The growing elderly population and increase in single-person households in Seoul have raised concerns about social isolation and solitary deaths. To address these issues, Seoul has implemented a “smart plug” system that monitors household electricity usage and light levels. If no changes are detected over a specific period, alerts are sent to local monitoring centers. While this system has proven effective for early detection of deceased individuals, it underscores the need for a proactive approach to prevent solitary deaths. A comprehensive framework is necessary to identify signs of social isolation early, leveraging IoT sensors and welfare data to monitor at-risk households and offer tailored services. Integrating data from various care services could help bridge existing service gaps, significantly contributing to the establishment of a safety net.
This study analyzed data from 260 households using smart plugs, managed by Seoul’s Smart Care team, to evaluate the system's effectiveness. Users were categorized into regular and crisis situations, with the latter including instances of early death discoveries and other intervention cases. The analysis revealed that most early discovery cases involved men in their 60s, particularly in the Guro district. While smart plugs were effective for early death detection, the findings indicate a need for a more systematic approach to managing at-risk groups that extends beyond age-based criteria. Examination of time-series electricity usage data revealed distinct patterns based on appliance type, which could serve as indicators of irregularities in daily routines, facilitating early identification of crisis symptoms.
By analyzing changes in electricity usage to discern activity patterns, criteria for reinstalling smart plugs were established. Daily usage correlations were measured to differentiate between consistent and irregular patterns. Households with consistently low correlation values often indicated irregular behavior, suggesting opportunities for early crisis intervention. Clustering analysis using DBSCAN showed that while most users maintained regular routines, households in the care request group exhibited irregularities, reinforcing the need for regular wellness checks. Based on these results, a criterion for smart plug reinstallation was developed, identifying 30.8% of households as needing new or additional installations due to unusual usage patterns, thereby helping to prevent device malfunctions and enhancing service reliability.
An XGBoost model was developed to predict crisis situations based on smart plug data collected at 10-minute intervals. This machine learning model was trained to differentiate between regular and crisis scenarios, including early death discoveries, emergency dispatches, and care requests. XGBoost is renowned for its high performance with complex datasets, achieving an AUC value of 0.873 and an accuracy rate of 80.1% on the validation set, which indicates strong predictive capability. ROC analysis confirmed that the AUC of 0.873 signifies excellent model performance, as values above 0.8 are generally regarded as favorable, while values below 0.6 suggest limited predictive capacity. Based on these results, the model’s alert system successfully classified households at risk with an accuracy of 78.3%, providing timely notifications when crisis signals persisted for over 24 hours.
To further enhance early detection of prolonged inactivity, a time-series forecasting model was employed to monitor extended periods of inactivity (exceeding 50 hours). This approach yielded an impressive AUC value of 0.946, reducing the alert interval from the previous 50 hours to a more responsive 17-hour window. Among the three AI models developed, the first focused on detecting device malfunctions to recommend reinstallation as needed, thereby improving IoT sensor reliability and reducing unnecessary operational costs. The second model targeted irregular activity patterns, enabling preemptive wellness checks for households showing signs of crisis. The third model enhanced response times by leveraging time-series predictions to detect inactivity early, facilitating rapid intervention and optimizing support delivery for socially isolated households.
Despite the effectiveness of the smart plug system in crisis detection, several operational and data management challenges persist. The system often triggers false alarms when not installed on frequently used devices, which increases staff workload. Additionally, users occasionally unplug the smart plug to conserve electricity, which undermines the system's cost-effectiveness. Because alerts rely on inactivity detection, the system is limited in its ability to respond to immediate emergencies, such as strokes or heart attacks. Data management issues also exist, as data is inconsistently handled between municipalities and service providers. Municipalities retain only alarm logs, while raw data remains solely with service providers. This separation restricts secondary data use and analysis, since raw data is only preserved for one year, hindering long-term insights and service improvement.
To enhance the smart welfare check system, the study recommends standardizing data formats across service providers for consistent and secure data management. User information, safety status, and service usage patterns should follow a common data standard, enabling real-time data exchange through open APIs. Seoul or related organizations could establish an integrated information system to manage this standardized data, complemented by periodic audits to ensure compliance with the standards. Furthermore, data handling guidelines must be implemented at every service stage: from initial data collection based on user consent and real-time monitoring for abnormal signs to secure data deletion or archival upon service termination. Personalized services should be tailored to factors such as health status, Wi-Fi availability, and smart meter compatibility. For example, high-risk individuals could benefit from wearable fall-alert devices, while households without Wi-Fi might use smart plugs with built-in communication features. Regularly updated risk criteria, based on surveys and risk pattern data, will aid in delivering customized care services for socially isolated households, thereby ensuring a more responsive and effective welfare system.