliance with the Declaration of Helsinki. Written informed consent was obtained from all participants. Clinical trial number: not applicable.

Data collection

The wristwatch devices used were Silmee W22 (TDK Corporation, Tokyo, Japan) and Fitbit Sense2 (Fitbit Inc., San Francisco, CA, USA), with a Holter ECG (Cardy303 Pico+, SUZUKEN Co., Ltd., Aichi, Japan) used to exploratorily examine the correlation between the wristwatch’s pulse rate and ECG’s heart rate. We used two wristwatch devices to assess which might be more suitable for studies involving patients. Silmee has been used in Japan for emotion analysis [16]. Although the Fitbit Sense2 can measure electrodermal activity, raw data access is currently unavailable; thus, it was included with future use in mind.

Participants wore both wristwatch devices on the same arm 3 days before and after menstruation, with Holter ECG worn for 24 h on the first day of menstruation. The wristwatches were worn as consistently as possible, except during charging or bathing.

We used data obtained from the wristwatch devices at 1-min intervals. Pulse rate, skin temperature, and motion frequency were obtained from Silmee, and pulse rate from Fitbit. Motion frequency is a value calculated by Silmee that expresses the frequency of body movement from 0 to 15. The calculation algorithm is not disclosed. We used the Holter ECG data for the R-wave interval output using a Cardy Analyzer 05 (SUZUKEN Co., Ltd., Aichi, Japan). Silmee’s pulse rate and skin temperature were excluded from the measurement if the pulse rate was below 50 beats/min and the skin temperature was below 25 °C, indicating a measurement error. The pulse rate and skin temperature were linearly interpolated after noise removal using a Hempel filter that excludes data with values three times the standard deviation within a 5-min period.

We explained the aim of this study to the participants and asked the participants to accurately record pain changes. On registration for the study, participants completed a questionnaire providing background information such as age, height, weight, and medical history. Participants recorded the time period in minute intervals during which pain occurred, increased, decreased, or disappeared (when they felt pain intensity had changed), logging pain intensity and location on another paper questionnaire at home. The questionnaires are shown in Appendix 1.

Statistical analysis

We calculated the Spearman’s rank correlation coefficients for each participant for the pulse rate obtained from each device. These analyses were conducted using data from the entire period during which the Holter ECG was worn. We also calculated the Spearman’s rank correlation coefficient between the acquired index indicators and the NRS. We considered the period from the start of data collection to the first pain record, and that from pain disappearance to the next pain record, as NRS = 0.

To focus on changes during the occurrence of pain, we analyzed data from 5 min before (pre) and after (post) the occurrence of pain. We defined pain-free baseline (pre-pain) data as 5 min before all pain occurrences. We defined post-pain data as data from the time of pain occurrence to five min later. We analyzed pre- and post-comparisons for all pain and abdominal pain only where the pain threshold was NRS ≥ 4, ≥3, or ≥ 2. Pre- and post-comparisons were performed using either Welch’s t-test or the Wilcoxon signed-rank test, depending on the distribution of each index. We confirmed normality using the Shapiro-Wilk test. A 5-min window was used because when performing frequency domain analysis in heart rate variability analysis, 5 min of data collection is recommended [17]. In this study, heart rate variability (HRV) could not be analyzed due to Fitbit’s inability to output pulse wave interval. However, since HRV analysis is planned for future studies, a 5-min window was adopted to assess whether pain-related changes could be captured. Data from the three days prior to menstruation were not included. The significance level was set at p < 0.05. Analyses were performed using MATLAB R2022b (MathWorks Inc., Natick, MA, USA).

Results

Five healthy females (mean age = 27.8 ± 3.9 (SD) years; height = 160.2 ± 6.1 cm; weight = 50.5 ± 6.2 kg) participated in the study. Four participants had pain data while wearing both wristwatches and Holter ECG; one had data only from the wristwatches. Analysis was performed on all five participants. All participants were able to wear the wristwatch devices even while sleeping, and all reported that they were able to record all pain instances while awake. Table 1 shows details of the pain characteristics. In total, 60 reports of pain changes were obtained, including 32 occurrences of pain from no pain. Additionally, there were 18 records of pain disappearance. We confirmed that none of the participants were engaging in exercise beyond their daily activities during pain episodes. The correlation coefficients for heart and pulse rates were: Holter ECG vs. Fitbit, r = 0.79–0.95; Holter ECG vs. Silmee, 0.32–0.74. Table 2 shows the distribution and correlation coefficients for each pain intensity indicator. Motion frequency and Fitbit pulse rates were weakly correlated with NRS.

Table 1 Characteristics of the pain

Table 2 Correlation between the pain intensity and each indicator

Table 3 shows the pre- and post-pain comparisons of indicators obtained from wristwatch devices. Of these, two reports were excluded because there was no 5-min baseline in the minute interval record, as pain data was also recorded 5 min before the pain occurred. When comparing baseline and post-pain data of all pain occurrences, we found a statistically significant difference only in motion frequency (mean ± SD of baseline: 12.2 ± 4.9, pain: 13.5 ± 3.4, p < 0.04). When only lower abdominal pain was analyzed, the comparison of baseline data to data at pain with an NRS of ≥ 4 indicated significant differences in motion frequency (baseline: 12.2 ± 4.9, pain: 14.9 ± 0.3, p < 0.001) and pulse rate (baseline: 76.7 ± 10.4, pain: 79.6 ± 6.7, p = 0.02). Subsequently, we analyzed a lower cutoff value for NRS and found significant differences only in motion frequency (NRS ≥ 3: mean ± SD = 15.0 ± 0.2, p < 0.001, NRS ≥ 2: mean ± SD = 14.4 ± 1.93, p < 0.001, with similar baseline to that in Table 3).

Table 3 Pre- and post-pain comparisons of indicators obtained from the wristwatch devices

Discussions

We examined the feasibility of wearing wristwatches in daily life, and obtained self-reported pain data in real-time, focusing on healthy females’ menstrual pain. As a result, participants were able to wear devices and record their pain. Unlike previous studies with low night-time usage (1.7%) [11], our participants used the devices continuously, including at night, suggesting that explaining the device’s importance might improve compliance in future studies. While previous studies collected data once a day [11, 12], our participants recorded pain in real time over three days. In future studies, we plan to include patients who will wear wristwatches for several days to record their pain in real time, and we aim to develop machine learning models for pain detection.

While the correlation between pulse and heart rates alone does not indicate that wristwatches can replace ECG, the correlation between Fitbit pulse rate and Holter ECG heart rate suggests that wristwatches may reliably track pulse rate. Additionally, moderate pain (NRS ≥ 4) showed marked increases in pulse rate and motion frequency, indicating potential use in pain detection. Previous research supports this correlation between pain intensity and pulse rate [18,19,20], likely driven by sympathetic nervous system activity during pain [20]. Considerable differences were found when there was moderate pain (NRS ≥ 4), and no significant differences when the threshold was lower, suggesting that detectable changes may not be observed without somewhat strong pain. Previous studies reported that the accuracy of pain classification using biological signals such as ECG, surface electromyography, and electrodermal activity is reduced in mild pain [20, 21]. Since clinical intervention is often needed for moderate to severe pain, focusing on these levels may be more appropriate for pain detection [22].

Movement has been used in pain detection studies through image recognition of videos depicting pain [21] and behavioral pain assessment tools for patients [4, 23]. Furthermore, studies investigating behavioral changes in healthy individuals in response to stress have focused on changes in posture and increased self-touch [24]. Therefore, our results suggest that body movements may occur even in healthy individuals in response to pain. Providing detailed explanations for this change in movement is challenging because we did not collect data on how participants coped with their pain.

Conclusion

This pilot study explored pain detection using wristwatch-derived biological signals during menstrual pain in healthy females. Our findings show that continuous data collection is feasible and that moderate pain may be identifiable via pulse rate and motion frequency. Larger-scale research with diverse patient populations is essential for developing robust wristwatch-based pain detection systems.

Limitations

This study has five limitations. First, it is a preliminary study with a small sample size aimed at assessing the feasibility of collecting time-series pain data from wearable devices; therefore, the biological signals associated with pain require examination in a larger sample. Furthermore, pre-post pain comparisons of NRS ≥ 4 pain were based on an unbalanced sample size. Second, this study focused on menstrual pain in young females, limiting applicability to other sexes or age groups. Third, we did not gather detailed information on daily activities or postures during pain, limiting our ability to analyze changes in motion frequency. To build a reliable wristwatch-based pain detection system for daily use, algorithms must account for normal movement instead of only controlled data. Fourth, pulse rate changes may result from factors such as stress and movement, which cannot be fully eliminated in daily life, making high-accuracy detection essential. Finally, pain is subjective, and NRS scores can vary between individuals. Therefore, in clinical practice, establishing a personalized pain goal, tailored to each individual’s comfort level, could be beneficial for treatment. Although we did not implement such measure in this study, future research should consider adjusting the pain threshold to be detected for each subject.