A Real-World Smart Home Dataset Integrating Sleep, Environmental, Physiological, and Ambient Sensing for Homecare Research

1. Summary

The growing demand for homecare technologies and assisted living systems has accelerated the development of smart home environments capable of continuously monitoring daily activities, physiological conditions, and environmental information [1,2]. Multimodal smart home datasets play a central role in supporting research on behavior monitoring, sleep analysis, activity recognition, health-centric home healthcare systems, and intelligent decision support applications [3,4,5]. However, many publicly available datasets are limited by short monitoring durations, controlled laboratory environments, limited detection methods, or inadequate integration of physiological and environmental information [6,7]. The comparison with publicly available datasets refers to research datasets released for reuse, including laboratory or scripted-activity datasets and smart-home datasets with more limited duration or modality coverage. Longitudinal datasets collected in real-world residential environments remain relatively scarce, especially those that integrate ambient sensing, wearable device data, sleep data, and environmental information into a unified framework [8,9].

This paper introduces a longitudinal multimodal dataset, constructed from data collected over approximately one month from 10 participants in their real-world living environments. Ten participants were selected purposively from older adults receiving homecare services who were willing to contribute residential sensor data. Preference was given to people aged 65 years or older who lived alone, consistent with the intended homecare context. One participant living with a spouse and two children was retained to provide a contrasting multi-person household scenario. The sample size reflected the feasibility-oriented nature of the deployment and the practical requirements of installing, supporting, and monitoring multiple sensing systems in private homes. Participants knew that sensing devices had been installed, their behavior may initially have been influenced by the deployment. Data collection began during the deployment period, and no separate adaptation phase was excluded. Consequently, a short-term observation or Hawthorne effect cannot be ruled out, particularly during the initial days of monitoring.

The sensor infrastructure integrates various IoT devices, including motion detectors, door sensors, environmental sensors, wearable devices, and under-mattress sleep tracking mats. Environmental sensing refers to measurements describing the physical indoor environment, including temperature and humidity. Ambient sensing refers to unobtrusive sensors embedded in the home environment that monitor occupant interactions and activities, such as motion sensors and door contact sensors. These sensors collectively provide continuous, timestamped measurements of room occupancy, routine patterns, indoor interactions, sleep states, heart rate, respiratory rate, snoring, temperature, humidity, and other environmentally relevant data.

This data was collected to support data analytics research in smart home, homecare monitoring, behavioral analysis, home assistance, and digital health applications. The published data retains the natural variability and operational characteristics of real-world residential sensor deployments, including heterogeneity in event frequencies and real-world user behavior. For ease of reuse and interoperability, the raw sensor observation data has been converted from JSON (JavaScript Object Notation) event streams to structured CSV files suitable for time series analysis, statistical modeling, and machine learning applications.

2. Data Description

The dataset is organized as timestamped records and stored in CSV format. Each observation corresponds to a sensor measurement collected by a specific sensing device at a specific time. All participants were anonymized using non-identifiable identifiers, and no personally identifiable information is retained in the published dataset. Recorded observations include a timestamp, anonymized participant identifier, sensor identifier, measurement type, and corresponding sensor value. Table 1 lists the main attributes included in the dataset structure.

This detection infrastructure integrates a variety of heterogeneous sensors deployed throughout the residence. Motion detectors are installed in some rooms to track occupant activity and movement. Once these sensors detect movement within the monitored area, they generate event-based observation data, enabling analysis of movement patterns, room occupancy, and transitions between rooms. Additionally, door open/close sensors are installed on specific objects and entry points within the residence, such as the refrigerator, bathroom door, and front door to capture interactions related to the occupants’ daily behaviors and habits.

Environmental sensors reported indoor temperature and humidity observations at device dependent intervals. This data provides contextual information about room conditions and changes in indoor comfort throughout the monitoring period. In addition to environmental sensing, wearable physiological monitoring was performed using smartwatch-based sensing devices capable of recording heart rate, step count, and calories measurements. The dataset also includes measurements taken under the mattress, capturing sleep-related behaviors such as transitions between different sleep states, breathing information, heart rate, and snoring. Table 2 describes the sensing devices, published variables, measurement units, acquisition mechanisms, reporting behavior, triggering conditions, and value coding. Because internal hardware sampling rates were not recorded in the deposited files, the table distinguishes device-dependent reporting from the intervals observed in the normalized CSV export.

Figure 1 shows the total number of records collected by each participant in the dataset. The number of observations varied among users due to differences in daily routines, house layout, sensor triggering frequency, and temporary interruptions in data collection during deployment. Observation counts varied with home layout, the number and placement of installed sensors, sensor-triggering behavior, participant routines, and temporary data interruptions. Consequently, differences in record counts should not be interpreted directly as differences in participant activity. Users 3 and 5 generated the most observations, while users 1, 7, 2, and 9 generated relatively fewer. Nevertheless, all participants provided substantial multimodal longitudinal data, covering approximately one month of follow-up.

Table 3 shows the characteristics and monitoring information of the participants. The ’Homecare Visits’ column reports the participants’ typical scheduled homecare visits, not visits by the research team. The metadata indicate that all homes received the same core modalities; motion-sensor counts varied from three to five according to residence layout, while each participant had three door-contact sensors, two environmental sensors, one wearable, and one sleep sensor. During the monitoring period, user3 lived in multi-person households and user6 lived with a dog. Consequently, data collected by ambient sensors such as motion detectors and door sensors, at times reflect the activities of other household members rather than solely those of the monitored participant. This is an inherent limitation of passive ambient sensing systems deployed in shared living environments. In contrast, wearable devices and sleep tracking data obtained from under-mattress sensors are specific to the individual participant, as these sensors are dedicated to their exclusive use. Therefore, such physiological and sleep-related data remain unaffected by the activities of other household members.

Figure 2 illustrates the distribution of observational data across different detection methods. The WithingsIntraDay stream contributed the largest modality-specific count in the normalized release (181,725 records), followed by motion3 (141,865) and WithingsSleepRaw (141,380). Counts reflect measurement multiplicity, device-reporting behavior, sparsely populated measurement-specific fields, and event frequency and should not be interpreted as sampling-rate comparisons. In contrast, event-driven detection methods (such as door sensors) generated less observational data because data was only recorded during explicit interactions. Motion sensors generally generated a large amount of residential-related observational data, reflecting daily activities and room changes within the residential environment. Environmental detection methods (such as temperature and humidity sensors) generated a moderate density of observational data through periodic environmental measurements.

The published database contains raw sensor event files, sensor deployment information, and supplementary documentation to facilitate data reproducibility and reuse. The dataset also includes a metadata file describing each participant, sensor deployment, recording period, and sensor availability. This data is available in CSV format to ensure compatibility with common data analysis environments and machine learning frameworks, including Python, R, MATLAB, and other time series processing platforms.

3. Methods

3.1. Data Collection Environment

The data collection architecture is designed to ensure global consistency in deployment configuration across all residential environments, while accommodating differences in apartment size, room layout, and participant living conditions. All participants are equipped with the same core detection devices, including wearable physiological monitoring systems, under-mattress sleep tracking mat, environmental sensors, and door contact sensors. However, the number of motion sensors deployed varies depending on the physical characteristics of each residence. Smaller apartments are equipped with at least three motion sensors, while larger residences are equipped with up to five, ensuring adequate coverage of the living room, bedrooms, kitchen, and bathroom.

Figure 3 is an illustrative layout rather than a common floor plan. Residence size, room configuration, household composition, and installed motion-sensor count varied. These contextual differences limit direct comparison of raw event counts across participants. Motion sensors were deployed in each room such as bedrooms, bathrooms, living rooms, and kitchens to detect the presence and activity of people. Door sensors were installed at specific doors and entrances within the house, such as the refrigerator, bathroom door, and front door, to track interactions related to daily activities. Environmental sensors continuously monitored the temperature and humidity of specific rooms. Wearable physiological monitoring devices recorded heart rate, calories, and steps. Finally, sleep tracking mat placed under the mattress monitored sleep-related activities, including bed occupancy, breathing signals, and different sleep states. Figure 4 shows the one instance of each sensor type that is used to collect data.

3.2. Data Acquisition and Timestamp Standardization

Throughout the monitoring period, the sensing infrastructure continuously generates event-related observational data with timestamps. Sensor data is acquired asynchronously based on changes in environmental conditions, participant movement, physiological activity, or specific sensor behaviors. Sensor data is transmitted via gateway-based local communication infrastructure and then stored in a cloud-based digital repository for subsequent processing and analysis as shown in Figure 5. The temporal density of the observation data varies depending on the sensing method, as different sensing devices operate at different sampling rates and triggering conditions.

Available timestamps from the different sensor streams were converted to UTC during preprocessing. This conversion provides a common time representation across participants and sensing modalities but does not, by itself, demonstrate that the internal clocks of all sensing devices were synchronized. Device-level clock synchronization and clock drift were not independently assessed, and no temporal offsets were estimated or corrected during preprocessing. In addition, delays may have occurred while observations were transmitted from the sensing devices through the local gateway to the cloud platform. Therefore, the dataset should be described as having timestamps standardized to UTC rather than as being fully time-synchronized.

3.3. Data Preprocessing and Cleaning

Sensor data is collected and stored in JSON format as a timestamped event stream generated by heterogeneous sensing devices. Each JSON record contains information such as a timestamp, anonymous participant ID, sensor and gateway IDs, measurement category, and corresponding sensor value. During preprocessing, a Python preprocessing script parses the JSON records and converts them to CSV format. The nested data structure is flattened into a standardized tabular representation with uniform columns for storing timestamps, users, sensor IDs, measurement types, and sensor values. The timestamp field for all sensing modes is converted to a standardized UTC date and time representation to provide a common timestamp representation across devices and participants.

After conversion, sensor identifiers and measurement labels were standardized to unify the naming conventions across the entire dataset. Observational data from different detection methods (motion, contact, and environmental sensors, wearable physiological monitoring devices, and sleep monitoring systems) were then merged into a unified tabular structure, preserving their original timestamps and sensor associations. Because the detection infrastructure employs an event-driven acquisition strategy, different sensor methods generate observational data at varying frequencies based on participant activity and environmental conditions. Consequently, physiological and sleep monitoring devices generate relatively dense, continuous data streams, while motion and contact sensors generate sparse, asynchronous, event-driven observational data.

3.4. Data Validation

Data quality was evaluated separately for each participant and sensor stream. The validation examined whether each expected sensor was present, the first and last observation times, the number of observations, the number of active recording days, prolonged interruptions, exact duplicate rows, malformed records, and invalid timestamps. Expected monitoring duration was obtained from the participant metadata, whereas actual data availability was determined from the timestamps contained in the participant CSV files.

Empty fields within a valid Withings record were not classified as missing sensor data. The Withings export may provide a value for one measurement, such as heart rate, while the corresponding steps or calories fields are empty at the same timestamp. In these cases, the device produced a valid observation, and at least one measurement was available. Missing or unavailable data were instead defined at the sensor-stream level: an expected sensor was considered unavailable when it was entirely absent, stopped reporting before the end of the observation period, or showed a prolonged interruption relative to its usual reporting pattern. For event-driven motion and door-contact sensors, low event counts alone were not considered evidence of missing data because a lack of events may reflect genuine inactivity. Complete absence and prolonged interruptions were assessed in relation to the other sensors installed in the same residence.

The validation identified several participant-specific availability issues. For user1, no observations were available from doorcontact1 during the monitoring period. For user4, doorcontact2 contained an interruption of approximately 11 days, and the temp1 environmental stream contained an interruption of approximately 2.2 days. For user5, the WithingsSleepRaw stream ended on 24 November 2025, before the end of the monitoring period, and sleep data were available for 18 nights. In contrast, all installed sensor streams for users6-user10 were represented on every available dataset day. Wearable-data days and available sleep-monitoring nights were reported as data-availability indicators and should not be interpreted as verified device adherence.

The normalized participant files contained 1,033,970 timestamp-valid observations. A total of 1,654 exact duplicate rows were identified, while no malformed normalized rows or invalid timestamps were detected. The duplicate rows were retained in the distributed files and are documented by the accompanying validation code. No interpolation, synthetic padding, or resampling was performed to replace unavailable periods. Participant-level availability and quality results are summarized in Table 4.

4. User Notes

This dataset is intended for academic and research use in data analysis related to smart homes, home assistance, homecare monitoring, behavioral analysis, human activity recognition, and longitudinal health sensor research. The published data preserves the original temporal structure and multimodal characteristics of real-world residential sensing environments, supporting downstream applications such as statistical analysis, machine learning, deep learning, time series modeling, and digital twin development. No independently annotated activity ground truth is included. Sensor outputs such as motion occupancy, door state, sleep state, and steps are device-derived observations rather than external activity labels. Consequently, the dataset directly supports unsupervised, self-supervised, routine-modeling, and anomaly-detection [5,10].

To improve accessibility and interoperability, the JSON event streams were converted to participant-level CSV files while preserving exported timestamps, sensor identifiers, measurement types, and values. Available timestamps were standardized to UTC; this common representation does not demonstrate device-level clock synchronization. No interpolation, synthetic padding, or resampling was performed. The distributed CSV files retain missing values and 1,654 exact duplicate rows, which are identified by the accompanying validation code rather than silently removed. The preprocessing and validation code is deposited alongside the dataset. It validates the five-column schema, parses UTC timestamps, identifies empty measurement fields and stream-level availability interruptions, quantifies temporal gaps, and generates participant and sensor-level quality tables without overwriting the released CSV files.

5. Conclusion

This dataset contributes longitudinal multimodal observations from real residences, combining ambient, environmental, wearable, and sleep sensing while preserving asynchronous temporal structure and real-world missingness. The accompanying code and quality tables make duplicates, unavailable periods, temporal interruptions, and modality-specific coverage explicit. Reuse should account for heterogeneous home layouts, household attribution, device-dependent reporting, the absence of independently annotated activities, and the distinction between UTC standardization and device-clock synchronization.