A Two-Tier Approach for Fall Detection Systems

Abstract

- A critical performance drawback of most fall detection systems is high false alarms. These false alarms are due to the imbalanced mix of the "fall" and "non-fall" data contained in the processed datasets on one hand, and the inherent limitation of the processing algorithms, on the other hand. To tackle this false alarm problem, a two-tier solution approach which entails Synthetic Minority Over-Sampling Technique (SMOTE) and hybrid of two machine learning algorithms (Multiple-Kernel Support Vector Machine (MK-SVM) and Multinomial Naive Bayes (MNB), hereafter known as SMOTE-based MKSVM-MNB is proposed. The results of simulation experiments performed using two open-source datasets namely SisFall Dataset and UMAFall Dataset show that SMOTE-based MKSVM-MNB significantly outperforms MKSVM, MNB and MKSVM-MNB in terms of the number of False Negatives (FN) recorded. Also, MKSVM-MNB significantly outperforms MKSVM and MNB in terms of FN.

Headnote

Abstract - A critical performance drawback of most fall detection systems is high false alarms. These false alarms are due to the imbalanced mix of the "fall" and "non-fall" data contained in the processed datasets on one hand, and the inherent limitation of the processing algorithms, on the other hand. To tackle this false alarm problem, a two-tier solution approach which entails Synthetic Minority Over-Sampling Technique (SMOTE) and hybrid of two machine learning algorithms (Multiple-Kernel Support Vector Machine (MK-SVM) and Multinomial Naive Bayes (MNB), hereafter known as SMOTE-based MKSVM-MNB is proposed. The results of simulation experiments performed using two open-source datasets namely SisFall Dataset and UMAFall Dataset show that SMOTE-based MKSVM-MNB significantly outperforms MKSVM, MNB and MKSVM-MNB in terms of the number of False Negatives (FN) recorded. Also, MKSVM-MNB significantly outperforms MKSVM and MNB in terms of FN.

Keywords: activities of daily living, fall detection, false negatives, imbalance, machine learning, oversampling.

I.INTRODUCTION

Adverse effects of falls include injuries such as; fractures, open wounds, bruises, sprains, joint dislocations, brain injuries, and strained muscles [11]. As falls occur more frequently amongst the elderly population, a twelve month study of 2,096 elderly people aged over 65 years in Nigeria discovered that, 23% of the all elderly people in the country experience a fall every year. From this estimate, 45% of women and 30% of men sustain serious injuries, including hip fractures [2].

Falls have also been reported as part of the common occurrences in COVID-19 patients undergoing intensive care, while being monitored in hospital isolated units or homes [7]. Besides the cost that could be involved in treating a fall victim, a lot of the negative impacts emerge when a fall is not detected early enough. A fall victim untended after a fall is more likely to develop complications such as, gastrointestinal bleeding, urinarytract infections, pneumonia, ulcer, myocardial infarction, and chest pain [3].

Moreover, most fall victims who were unable to get up in time, have been reported to have a decrease in their ability to perform basic activities of daily living for three consecutive days, after the fall incidence [9]. This means that help must be provided as soon as the victim experiences a fall in order to reduce its resultant complications.

In order to assist in providing prompt medical attention for fall victims, diverse fall detection systems have been proposed over the years [6]. These fall detection systems make use of data generated from monitoring the ADL of the subjects in order to identify an unusual activity such as a fall.

Fall detection systems can be grouped into the following three categories: wearable, ambient, and vision based systems. This grouping is based on the hardware platform through which the set of data being processed to detect a fall is gathered. Wearable based systems make use of accelerometers and gyroscopes in detecting falls, ambient based systems analyze vibrations and sound in order to distinguish a fall from normal activity, while vision based systems use cameras and image processing algorithms to monitor a person's body posture or motion so as to detect a fall. The problem with ambient device and vision based systems is the restriction of its use to the room where the sensors or cameras are placed.

However, with the wearable system, data can be gathered at whatever position the subject is, within the environment where the ADL is carried out. Approaches based on wearable sensors, even though being more intrusive, are therefore more suitable for research in this problem domain, as they provide a more realistic dataset [6].

Fall detection systems often identify data indicating a fall as positive and data indicating a usual ADL as negative. A defect in the processing of the ADL dataset can result in the production of a high rate of false positives (false alarms) or false negatives (a situation where the system does not give an alarm when there is a fall occurrence).

The more accurately this dataset can be processed, the more effective the system will be in detecting falls, thereby assisting caregivers in providing better assistance to the patients under their care, and reducing the time expended on attending to false emergency (fall) alarms. With the aim of minimizing false alarms, this paper proposes a two-tier approach of data mining namely SMOTE-based MKSVM-MNB for fall detection systems.

II.RELATED WORK

[1] presents an unobtrusive smart phone based fall detection system that uses a combination of information derived from machine learning classification using Decision Trees. The data from the smart phone built-in accelerometer is continuously screened when the phone is in the user's belt or pocket in order to detect a fall. [10] proposed a real-time, multi-space, multi-camera tracking system for a fall detection system with all the deriving qualities, i.e. monitoring any area regardless its size by the use of multiple depth sensors that retain workers' individual privacy. The introduced detection system is based on key features that characterize a fall, such as vertical velocity and area variance, while the falling process is modeled by HMM. [12] proposed an enhanced fall detection system based on body smart sensors, implementing and deploying it successfully to detect accidental falls in homes. By using data from an accelerometer, smart sensor and cardio tachometer, the impacts of falls were successfully distinguished from ADL, reducing the false detection of falls. [5] proposed a fall detection algorithm which uses a very simple and computationally efficient set of features extracted from a publicly available dataset. The extracted features are used to train and test four machine learning classifiers namely: Decision Tree (DT), Support Vector Machine (SVM). Logistic Regression (LR) and K-Nearest Neighbor (KNN) with SVM giving the highest accuracy. [8], presents a pervasive fall detection system (FallDroid), used on a Smartphone. The system exploits a two-step algorithm proposed to monitor and detect fall events using the embedded accelerometer signals. Comprising of the threshold based method and multiple kernel support vector machine, the proposed algorithm is capable of identifying fall-like events (such as lying on a bed or sudden stop after running) and reduce false alarms. [4] addresses the detection of human falls using relevant pixel-based features reflecting variations in body shape. Specifically, the human body is divided into five partitions that correspond to five partial occupancy areas. For each frame, area ratios are calculated and used as input data for fall detection and classification. In this regard, an effective fall detection approach using generalized likelihood ratio (GLR) scheme is designed. However, a GLR scheme cannot discriminate between true falls and like-fall events, such as lying down. To mitigate this limitation, they applied the support vector machine algorithms on features of the detected fall, in order to recognize the type of fall. Tests on two publicly available datasets show the effectiveness of the proposed approach to appropriately detecting and identifying falls.

Algorithm 1: SMOTE Algorithm

Input: Number of minority class samples T; Amount of SMOTE N%; Number of nearest neighbors k Output: (N/100) · T synthetic minority class samples

1. (· If N is less than 100%, randomize the minority class samples as only a random percent of them will be SMOTE d. ·)

2. if N < 100

A critical drawback in fall detection systems is the imbalanced ratio of 'fall' data to 'non-fall' data within the ADL dataset. This is due to an uneven distribution of classes present in the ADL data as a whole, with the non-fall data being more. This results in the generation of large number of false negative alarms, which occur when the fall detection system indicates a 'non-fall' even when a fall has actually occurred (naive result), answering to the class with more data, and sometimes giving the fall detection system a false level of accuracy (which is often high). A 'fall' incident, indicated as 'non-fall' incident, can prevent the fall victim from getting the necessary assistance from a caregiver. This can lead to the fall victim experiencing some of the negative impacts of falling. This paper addresses this issue by using a Synthetic Minority Over-Sampling Technique (SMOTE)-based machine learning approach.

III.RESEARCH METHODOLOGY

A.Description of the Proposed Model

The methodology employed in this study is depicted in the block diagram shown in Figure 1. The methodology developed for this study is a SMOTEbased FDS model which makes use of SMOTE-based MKSVM-MNB (Multi-Kernel Support Vector Machine - Multinomial Naive Bayes) algorithm. It consists of an over-sampling technique called SMOTE and a hybrid of two machine earning techniques namely; a MultipleKernel Support Vector Machine and a Multinomial Naive Bayes Classification Algorithm. The data mining techniques adopted are shown in the order of the data acquisition, data normalization, feature extraction, oversampling and classification. The experimental approach used in this study is described in two stages, which are: The data pre-processing stage oversampling and the data classification and performance evaluation stage. The fall detection process depicted in the model involves: the removal of outliers from the dataset using the z-score technique, extraction of the most significant features from the dataset, using the Minimum Redundancy Maximum Relevance Algorithm and classification of the dataset using the SMOTE-Based MKSVM-MNB approach which involves the use of SMOTE and a hybrid of the MK-SVM and MNB Algorithms.

SMOTE: This technique provides the oversampling of the dataset in order to Balance the dataset.

MKSVM-MNB: this is the Hybrid the MKSVM and MNB Algorithms without Data Oversampling feature.

The algorithm 1 shows the processes performed by the SMOTE algorithm in balancing the dataset. In algorithm 2, the balanced dataset is obtained and the set of pairs of data representing fall and non-fall data is identified. The initial SVM parameters and the kernel function used is defined and the minimum margin is computed. The training process then begins. The dataset is divided into two folds: the training and testing set, applying a 10-fold cross validation. The error function is computed and the SVM kernels are combined. The Multi Kernel SVM optimizes the synthetic (balanced) dataset to obtain the most optimal SVM kernel (of the Linear, Gaussian and Sigmoid). The MNB function interprets the kernel combination of the parameters as random variables (puts priors on them and varies them) in order to obtain an optimal result (subset) from the dataset. This optimal result is then used to predict a new class label in order to obtain data indicating a fall or non-fall event. Figure 2 shows the flowchart of MKSVM-MNB.

IV.RESULTS

A. Classification of SisFall Dataset Using the MKSVM Algorithm

In the classification of the SisFall dataset using the MKSVM, a list of predicted and actual fall class values with zero (0) depicting "non-fall" and one (1) depicting a "fall" is generated. Figure 3 shows some of these class values, and also the values depicting the 3-dimensional orientation of the subject of the fall while Figure 4 shows the graphical representation of the predicted and actual values.

B. Classification of SisFall Dataset Using the MNB Algorithm

In the classification of the SisFall dataset using the MNB, a list of predicted and actual fall class values with zero (0) depicting "non-fall" and one (1) depicting a "fall" is generated. Figure 5 shows some of these class values, and the values depicting the 3-dimensional orientation of the subject of the fall while Figure 6 shows the graphical representation of the predicted and actual values.

C. Classification of SisFall Dataset Using MKSVMMNB

In the classification of the SisFall dataset using MKSVM-MNB, a list of predicted and actual fall class D. Classification of SisFall Dataset Using SMOTEbased MKSVM-MNB

In the classification of the SisFall dataset using the SMOTE-based MKSVM-MNB list of predicted and actual fall class values with zero (0) depicting "non-fall" and one (1) depicting a "fall" is generated. Figure 9 shows some of these class values, and also the values depicting the 3-dimensional orientation of the subject of the fall while Figure 10 shows the graphical representation of the predicted and actual values.

E. Classification of UMAFall Dataset Using the MKSVM A lgori thm

In the classification of the UMAFall dataset using the MKSVM, a list of predicted and actual fall class values with zero (0) depicting "non-fall" and one (1) depicting a "fall" is generated. Figure 11 shows some of these class values, and also the values depicting the 3dimensional orientation of the subject of the fall while Figure 12 shows the graphical representation of the predicted and actual values.

F.Classification of UMAFall Dataset Using the MNB Algorithm

In the classification of the UMAFall dataset using the MNB, a list of predicted and actual fall class values with zero (0) depicting "non-fall" and one (1) depicting a "fall" is generated. Figure 13 shows some of these class values, and also the values depicting the 3dimensional orientation of the subject of the fall while Figure 14 shows the graphical representation of the predicted and actual values.

In the classification of the UMAFall dataset using MKSVM-MNB, a list of predicted and actual fall class values with zero (0) depicting "non-fall" and one (1) depicting a "fall" is generated. Figure 15 shows some of these class values, and also the values depicting the 3dimensional orientation of the subject of the fall while Figure 16 shows the graphical representation of the predicted and actual values.

H. Classification of UMAFall Dataset Using SMOTEbased MKSVM-MNB

In the classification of the UMAFall dataset using the MKSVM-MNB with the SMOTE oversampling technique, a list of predicted and actual fall class values with zero (0) depicting "non-fall" and one (1) depicting a fall is generated. Figure 17 shows some of these class values, and also the values depicting the 3dimensional orientation of the subject of the fall while Figure 18 shows the graphical representation of the predicted and actual values.

From the results in Table 1, the following observations are evident:

(1)The number of FNs obtained for the MKSVMMNB is considerably lower than that obtained for MKSVM and MNB.

(2) Observation in (1) shows that in using the MKSVM-MNB the system performs much better at indicating a situation where a fall event has occurred, and when in reality it has truly occurred, than indicating a situation where no fall has occurred, when in reality a fall event has occurred, as against using MKSVM and MNB individually. This justifies the use of the hybrid algorithm.

(3) The number of FNs obtained for the SMOTE-based MKSVM-MNB is considerably lower than that obtained for MKSVM-MNB, MKSVM and MNB.

(4) Observation in (3) shows that in using the SMOTE-based MKSVM-MNB, the system performs much better at indicating a situation where a fall event has occurred, and when in reality it has truly occurred, than indicating a situation where no fall has occurred, when in reality a fall event has occurred, as against using MKSVMMNB or MKSVM and MNB individually.

From the observations above, the numbers of FNs obtained for the SMOTE-Based MKSVM-MNB algorithm is the lowest, as compared those obtained when the MKSVM-MNB algorithm was used, and in the case where the MKSVM and MNB were used individually. This shows that the developed SMOTEbased system indicates more often, a situation where a fall event has occurred, when in reality, it has truly occurred; than it indicates a situation where a fall event has not occurred when in reality a fall event has occurred, thereby providing more reliable fall detection.

From the results in Table 2, the following observations are evident:

(1) The number of FNs obtained for the MKSVMMNB is lower than that obtained for MKSVM and MNB.

(2) Observation in (1) shows that in using the MKSVM-MNB the system performs much better at indicating a situation where a fall event has occurred, and when in reality it has truly occurred, than indicating a situation where no fall has occurred, when in reality a fall event has occurred, as against using MKSVM and MNB individually. This justifies the use of the hybrid algorithm.

(3) The number of FNs obtained for the SMOTE-based MKSVM-MNB is considerably lower than that obtained for MKSVM-MNB, MKSVM and MNB.

(4) Observation in (3) shows that in using the SMOTE-based MKSVM-MNB, the system performs much better at indicating a situation where a fall event has occurred, and when in reality it has truly occurred, than indicating a situation where no fall has occurred, when in reality a fall event has occurred, as against using MKSVMMNB or MKSVM and MNB individually.

From the observations above, the numbers of FNs obtained for the SMOTE-Based MKSVM-MNB hybrid algorithm is the lowest, as compared those obtained when the MKSVM-MNB hybrid algorithm was used, and in the case where the MKSVM and MNB were used individually. This shows that the developed SMOTEbased system indicates more often, a situation where a fall event has occurred, when in reality, it has truly occurred; than it indicates a situation where a fall event has not occurred when in reality a fall event has occurred, thereby providing more reliable fall detection.

V. CONCLUSIONS

The developed SMOTE-based system indicates more often, a situation where a fall event has occurred, when in reality, it has truly occurred; tha has not occurred when in reality a fall event has occurred, making it more reliable. Thus, incorporating SMOTE into MKSVMMNB holds a great promise for developing a reliable fall detection system.

References

REFERENCES

[1] Aguiar B., Rocha T., Silva J. & Sousa I. (2014). Accelerometer-based fall detection for smartphones. In Proceedings of the 2014 IEEE International Symposium on Medical Measurements and Applications (MeMeA), Lisboa, Portugal, 11-12 June 2014; pp. 1-6.

[2] Bekibele C. & Gureje O. (2010). Fall incidence in a population of elderly persons in Nigeria. Gerontology, 56(3):278-283.

[3] Fleming J. & Brayne C. (2008). Inability to get up after falling subsequent time on floor and summoning help: Prospective cohort study in people over 90, Biomed. J., 337 (2227), 1279-1282.

[4] Harrou F., Zerrouki N., Sun Y. & Houacine A. (2019). An Integrated Vision-Based Approach for Efficient Human Fall Detection in a Home Environment, IEEE Access, Volume 7,pp 114966 - 114974, 10.1109/Access.2019.2936320.

[5] Hussain F., Umair M., Muhammad E., PiresI., ValenteT., GarciaN., & Pombo N, (2018). 9th International Conference on Sensor Device Technologies and Applications, At Venice, Italy.

[6] Mubashir M., Shao L., & Seed L. (2013). A survey on fall detection: Principles and approaches, Neurocomputing, 15(6), 16-23.

[7] Norman R.,Stall N., & Sinha, S. (2020). Typically Atypical: COVID-19 Presenting as a Fall in an Older Adult. Journal of the American Geriatrics Society,68(7).https://doi.org/10.1111/jgs.16526.

[8] Shahzad A. & Kim K., (2019). FallDroid: An Automated Smart Phone based Fall Detection System using Multiple Kernel Learning. IEEE Trans. Ind. Inform. 2019, 15, 3544.

[9] Tinetti M. (2003). Preventing falls in elderly persons. National England Journal of Medicine. 348, 42-49.

[10] Triantafyllou D., Krinidis S., Ioannidis D., Metaxa, I., Ziazios C., Tzovaras D. (2016). A real-time fall detection system for maintenance activities in indoor environments. IFAC-PapersOnLine 49 (28), 286-290.

[11] Vianda S., Johannes H., Smith M., & Paul L. (2004). Consequences of falling in older men and women and risk factors for health service use and functional decline. Age and Ageing, 33(1), 58-65.

[12] Wang Y., Wu K., & Ni L. (2017). WiFall: Device-Free Fall Detection by Wireless Networks. IEEE Trans. Mob. Comput. 2017, 16(2), 581-594.