Estimating time of death from smartwatch data: initial empirical results Authors: Tanja Gosseling, Bartholomeus G H Latten, Jan Peter van Zandwijk Published: 2026 ABSTRACT AbstractDetermining the time of death is a critical factor in forensic death investigations. Currently, the most widely used method is rectal temperature measurement. However, there are many, often unknown, factors influencing the results of this method and furthermore it has the limitation that it can only be applied within a restricted timeframe. Therefore, there is a need for new alternative methods. The research question of this exploratory study is whether digital information from smartwatches allows for a reliable determination of the time of death. With this purpose in mind, we collected data from Garmin, Apple and Samsung smartwatches, worn by participants dying at a scheduled moment during a euthanasia procedure. Analysis of experimental data (N = 27) indicates that data registered by smartwatches showed changes around the participant’s death, allowing for an estimation of time of death. Although the number of variables and time granularity of registrations varied across types of devices, in most cases the difference between the time of death determined by the physician and the estimates derived from smartwatch data, was smaller than roughly 30 min. For unexplained reasons, however, in some instances there were gaps in the data recorded by smartwatches. This highlights the need for caution when applying this data for determining the time of death. Nonetheless, current first results do indicate that data from smartwatches has potential of providing relevant information about time of death in addition to existing methods.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-026-61387-9.Keywords: Post-mortem interval, Time of death, Forensic death investigation, Smartwatches, Wearable device forensicsSubject terms: Public health, Medical research, Mathematics and computing Introduction Determining the time of death (TOD) can be an important factor in both medical and forensic death investigations. Information about the TOD is often crucial in a forensic setting of a (potential) homicide, where a clear timeline of events needs to be established. Additionally, information about the time and the date somebody has died may help the bereaved in their grieving process. Without reliable eyewitnesses, the TOD can be difficult to determine. In such cases, tactical information, for instance the time the person was last seen or camera footage, along with post-mortem changes in the body—rigor mortis (stiffness), livor mortis (lividity), and algor mortis (cooling of the body)—are used to estimate the time interval between death and the examination of the body, the so-called post-mortem interval (PMI). Currently, rectal body temperature measurement is a widely applied method for estimating PMI. Many models have been created[e.g.1–4], with the one developed by Henssge5,6 being the most common in practice. However, the results of such methods can be affected by various factors, including fluctuation of ambient temperature, wind flow, type of surface on which the body is found, clothing and whether the body is wet. As a result, calculation of PMI using temperature measurement often produces results with an interval of several hours (from ± 2.8 h up to ± 7 h5–7), which may often be too wide for practical purposes. Furthermore, once rectal temperature equals the ambient temperature – at room temperature typically occurring around three days post-mortem - the PMI cannot be calculated using this method. This also applies to a recently introduced more advanced method for determining PMI using a thermal photogrammetry and modelling approach, which shows an interval of 0.26 ± 1.38 h7. Also, other methods for determining the PMI relying on thanatochemical changes are being studied, such as changes in ocular potassium concentration8 and proteomics9–11. Given the uncertainties associated with the currently employed methods, there is a critical need for alternative methodologies to ascertain the PMI or, preferably, the TOD. An interesting new opportunity for this may be emerging from the field of health-related consumer electronics. In the present era of technological innovation, wearable devices such as smartwatches, smart rings or other activity trackers are enjoying popularity and consequently deceased persons are increasingly often found wearing such devices. In a broader forensic context, digital traces from various types of consumer electronics have been found to have potential for inferring information about activities in the physical world12. In this study, we explore the possibility of wearable health-related consumer electronics in providing information concerning the TOD. Using photoplethysmography (PPG)13 and information from sensors such as accelerometers and gyroscopes, wearable devices can collect several types of biometric data, including heart rate, blood oxygen saturation, or measures of physical activity, which are, importantly, all registered together with detailed timestamps. As biometric data collected by smartwatches are expected to show changes after death, these devices might offer interesting new investigative opportunities for estimating TOD. As data can remain stored on devices for longer periods of time, this could also potentially extend the time range for which estimates of TOD can be obtained beyond what is currently achievable with existing temperature-based techniques. In living people, smartwatches have already been studied for their use in the diagnostics of medical conditions, such as cardiac arrhythmias14–16 and detection of epileptic seizures17. Furthermore, it has been investigated whether data from smartwatches can be used to make a model, that can predict death events in terminally ill patients18. In a clinical setting, a recent study19 has shown that specifically developed cardio watches can detect a medically induced circulatory arrest. However, the reliability and time range of estimates of the TOD using data from commonly available types of smartwatches is currently unknown. To assess the potential of data from wearable devices for determining TOD, it is essential to collect data in a setting with known ground-truth, i.e. where TOD is independently determined by a physician. Such a study design presents substantial practical and ethical challenges, since it requires collecting data from individuals wearing a device in situations where death is anticipated. In this regard, cases in which a patient receives euthanasia could offer a unique opportunity, as the date and time of death are scheduled. In the Netherlands, euthanasia is a strictly regulated legal procedure, defined as “termination of life by a physician at the request of a patient”20. There are certain requirements that must be met before euthanasia can take place, among which there must be unbearable suffering without prospects of improvement. In the Netherlands, euthanasia is performed by first administering a coma inducing drug (thiopental or propofol) intravenously. Next, when the patient is in a deep coma, a muscle relaxant (rocuronium) is administered intravenously, which causes cessation of breathing and ultimately death. The euthanizing physician is legally required to record the exact times of administration of the drugs and the time of death. The purpose of this exploratory study is to take the first steps in determining the applicability of data from commonly available wearable devices for estimating TOD. Hence, the research question of this paper is whether digital information from smartwatches allows for a reliable determination of the TOD, and, if so, with what margin? For this, patients who were scheduled to receive euthanasia were asked to wear a smartwatch during the procedure. Afterwards, we examined data recorded by these smartwatches for information allowing for estimation of TOD. Preliminary findings of this study have been published as an abstract21. Results Participants and euthanasia procedure In the period between June 2024 and August 2024 a total of 75 euthanizing physicians were approached for the study and eventually 28 participants agreed to be included in the study. Details on inclusion criteria, participant recruitment and ethical approval are described in “Methods”. One of the participants had deceased of natural causes shortly before the euthanizing procedure. Since the time of death was not observed and confirmed by the physician, this patient was excluded from the further data analysis. Table 1 gives an overview of the participant characteristics as obtained from the registration forms. Table 1Participant characteristics.Total (n = 27)Garmin (n = 9)Apple (n = 9)Samsung (n = 9)Sex (%)Male12 (44%)3 (33%)3 (33%)6 (67%)Female15 (56%)6 (67%)6 (67%)3 (33%)Age (years)Median77708177Range58–9258–9268–8764–92Underlying medical condition (%)Malignancy18 (67%)6 (67%)4 (45%)8 (89%)Neurological (including dementia)4 (15%)1 (11%)3 (33%)0Cardiovascular2 (7%)01 (11%)1 (11%)Pulmonary2 (7%)1 (11%)1 (11%)0Accumulation of old age-related diseases1 (4%)1 (11%)00 Participant characteristics. Besides the characteristics displayed in Table 1, information provided on registration forms showed that two participants had a pacemaker, which was active during euthanasia. Two participants received oxygen suppletion during the procedure. Information provided by the euthanizing physicians further showed that most of the participants had a light skin tone, except for one participant, who had a tanned skin. One of the participants was icteric. None of the participants had tattoos on the wrist the smartwatch was worn. The majority of participants (16) had a normal physique, eight were underweight, one overweight and two had edema. Information provided by euthanizing physicians showed that the mean time between administering the coma inducer and the TOD amounted to 6 min. In all cases there was a short period of on average 3 min between administering the coma inducer and the muscle relaxant. The registration forms showed that the physicians established death by testing neurological reflexes, observing respiration, palpating the pulse or auscultation of the heart. Information from the registration forms showed that some physicians relied on a single observation to determine the time of death, while others used a combination of multiple observations. Based on information reported by the physicians, neurological reflexes (such as the pupil reflex or corneal reflex/eyelid reflex) were tested 14 times, respiration was observed nine times, palpating the pulse (at the carotid artery or at the wrist) nine times, and the heart was auscultated 20 times. Smartwatch data Garmin Vivoactive 5 In nine experiments the participant wore the Garmin watch. In all cases, the watch produced consistent measurements throughout the period it was worn. Details on smartwatch configuration, data acquisition and processing can be found in “Digital methodology”. From analysis of processed data, it follows that the Garmin smartwatch has the capability to record data every minute and that time information is stored in the variable ‘timestamp’. Some biometric variables are collected every minute, while others are collected at a lower frequency. The following variables were identified as having a change in the period around the TOD, from which a Time of Last Registration (TLR) can be derived. The method employed to identify relevant variables from smartwatch data and the definition of TLRs can be found in “Data analysis”. heart_rate: during wearing, (almost) every minute there is a value corresponding to number of heart beats per minute. When not worn, the device registers a value of 0 every 16 min, with no data being registered in between. For TLR we use the value of ‘timestamp’ corresponding to the last registration of a heart_rate > 0 in the period that the watch is worn. reading_spo2: during wearing, (almost) every minute there is a value registered corresponding to the percentage of oxygenation. When not worn, the device produces no registration of reading_spo2. For TLR we use the value of ‘timestamp’ corresponding to the last registration of reading_spo2 in the period the smartwatch was worn, irrespective of the value being registered. respiration_rate: during wearing, (almost) every minute there is a value registered corresponding to number of breaths per minute. When not worn, the device produces registrations of respiration_rate equal to -2.0 every minute. For TLR we use the value of ‘timestamp’ corresponding to the last registration of respiration_rate > -2.0 in the period that the watch is worn. intensity: during wearing, there are registrations with variable time intervals in the range 0–10 min. When not worn, the device produces registrations with value 0 about every hour. For TLR we use the value of ‘timestamp’ corresponding to the last registration of intensity with any value. body_battery_status: during wearing, (almost) every minute there is registration with constant value of 1.0. When not worn, the device produces registrations with a constant value of either 2.0 or 126.0 every minute. For TLR we use the value of ‘timestamp’ corresponding to the last registration of body_battery_status with value 1.0 in the period the watch was worn. Figure 1 gives a typical example of registrations of these variables produced by the Garmin device during an experiment. Note that for all variables, the TLR derived from the data is later than TOD as determined by the physician. Typical example of the selected relevant variables from Garmin Data. The two dotted vertical lines indicate the times the smartwatch was put on and taken off. The solid vertical line indicates the TOD as determined by the physician. Time axis is relative to the time the watch was put on. For each variable, the blue dot indicates the datapoint selected for determining TLR. Data pertains to participant #4 in Table S1 of the supplementary material. Note that for this subject, there are registrations of variables after the watch is taken off, which are likely caused by movement of the watch The differences between TLR of the five variables described above and TOD determined by the physician are shown in the boxplots of Fig. 2. A table showing the differences between TOD determined by the physician and TLR as derived from the five variables described above for all participants is provided in Table S1 of the supplementary material to this paper. Boxplot of the difference (TLR-TOD) between TLR derived from Garmin Vivoactive 5 data and TOD as determined by physician for the five variables described in the main text. Outliers in the data are indicated by open dots. Apple Watch 9 In nine experiments the participants wore an Apple watch. Details on smartwatch configuration, data acquisition and processing can be found in “Digital methodology”. From analysis of processed data, it follows that the Apple smartwatch dynamically records data at variable time intervals ranging between a few seconds to ten minutes. For time information, we use the variable start_date. Based on the criteria from “Data analysis”, the following variables were identified as having a change in the period around the TOD. Heart Rate: during wearing, consecutive values are recorded at variable time intervals up to ten minutes. Most frequently, the time intervals were a few seconds and roughly five minutes. Heart rate is registered in beats per second, which we converted to beats per minute for comparison with the Garmin and Samsung devices. When not worn, the device registers no data. For TLR we use the value of ‘start_date’ corresponding to the last registration of Heart Rate in the period that the watch is worn. Active Energy Inspection of data reveals that the variable Active Energy shows some changes around the TOD. This variable contains noisy data which shows a slight trend transitioning from higher values before TOD towards lower values after TOD. Due to the noisy nature of the data, it is difficult to assign a TLR to this data. We did this by setting a threshold and used the value of ‘start_date’ corresponding to the last registration with a value of Active Energy < 0.5 in the period the watch was worn, but this is not necessarily very robust. Figure 3 gives a typical example of registrations of the variables produced by the Apple device during an experiment. In all experiments, Active Energy was measured continuously during the period the watch was worn. In six cases, the devices produced Heart Rate data continuously, while in the other three cases, little to no Heart Rate data was registered during the period the watches were worn. A clear explanation for this lack of data was not found. In this case, the TLR from the Apple data can be both earlier and later than TOD, as is indicated by positive and negative numbers. It is, however, important to note that looking at the Heart Rate data alone, it is currently difficult to distinguish between cases of continuous and discontinuous registrations. Typical example of the selected relevant data from Apple Watch Data. For explanation: see legend of Fig. 1. Data shown pertains to participant #5 in Table S2 in the supplementary material. Table S2 in the supplementary material contains a complete overview with data for all participants on the difference between TOD determined by the physician and TLR as derived from the variables Heart Rate and Active Energy. Figure 4 contains a boxplot of the differences between TLR from the Heart Rate and Active Energy and TOD as determined by the physician. Boxplot of the difference (TLR-TOD) between TLR derived from Apple Watch 9 data and TOD as determined by physician for the variables Heart Rate and Active Energy. Note difference in number of participants between the two. For explanation: see the main text Samsung Watch 6 Nine experiments were conducted, where the participant wore the Samsung watch. All experiments produced valid data. Details on smartwatch configuration, data acquisition and processing can be found in “Digital methodology”. From analysis of processed data, it follows that the Samsung smartwatch registers data at time intervals of ten minutes in the variable start__time. Sometimes a measurement can be missing, leading to larger intervals between consecutive data points. Based on the criteria from “Data analysis”, the following variable was identified as having a change in the period around the TOD. _heart__rate_heart__rate: during wearing, every 10 min there is a value corresponding to number of beats per minute. When not worn, the device registers no data. For TLR we use the value of ‘start__time’ corresponding to the last registration of a heartrate in the period that the watch is worn. Figure 5 gives a typical example of registrations of the variable produced by the Samsung device during an experiment. Typical example of the selected relevant variable from Samsung Data. For explanation: see legend of Fig. 1. Data shown pertains to participant #4 in Table S3 of supplementary material. Figure 6 shows a boxplot for the difference between TLR of _heart__rate_heart__rate and TOD determined by the physician. As can be seen from Fig. 6, the TLR from the Samsung data can be both earlier and later than TOD. In one participant there were still registrations at the moment the watch was removed from the wrist, about 30 min after TOD. In this case, it is unknown whether the TLR derived from the data is due to ceasing of heart rate or the removal of the watch. Therefore, this participant was excluded from further analysis. Table S3 in the supplementary material to this paper contains a complete overview with data for all participants on the difference between TOD determined by the physician and TLR as derived from the variable _heart__rate_heart__rate. Boxplot of the difference (TLR-TOD) between TLR derived from Samsung Watch 6 data and TOD as determined by physician for the variable _heart__rate_heart__rate. Effects of post-mortem movements, medical conditions and physical characteristics Post-mortem movement of the arm of the participant by physician or family was reported seven times on the registration forms. These were small movements such as lifting the arm. In all cases, there were no noticeable phenomena present in the corresponding smartwatch data that could be connected to these events. After removal and during transport of the watches, sometimes registrations of biometric data were observed in the Garmin device (see e.g. Figure 1). A likely explanation for this phenomenon is that measurements such as heart rate are triggered by movement of the device and therefore might show up as false positive registrations in such cases. From inspection of the currently available data, no marked differences in TLRs were identified when separating the study population into categories based on medical conditions and physical characteristics, such as skin color and physique. Due to the small sample size no statistical analysis was performed. Statistical analysis Table 2 provides basic descriptive statistics of the difference TLR-TOD for the variables identified in “Smartwatch data” for the three types of devices investigated in this study. Table 2Descriptive statistics of difference (TLR-TOD) for the variables identified in “Smartwatch data”.DeviceVariableMean (min)SD (min)95% CI (min)Garmin Vivoactive 5 heart_rate 7.45.0(3.6, 11.3) reading_spo2 9.26.5(4.2, 14.2) respiration_rate 10.86.7(5.6, 16.0) intensity 28.228.2(6.6, 49.9) body_battery_status 21.221.2(4.9, 37.5)Apple Watch 9 Heart Rate 2.35.3(-3.2, 7.9) Active Energy 8.521.0(-7.6, 24.6)Samsung Watch 6 _heart__rate_heart__rate -2.816.6(-16.6, 11.0)Data is represented in fractions of minutes. Positive numbers mean that the estimate from smartwatch data is after TOD and negative numbers that the estimate is before TOD. Mean = arithmetic mean, SD = standard deviation, 95% CI = 95% confidence interval of the mean. Descriptive statistics of difference (TLR-TOD) for the variables identified in “Smartwatch data”. Data is represented in fractions of minutes. Positive numbers mean that the estimate from smartwatch data is after TOD and negative numbers that the estimate is before TOD. Mean = arithmetic mean, SD = standard deviation, 95% CI = 95% confidence interval of the mean. Given the number of identified variables per smartwatch, Friedman test on differences between variables (TLR-TOD) from a single device was only carried out for the Garmin Vivoactive 5 device. This test was significant (p = 0.0043). Subsequent post-hoc pairwise comparisons showed significant differences between the variable body_battery_status and the variables heart_rate, respiration_rate and reading_spo2. Additionally, there was a significant difference between reading_spo2 and respiration_rate. In combination with Table 2, this indicates that the variable body_battery_status is the least accurate variable for this type of watch. For comparison between devices, we selected for Garmin Vivoactive 5 heart_rate, for Apple Watch 9 Heart Rate and for Samsung Watch 6 _heart__rate_heart__rate. Kruskal-Wallis test on these differences was not significant. Discussion In this study, we set out to collect experimental data to examine to what extent TOD can be estimated from smartwatch data. The results of this study indicate that, in cases where the smartwatch produced continuous measurements during the period it was worn, times of last registration (TLR) of variables fall within a time interval of roughly 30 min before or after TOD as determined by the physician. This suggests that smartwatch data has the potential to provide estimates of the TOD with smaller time intervals than the traditional methods based on rectal body temperature (Henssge), which usually yield time intervals of several hours (from ± 2.8 h up to ± 7 h). While thermal photogrammetry and modelling approaches show a relatively small interval of 0.26 ± 1.38 h for calculating the TOD, the applicability of these temperature-based methods is limited to the early postmortem period, as it can only be applied before the body temperature reaches the ambient temperature. In contrast, smartwatch data can be used in cases with a prolonged postmortem period. This suggest that the TOD may be estimated weeks, months of even years after death, provided that the data has been stored on the devices. Therefore smartwatch data could especially be useful in cases of a prolonged postmortem interval, where temperature-based methods are no longer applicable. In such cases, smartwatch data can be substantially more informative than TOD estimates based on observations of the body decomposition stages, which estimates the PMI in multiple days22. In this study, differences between types of examined smartwatches were found concerning time granularity of registrations and the number of identified variables providing information on the TOD. Such heterogeneity between devices is fairly common in the digital forensic domain. Together with the rapid succession of hardware and software versions of the same device it poses substantial challenges as it complicates direct comparison and limits generalizability. Besides this, our findings contain some currently unexplained phenomena that could potentially impact applicability. The Garmin Vivoactive 5 was found to register data with the highest time resolution and its data contains the most variables related to the TOD. TLRs derived from the data show a systematic tendency as they are almost all later than the TOD. For the Apple Watch 9, in three cases unexplained discontinuous Heart Rate registrations occurred. Since such cases are difficult to identify by looking at smartwatch data only, it might be somewhat risky to base an estimate of TOD solely on this data. Due to the noisy nature of Active Energy data, the estimate of TOD derived from this data involves a certain degree of uncertainty. For the Samsung Watch 6, only one variable related to the TOD was identified and the default time resolution of registration (10 min) was larger than for the other watches, thereby introducing a larger uncertainty in the estimate of TOD. An element of uncertainty in this study might be the time of death as determined by the physician. Since dying is a gradual process, time of death is not a specific moment in time and therefore, the reported value of TOD may depend on the method used to determine it. In the Netherlands, there is no fixed protocol on how to determine the moment at which death has occurred, leaving this at the discretion of the physician. The variability of methods employed by physicians in this study may lead to some spread in the ground-truth values of TOD, the size of which is difficult to quantify. This spread may account to some extent for the differences between ground-truth TOD and estimates of TOD obtained from smartwatch data. One of the limitations of this study is the relatively small number of participants. Nonetheless we find it particularly noteworthy that so many patients gave consent under the delicate circumstances associated with euthanasia. Small sample sizes limit the possibility for statistical analyses to determine the effects of, for example, physical characteristics. Therefore, some caution is warranted when interpreting the results and drawing strong conclusions should be avoided. Due to the study design, our study population likely differs in several aspects from the population encountered in forensic death investigations. First, the study population consisted of mostly elderly and terminally ill participants who had reduced mobility or often were bedridden, while the forensic population will be younger, more healthy and more physically active. Therefore, variables related to physical activity, such as number of steps or walking velocity, were not registered frequently by the smartwatches during the observation period. In forensic cases these variables potentially could provide additional information to estimate the TOD. Also, participants in our population died of respiratory arrest leading to cardiac arrest due to the administration of drugs during the euthanasia procedure, while in the forensic population there might be other causes, such as extensive bleeding leading to drastically reduced blood pressure with an absent peripheral pulse. It is currently unknown how smartwatches react under such conditions and to what extent this affects data produced by devices. All these considerations impact the generalizability of our study results and consequently, caution is warranted when extrapolating the present results directly to a forensic population. However, despite these differences, we believe current findings nonetheless provide meaningful insights into data registered by smartwatches during the dying process, the mechanisms of which are also likely to occur in forensic cases. In this study, we adopted a data-driven approach to selecting variables from smartwatch data providing information about the TOD. While this approach aligns with common digital forensic methodology, where traces are first identified and subsequently analyzed and validated to determine the inferences that can be drawn from them, at the same time it makes our findings mainly descriptive rather than explanatory. The extent to which observed changes in registrations by smartwatches are connected to underlying physiological changes or are caused by other reasons is currently unknown. Therefore, caution should be exercised when assigning a physiological or medical interpretation to these changes. This research presents the first experimental data connecting registrations from smartwatches to ground-truth TOD and indicates the potential of using smartwatch data in post-mortem investigations. Given our current level of understanding and the limitations discussed above, the method shows potential to provide information on the TOD and to serve as a complementary source of information when combined with other techniques. All of the above considerations suggest that more research needs to be conducted in order to obtain a more complete understanding of the value of smartwatches for death investigations. Methods Ethics approval In the Netherlands, scientific research involving patients is subject to the Medical Research Involving Human Subjects Act (Dutch: Wet Medisch Wetenschappelijk Onderzoek, WMO) if there is an infringement of the physical and/or psychological integrity of the participant23. In those cases, the study must be reviewed by an accredited Medical Research Ethics Committee (MREC). In cases where it is not clear immediately if the WMO applies to that study, an MREC can perform a review to establish if a study is subject to the WMO. For this research, the Research Ethics Committee of the Maxima Medisch Centrum reviewed the study protocol (reference N24.017) and subsequently confirmed that the WMO does not apply to this research meaning further official approval by the committee was not required. Nevertheless, due to potential ethical considerations, the study protocol was reviewed and approved by the Director Science and Technology of the Netherlands Forensic Institute (reference 25-DIR-131). The study was conducted according to the principles of the Helsinki Declaration. Participant recruitment For inclusion in this study, patients had to be at least 18 years of age, have an approved euthanasia procedure scheduled, and had to be able to wear a smartwatch around their wrist. In the Netherlands, physicians are required to notify forensic physicians of a planned euthanasia. It was at this stage, that information about the study was provided to the euthanizing physicians. Physicians who showed interest in the study were subsequently approached by the researcher. The physician was asked to make an assessment whether the patient might be willing to consider participation and if the patient was sound of mind and capable of making an informed decision regarding participation. If so, the patient was approached by their physician about the study. In case of any doubt, the patient was not approached. Information folders about the study were provided and patients could make the decision whether or not to participate. Informed consent was obtained from all patients who decided to participate in the study. Digital methodology Selected smartwatches For this study, three types of smartwatches were selected: Garmin Vivoactive 5 (software version 10.20), Apple Watch 9 (WatchOS 10.2) and Samsung Watch 6 (Wear OS 4.0). Selection of these specific devices was made based on input from digital specialists from the Dutch police as to which devices were the most popular, widely used, and where there was a method for data extraction. For practical reasons and to reduce the possibility of irregularities due to individual device behavior, multiple identical devices were used in this study for each type of smartwatch. Smartwatch configuration Before each experiment, the smartwatch was fully reset to factory settings to avoid possible influence of previous experiments on device behaviour. For the same reason, data from previous experiments was removed from the paired mobile phones for the Apple and Samsung watches (see “Smartwatch data acquisition and - processing”). After factory reset, smartwatches were manually configured using a fixed protocol, which included switching off all device behaviour that could disturb the participant, such as visual or vibrational notifications or alarms. Measurements which needed to be started manually (e.g. workouts or ECG measurements) were not used, so that no further actions regarding the smartwatch were required from the participants or their families. We used the default settings for the frequency of registrations of measurements by the smartwatches. Types of data collected by the devices were left as they were after the factory reset, with the exception that we enabled continuous measurement of oxygen saturation on the Garmin smartwatch and continuous measurement of Stress on the Samsung smartwatch to collect these potentially relevant data. After configuration, time on the device was synchronized. For Apple and Samsung devices, this was done by connecting to the paired mobile phone, which was connected to a Wi-Fi network. For the Garmin device, time was synchronized through GPS using the Timesync option in the configuration menu. Smartwatch data acquisition and - processing For the Garmin Vivoactive 5 smartwatch, data was acquired directly from the smartwatch by connecting it to a computer, where via the Media Transfer Protocol (MTP) the watch appears as separate drive ‘Vivoactive 5’. All relevant data is in the folder named GARMIN, which was copied as a whole from the drive. From pilot experiments it follows that the subfolder GARMIN/Monitor contains files with information that is automatically collected by the smartwatch in proprietary Garmin .FIT format. Files in this folder were processed and converted to .csv files for further analysis using scripts developed for this study. Here, variable names were assigned, based on information from the software developer kit provided by Garmin and analysis of the Garmin Connect App. As it is currently not possible to acquire health data directly from Apple Watch 9 and Samsung Watch 6 devices, these devices were paired to a mobile phone. For Apple smartwatches, an iPhone 11 (i0S 17.4) and an iPhone 15 (iOS 17.1) were used for this purpose and for Samsung smartwatches a Samsung Galaxy S22 (Android 13) and Samsung Galaxy S22+ (Android 12) were used. After pairing via Bluetooth, data from the smartwatches was synchronized with the mobile phones, where it was stored in databases connected to health-related apps. For iPhone this database is healthdb_secure.sqlite, related to the Apple Health App, and for Samsung telephones it is SecureHealthData.db, related to the Samsung Health App. After synchronization, full file system extractions of the mobile telephones were produced using commercial forensic tools and from these extractions the databases mentioned above were obtained. Data in the database SecureHealthData.db is encrypted and was therefore first decrypted using commercial forensic tools before further processing. Finally, information from the databases healthdb_secure.sqlite and SecureHealthData.db pertaining to the smartwatches was further processed and converted to .csv files for further analysis using scripts developed for this study. In processing and data conversion of Apple and Samsung devices, original variable names as assigned by the manufacturer were retained in the processed data. Protocol during euthanasia Participants were given a randomly selected smartwatch (configured as described in “Smartwatch configuration”) by the researcher well in advance of the planned euthanasia and were asked to wear it at least three hours before the euthanasia procedure. Participants were instructed to wear the smartwatch on the opposite arm to the one where the IV for drug administration was located. The euthanizing physician was asked to remove the smartwatch approximately half an hour after the participant’s death. The euthanizing physician was requested to fill out a registration form, providing information about: the participant’s characteristics (e.g. sex, age, relevant medical history and possible use of medical or therapeutic aids, such as a pacemaker or oxygen suppletion), the participant’s characteristics (e.g. sex, age, relevant medical history and possible use of medical or therapeutic aids, such as a pacemaker or oxygen suppletion), the participant’s physique, skincolor and presence of tattoos on the wrist where the smartwatch was worn, as these might potentially affect measurements from the PPG sensor of the smartwatch. the participant’s physique, skincolor and presence of tattoos on the wrist where the smartwatch was worn, as these might potentially affect measurements from the PPG sensor of the smartwatch. the specific drugs administered during the euthanasia and the times of administration. the specific drugs administered during the euthanasia and the times of administration. the time of death as determined by the physician and the method used to determine this. The latter is relevant, since there is no standardized protocol for this in the Netherlands. the time of death as determined by the physician and the method used to determine this. The latter is relevant, since there is no standardized protocol for this in the Netherlands. time of putting on and taking off the smartwatch from the participant, as well as any post-mortem movement of the arm wearing the smartwatch. time of putting on and taking off the smartwatch from the participant, as well as any post-mortem movement of the arm wearing the smartwatch. After the experiment, the device was collected from the euthanizing physician and was re-checked to see whether the device’s clock showed the correct time. After that, data was secured, stored and processed under a code that could not be traced back to a specific individual. Data analysis Data analysis was limited to the period in which the smartwatch was worn by the participant, as derived from information on the registration form. After the extraction and processing of the data, the following steps were taken when analyzing the produced .csv datafiles to derive an estimate of the TOD from smartwatch data. Identification of relevant variables from smartwatch data:For each type of smartwatch separately, we identified by visual inspection of the data registered variables that could potentially be suitable to provide information about the TOD. For this, we looked in the .csv files for variables that:have multiple registrations with associated timestamps during the period the smartwatch was worn.show changes in values around the period of the TOD as determined by the physician. Identification of relevant variables from smartwatch data:For each type of smartwatch separately, we identified by visual inspection of the data registered variables that could potentially be suitable to provide information about the TOD. For this, we looked in the .csv files for variables that:have multiple registrations with associated timestamps during the period the smartwatch was worn.show changes in values around the period of the TOD as determined by the physician. For each type of smartwatch separately, we identified by visual inspection of the data registered variables that could potentially be suitable to provide information about the TOD. For this, we looked in the .csv files for variables that:have multiple registrations with associated timestamps during the period the smartwatch was worn.show changes in values around the period of the TOD as determined by the physician. For each type of smartwatch separately, we identified by visual inspection of the data registered variables that could potentially be suitable to provide information about the TOD. For this, we looked in the .csv files for variables that:have multiple registrations with associated timestamps during the period the smartwatch was worn.show changes in values around the period of the TOD as determined by the physician. have multiple registrations with associated timestamps during the period the smartwatch was worn. have multiple registrations with associated timestamps during the period the smartwatch was worn. show changes in values around the period of the TOD as determined by the physician. show changes in values around the period of the TOD as determined by the physician. Feature extraction from variables around the TOD: Feature extraction from variables around the TOD: For each identified variable, a criterion was defined to objectively quantify observed changes around the TOD from which a timestamp could be derived. Henceforth, this timestamp will be referred to as Time of Last Registration (TLR). Note that for one device, there can be multiple TLRs, derived from different variables registered by that device. For each device, results are visualized in the form of boxplots of the difference TLR-TOD for each of the identified relevant variables. For each device, results are visualized in the form of boxplots of the difference TLR-TOD for each of the identified relevant variables. Note that this approach is purely data-driven and that selection of variables was based on the criteria described above rather than on the physiological meaning of variables (such as, for instance, heart rate). Because of this approach, our focus was restricted to identifying variables showing changes around the TOD and to describing the observed changes to derive a TLR from that, without addressing the underlying origin of these changes. Figure 7 illustrates the process of identification of relevant variables and determining TLR. The process of identification of relevant variables and determination of Time of Last Registration (TLR) from that variable. The solid blue line shows a variable having multiple registrations around the time of death, indicated by TOD. Note that this variable shows a change around the instant of TOD which is indicated by the dotted line, where the variable changes from a low value to a higher value. This change makes it a potentially relevant source of information about TOD. From the change, a time of last registration (TLR) can be derived, which then serves as an estimate of TOD. In this case, TLR would be the timestamp associated with the last registration of the low value. The corresponding datapoint is marked as the dark blue circle in the figure with the text TLR. Observed changes around TOD and the criterion developed in this research to derive a TLR from that change can be different for various variables registered by smartwatches and are described in detail in “Smartwatch data”. Code implementing determination of TLRs for identified variables for the three types of smartwatches examined in this study is available online (10.5281/zenodo.21235744). Statistical analysis In this study, a power analysis could not be performed because reliable estimates of effect size were unavailable at the time of study design, given the exploratory nature of this research and non-existent literature on the topic. To evaluate applicability for estimating TOD and time ranges basic descriptive statistical analysis was performed on the differences between the TLRs derived from the features identified as described in “Data analysis” and TOD, as determined by the physician. Mean, standard deviation and confidence interval for the mean were calculated. Inferential statistical analysis focused on the difference (TLR-TOD) for each variable. A non-parametric Friedman test was used to test whether the differences have the same probability distribution across variables within a single brand of smartwatch if there are at least three features. Only in case the overall test result was significant, post-hoc analysis was done with a Wilcoxon signed-rank test for paired samples without correction for multiple comparisons. To compare differences (TLR-TOD) between devices, for each device the feature having the lowest mean was selected and a non-parametric Kruskal-Wallice test was performed. In case of a significant result, post-hoc pairwise comparisons were made using a Mann-Whitney U test without correction for multiple comparisons. We used a confidence level of 95% for all intervals and tests. Because of the small sample size, it was decided not to perform statistical tests aimed at investigating certain effects of post-mortem movements, medical conditions and physical characteristics on the variables. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Acknowledgements Assistance during data acquisition and processing by Loes Quirijnen and Thijs van Meurs is gratefully acknowledged. We are indebted to Marjan Sjerps for useful discussions and help on statistical analysis, to Susan Laraghy for technical advice concerning smartwatches and commenting on a draft of the paper, and to Bela Kubat for useful scientific discussions. We are grateful to employees from Public Health Service Oost-Nederland and the NFI for organisational and logistical support during patient inclusion and data collection. We would especially like to thank all of the participants, their families and their physicians for providing the possibility to conduct this research. Author contributions This research was conducted as part of the forensic physician training program of TG at NSPOH. TG: study design, data collection, data analysis, manuscript writing and editing. JPvZ: study design, data analysis, manuscript writing and editing. BL: conceptualization, study design, manuscript writing and editing. All authors reviewed the manuscript. Data availability The dataset generated in the current study is not publicly available due to privacy reasons but is available from the corresponding author on reasonable request. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1 Data Availability Statement The dataset generated in the current study is not publicly available due to privacy reasons but is available from the corresponding author on reasonable request.