A Non-Intrusive Method for Lonely Death Prevention Using Occupancy Detection and an Anomaly Detection Model

Abstract

In countries like Japan, Australia, France, Denmark, and South Korea, the numbers of single-person households and older adults living alone have been steadily increasing each year, leading to the social issue of lonely deaths among older adults. Against this backdrop, this study proposes a method to develop a system for preventing lonely deaths based on information technology, including the Internet of Things (IoT). IoT sensor data, which include nine environmental variables such as indoor temperature, relative humidity, CO₂ concentration, fine dust particle levels, illuminance, total volatile organic compound levels, and occupancy data collected from passive infrared sensors, provide empirical evidence so that anomalies can be detected in the behavior patterns of older adults when they remain in one place for an unusually long time. Detecting such risky situations for older adults living alone involves anomaly detection through occupancy monitoring. The data from occupancy monitoring were analyzed using four classification models, namely Logistic Regression, k-Nearest Neighbor, Decision Tree, and Random Forest, with the performance of occupancy detection being compared across these models. Furthermore, the method proposed in this study includes data processing for environmental variables to improve the performance of occupancy detection.

1. Introduction

The numbers of single-person households and older adults living alone have been increasing each year in several countries, including South Korea. Notably, the issue of lonely deaths among older adults living alone has become a social concern. Without the support of a spouse or children, older individuals living alone find it challenging to effectively address daily life risk factors such as illnesses or accidents. Furthermore, many older adults prefer to remain in their own homes, where they have spent much of their lives, rather than moving to care facilities. This preference often results in social isolation, and the delayed response to emergency situations could potentially lead to lonely deaths.

Among the population aged 65 years or above, the percentage of older individuals living alone, i.e., the percentage of single-person households, has steadily increased from 19% in 2017 to 21.1% in 2023 in Korea [1]. Consequently, the Ministry of Health and Welfare of South Korea has been actively working to establish a reliable safety net and develop measures to prevent lonely deaths.

Against this backdrop, many researchers investigated a method to develop a lonely death prevention system based on information technology (IT), including Internet of Things (IoT), without demanding much cost or labor. In this study, empirical evidence was found by identifying cases showing anomaly patterns in the behavior patterns of older adults using the IoT sensor data of nine environment variables: indoor temperature, relative humidity, CO₂ concentration, three levels of fine dust particles, illuminance, the level of total volatile organic compounds (TVOCs), and occupancy big data collected from a passive infrared (PIR) sensor. The significance of occupancy detection has been confirmed as it identifies an anomaly pattern in the behaviors of older adults when they remain stationary for an extended period. Additionally, an efficient method of carrying out occupancy detection was developed utilizing a machine learning (ML) technique.

For occupancy detection, collecting occupancy data using a radio frequency identification (RFID) tag [2,3,4] could lead to privacy intrusion, as participating older individuals are required to carry a wearable tag at all times. Similarly, collecting occupancy data using Bluetooth Low Energy (BLE) may also raise privacy issues and [5,6] could cause a dispute over privacy intrusion as older adults have to carry the BLE beacon. In this study, occupancy detection and anomaly detection based on occupancy detection were performed to detect emergency situations of older adults living alone by applying a non-intrusive method to collect indoor environmental data. The findings of this study would be invaluable in the implementation of a lonely death prevention system for older individuals living alone. Previous research has explored methods to reduce energy consumption by controlling illuminance and ventilation through occupancy detection. [7,8]. Sayed et al. [9] designed an edge-based occupancy detection system that utilizes non-intrusive ambient data and a deep learning model. Aliero et al. [10] compiled a public set of training datasets for building occupancy profile prediction and compared occupancy prediction accuracies of five ML algorithms. The authors of [11] introduced a novel platform architecture integrating an IoT platform to collect sensors’ streaming data and machine learning algorithms implemented in the server site for application to streaming and non-stationary data. While numerous studies have focused on occupancy detection to enhance energy management in buildings [12,13,14,15,16,17], none have specifically gathered occupancy data and utilized ML models to infer participants’ behaviors to predict risk situations for older adults living alone. Consequently, this study plays a crucial role in addressing the prevention of lonely deaths among older adults living alone. The occupancy detection method using ML algorithms proposed in this study and the anomaly detection method based on occupancy detection do not provide a high level of accurate risk detection. However, the significance of this study is that it provides a methodology for non-intrusively detecting risky situations for people living alone in a field where it is difficult to detect risky situations by analyzing complex human behavior patterns.

This paper is structured as follows: Section 2 describes the indoor environmental data collected in 1 min intervals using an IoT sensor for 38 days from the bedrooms, living rooms, and toilets in the homes of the participating older adults living alone. The models used in occupancy detection and anomaly detection for the collected data are also explained. Section 3 accounts for the measured performance of the models for occupancy detection, and Section 4 presents the conclusions of this study.

To derive these results, we established the following research questions:

▪

What algorithms are available that can develop a non-intrusive lonely death prevention system based on information technology, including Internet of Things, without demanding much cost or labor?

▪

If anomaly patterns are detected in the behaviors of the older adults, occupancy detection must be effective. Among ML algorithms, which model has excellent occupancy detection performance, and what data processing methods are available to improve occupancy detection performance?

2. Materials and Methods

2.1. Data

In this study, an IoT sensor was used to obtain environmental data from the bedrooms, living rooms, and toilets of the apartment homes of the older adults living alone. The data were collected in 1 min intervals for 38 days from 1 May to 7 June 2023, and the results are presented in

Table 1. Nine variables of the indoor environment measured in the homes of older adults living alone.

Variable (Variable Name in Input Data)	Definition and Unit
Temperature (temp)	Indoor air temperature (°C)
Relative humidity (humi)	Relative humidity (%)
CO2 (co2)	Carbon dioxide concentration (ppm)
Dust_pm_0.1 (dust_pm_1)	Ultrafine dust particles of ≤0.1 μm diameter (μg/m3)
Dust_pm_1.0 (dust_pm_10)	Fine dust particles of ≤1.0 μm diameter (μg/m3)
Dust_pm_2.5 (dust_pm_25)	Fine dust particles of ≤2.5 μm diameter (μg/m3)
Illuminance (illuminance)	Illuminance (lux)
TVOC (voc)	Total volatile organic compound level (ppb)
PIR_in_Room (in_room)	Occupancy status collected from PIR sensor

.

Table 2. IoT sensor specification.

Measurement Factors of Indoor Environment	Range	Image of the Sensor
Temperature (°C)	−40~125
Relative humidity (%)	0~100
CO2 (ppm)	0~5000
Dust_pm_0.1, Dust_pm_1.0, Dust_pm_2.5 (μg/m3)	0~5000
Illuminance (lux)	0~1000
TVOC (ppb)	0~29,206
PIR_in_Room	0 or 1

is a detailed table describing the performance of the IoT sensor used to collect indoor environmental data.

The IoT sensor was installed in the bedrooms, living rooms, and toilets in the homes of older adults living alone, as shown in Figure 1. The environmental data presented in Table 1 were collected, while occupancy label data were also obtained to develop a suitable ML model for occupancy detection by installing a CCTV in each room.

To examine the cases of risk situations faced by older adults living alone, anomaly phases were detected using a two-stacked long short-term memory (LSTM) anomaly detection (LSTM-AD) model that was used in a previous study [18]. For the enhanced performance of the LSTM-AD model through adequate learning, the occupancy vector was converted into a cumulative occupancy vector to conduct an experiment on anomaly detection. In the collected data(bedroom occupancy, living room occupancy, and toilet occupancy), a set of behavior vectors is defined as follows:

$$\left\{\left(i, j, k\right)| i=0 or i=1, j=0 or j=1, k=0 or k=1\right\}$$

The cumulative occupancy vector is obtained by converting the data based on the time the older adults stay in a single place for more than 1 min. With the cumulative occupancy vector $y_t=\left(y_{t,1},y_{t,2},y_{t,3}\right)$ as the input data and the cumulative occupancy vector $\widehat{y_t}=\left(\widehat{y_{t,1},} \widehat{y_{t,2}}, \widehat{y_{t,3}}\right)$ predicted by the LSTM-AD model, the anomaly score is defined as the sum of the MSE of each component as follows:

$$anomaly score=\frac{1}{n}{\displaystyle \sum }_{i=1}^n\left\{{\left(\widehat{y_{i,1}}-y_{i,1}\right)}^2+{\left(\widehat{y_{i,2}}-y_{i,2}\right)}^2+{\left(\widehat{y_{i,3}}-y_{i,3}\right)}^2\right\}$$

Figure 2 demonstrates the LSTM-AD model, and Figure 3 shows the experimental result for anomaly detection on the behavior patterns of the participating older adults.

As determined using the anomaly detection analysis, the cases where the anomaly score was high included spending ≥ 400 min in the bedroom and ≥25 min in the toilet. By setting the threshold of the anomaly score at a higher level, it becomes possible to detect more unusual cases as outliers. In summary, instances of older adults staying in one location for an exceptionally prolonged period were identified as the anomaly behavior pattern. Therefore, as depicted in Figure 3, the detection of risk situations for older adults living alone involves anomaly detection through occupancy monitoring.

2.2. Model

As described in Section 2.1, occupancy detection should be performed accurately to ensure that the result of anomaly detection through occupancy detection is reliable. In Section 2.2, the ML models used in occupancy detection are described. In occupancy detection, it is necessary to classify occupancy and non-occupancy from the input of environmental data. Hence, four classification models, namely Logistic Regression (LR), k-Nearest Neighbor (k-NN), Decision Tree (DT), and Random Forest (RF) models, were used to measure and compare the performance of occupancy detection. Each model, as described in a previous study, can be explained as described below [19].

2.2.1. LR Model

The LR model is a method of modeling the conditional probability $P\left(Y=1|x_1,x_2,\cdots ,x_k\right)$, as shown in Equation (1), where $X_1, X_2, \cdots , X_k$ are explanatory variables, and $Y$ ($Y$ is $0$ or $1$) is the binary response variable.

$$log\left[\frac{P\left(Y=1|x_1,x_2,\cdots ,x_k\right)}{1-P\left(Y=1|x_1,x_2,\cdots ,x_k\right)}\right]=\alpha +{\beta }_1{x}_1+{\beta }_2{x}_2+\cdots +{\beta }_k{x}_k$$

(1)

From the training dataset, intercept $\alpha$ and the effect of $x_i$, namely ${\beta }_i$ $\left(i=1, 2, \cdots ,k\right)$, are estimated with maximum likelihood estimation.

It is an algorithm that solves the problem of binary classification as it considers new data as $Y=1$ if $P\left(Y=1|x_1,x_2,\cdots ,x_k\right)$ obtained from prediction Equation (1) is greater than the pre-defined threshold and $Y=0$ if it is smaller than the threshold.

2.2.2. k-NN Model

The k-NN algorithm is an ML algorithm, and it is a non-parametric supervised learning method. When new data are given, the nearest $k$ is extracted from existing data using the distance measuring metric $d$. Based on this, the class of the new data for classification problems and the predicted value of the new data for regression problems are predicted. The hyperparameters of the k-NN model include the number of searchable neighbors $k$ and metrics $d$. The Euclidean distance, Manhattan distance, Mahalanobis distance, correlation distance, and rank correlation distance are denoted as $d$.

2.2.3. DT Model

The DT model is a supervised ML technique used to solve regression or classification problems. Predictions are achieved by organizing decision rules into a hierarchical tree structure. DT models are created using training data and consist of a hierarchy of branches where explanatory variables are represented as nodes and the feature spaces are categorized into non-overlapping groups based on certain conditions. Each internal node represents a test for an attribute (e.g., whether a coin toss results in heads or tails), each branch represents the test result, and each leaf node represents a class label (decision taken after computing all attributes). The paths from the root to leaf represent classification rules. The final nodes at the bottom display the dependent variables as classified groups.

2.2.4. RF Model

The RF model is an ensemble learning method, which is an ML algorithm that uses multiple DTs. The RF algorithm is a method for solving regression and classification problems. For classification problems, the class predicted by most DTs is produced as the output. For regression problems, the average of the predicted values of each DT is produced as the output. The RF algorithm is designed to solve the overfitting problem in the training dataset for the DT. A new DT is generated by dividing the training dataset into several parts and randomly selecting a predefined number of explanatory variables for each of the training dataset parts. This method reduces variance in the model by learning several DTs and averaging the prediction values from the DTs.

3. Performance Results

This section presents the measured occupancy detection performance of the four classification models, LR, k-NN, DT, and RF, which were previously described in Section 2.2. In this study, the models provided by the Scikit–Learn library were used. The confusion matrix in

Table 3. Confusion matrix.

		Prediction
		Positive (Non-Occupancy)	Negative (Occupancy)
Real	Positive (non-occupancy)	tp	fn
	Negative (occupancy)	fp	tn

was generated to obtain the scores of accuracy and recall on occupancy and the f1 scores of occupancy and non-occupancy.

The input data of the four models, LR, k-NN, DT, and RF, are shown in

Table 4. The input data of the four models: LR, k-NN, DT, and RF (data in the bedroom).

Regdate	Temp	Humi	co2	Dust_pm_1	Dust_pm_25	Dust_pm_10	Illuminance	voc	in_Room
1 May 2023 7:48	24.3	74	659	26	26	27	4	418	1
1 May 2023 7:49	24.3	74	653	25	25	27	4	417	1
1 May 2023 7:50	24.3	74	654	25	25	27	4	407	1
1 May 2023 7:51	24.3	74	657	25	25	27	4	402	1
1 May 2023 7:52	24.4	74	669	25	25	27	4	402	1
1 May 2023 7:53	24.7	73	676	25	25	27	4	402	1
1 May 2023 7:54	24.5	73	679	24	24	26	4	399	1
1 May 2023 7:55	24.4	74	700	24	24	26	4	394	1
1 May 2023 7:56	24.4	74	708	24	24	26	4	391	1
1 May 2023 7:57	24.4	74	707	23	23	24	4	391	1

, and the output data of the models are shown as the occupancy status for each time.

Table 5. Ratio of time that older adults living alone spent in each room.

Place Stayed	Outside	Toilet	Bedroom	Living Room
Percentage of time spent (%)	48.93	1.02	20.33	29.72
Daily standard deviation (%)	29.81	0.54	11.45	16.94

shows the average value of the ratio of time that the older adults living alone spent in each room of the house and outside and the standard deviation on a daily basis for the ratio of time spent.

Table 6. Occupancy detection performance of four models with raw data as input data.

Toilet
Model	tp	fp	fn	tn	accuracy	negative_recall	positive_F1_score	negative_F1_score
LR	25,943	724	0	0	0.9729	0.0000	0.9862	not defined
k-NN	25,937	724	6	0	0.9726	0.0000	0.9861	not defined
DT	25,066	706	877	18	0.9406	0.0249	0.9694	0.0222
RF	25,885	722	58	2	0.9708	0.0028	0.9852	0.0051
Living Room
Model	tp	fp	fn	tn	accuracy	negative_recall	positive_F1_score	negative_F1_score
LR	7519	6696	260	248	0.5275	0.0357	0.6837	0.0666
k-NN	4237	3758	3542	3186	0.5042	0.4588	0.5372	0.4661
DT	4099	3718	3680	3226	0.4975	0.4646	0.5256	0.4658
RF	4193	3837	3586	3107	0.4958	0.4474	0.5305	0.4557
Bedroom
Model	tp	fp	fn	tn	accuracy	negative_recall	positive_F1_score	negative_F1_score
LR	15,417	6489	2	0	0.7037	0.0000	0.8261	not defined
k-NN	12,971	5523	2448	966	0.6362	0.1489	0.7650	0.1951
DT	10,701	4573	4718	1916	0.5759	0.2953	0.6973	0.2920
RF	12,598	5346	2821	1143	0.6272	0.1761	0.7552	0.2187

shows the measured occupancy detection performance in the toilet, living room, and bedroom, with the raw data of environment variables given as the input data for the models in temporal order.

The results suggest that the model exhibiting generally high values for negative_f1_score and negative_recall that indicate the occupancy detection performance in the three rooms is the DT model. However, as the values were not high enough for the model to be used in an actual occupancy detection system, this study suggests a method to enhance the occupancy detection performance through data processing. To improve the occupancy detection performance, only data identified by the PIR sensor as occupancy during the period of 1–20 May were utilized as the training data input. The ratio of training data to testing data was set at 50:50, with data from 1 May to 27 May being utilized as the training data.

Table 7. The occupancy detection performance of the four models in the case of prioritizing the data of occupancy detected using the PIR sensor.

Toilet
Model	tp	fp	fn	tn	accuracy	negative_recall	positive_F1_score	negative_F1_score
LR	5846	419	0	0	0.9331	0.0000	0.9654	not defined
k-NN	5827	419	19	0	0.9301	0.0000	0.9638	not defined
DT	5407	380	439	39	0.8693	0.0931	0.9296	0.0870
RF	5781	411	65	8	0.9240	0.0191	0.9605	0.0325
Living Room
Model	tp	fp	fn	tn	accuracy	negative_recall	positive_F1_score	negative_F1_score
LR	0	1	568	1525	0.7283	0.9993	not defined	0.8428
k-NN	98	201	470	1325	0.6796	0.8683	0.2261	0.7980
DT	184	449	384	1077	0.6022	0.7058	0.3064	0.7211
RF	114	208	454	1318	0.6839	0.8637	0.2562	0.7993
Bedroom
Model	tp	fp	fn	tn	accuracy	negative_recall	positive_F1_score	negative_F1_score
LR	37	51	179	226	0.5335	0.8159	0.2434	0.6628
k-NN	72	98	144	179	0.5091	0.6462	0.3731	0.5967
DT	81	124	135	153	0.4746	0.5523	0.3848	0.5416
RF	89	122	127	155	0.4949	0.5596	0.4169	0.5546

presents the measured performance when prioritizing occupancy data detected using the PIR sensor for the training data.

With the data of occupancy detected using the PIR sensor being prioritized for the training data, the values of negative_recall and negative_f1_score that indicate the occupancy detection performance were greatly improved compared to the results shown in Table 6. In the case of the toilet, the values of negative_recall and negative_f1_score increased on average by 428% and 414%, respectively. In the case of the living room, the mean increases were 733% and 341%, respectively, and in the case of the bedroom, the mean increases were 213% and 148%, respectively. The positive_f1_score value was reduced in a trade-off relationship compared to the use of raw input data; nonetheless, as it is crucial for the occupancy detection performance to be enhanced, the method of prioritizing the data of occupancy detected using the PIR sensor for the training data was more suitable for the purpose of this study. However, the occupancy detection performance in the toilet was substantially low, suggesting that it would be more suitable to use the PIR sensor with a 98% accuracy. In the case of the living room, the occupancy detection performance of the RF model, represented as negative_recall and negative_f1_score, were high at 0.8637 and 0.7993, respectively. In the case of the bedroom, the k-NN model had high negative_recall and negative_f1_score values at 0.6452 and 0.5967, respectively. Occupancy in the living room and the bedroom, but not the toilet, was reliably detected through the k-NN, DT, and RF models. Having prioritized the occupancy data for the training data, the measured performance was shown to have been markedly enhanced.

4. Conclusions

In numerous countries, including South Korea, the numbers of single-person households and older individuals living alone have been steadily increasing each year, resulting in the social issue of lonely deaths among older adults. Therefore, this study aimed to explore a method for developing a lonely death prevention system based on IT, including the IoT. An analysis of IoT sensor data, which included nine environmental variables such as indoor temperature, relative humidity, CO₂ concentration, three levels of fine dust particles, illuminance, TVOC levels, and occupancy data collected from PIR sensors, provided empirical evidence that anomaly patterns in the behaviors of older adults could be detected when they remained in one place for an unusually extended period. Consequently, anomaly detection through occupancy monitoring emerges as a critical factor in identifying risk situations for older individuals living alone.

For occupancy detection, the performance of four classification models, namely LR, k-NN, DT, and RF, was measured and compared. The results demonstrate that, compared to using raw environmental data as the input data of the models in temporal order, prioritizing the data of occupancy detected using the PIR sensor for the training data led to enhanced occupancy detection performance, as represented in the values of negative_recall and negative_f1_score, which increased, on average, by approximately 492% and 293%, respectively. The model exhibiting a generally high performance of occupancy detection using raw input data was the DT model. The PIR sensor measurements revealed that the k-NN, DT, and RF models produced reliable and outstanding detection results on occupancy in the living room and the bedroom (but not the toilet) upon prioritizing the PIR sensor data of occupancy for the training data.

The proposed method of occupancy detection and anomaly detection through occupancy detection in this study is anticipated to be valuable in the development of a lonely death prevention system for older individuals living alone. However, it is worth noting that the performance of occupancy detection, particularly in the case of the toilet, was found to be low. This finding suggests the need to either utilize a PIR sensor with 98% accuracy or conduct further research in this area as a follow-up study.