CareFall: Automatic Fall Detection through Wearable Devices and AI Methods

Abstract

The aging population has led to a growing number of falls in our society, afecting global public health worldwide. This paper presents CareFall, an automatic Fall Detection System (FDS) based on wearable devices and Artificial Intelligence (AI) methods. CareFall considers the accelerometer and gyroscope time signals extracted from a smartwatch. Two diferent approaches are used for feature extraction and classification: i) threshold-based, and ii) machine learning-based. Experimental results on two public databases show that the machine learning-based approach, which combines accelerometer and gyroscope information, outperforms the threshold-based approach in terms of accuracy, sensitivity, and specificity. This research contributes to the design of smart and user-friendly solutions to mitigate the negative consequences of falls among older people.

Keywords

• eol>fall detection system
• accelerometers
• classification algorithms
• machine learning
• wearable sensors

1. Introduction

Population aging is increasing worldwide. The World Health Organization considers falls among the elderly to be a major global public health challenge [ 1 ]. In fact, falls can adversely afect the quality of life in older people, causing them serious physical, psychological, and social consequences, such as contusions, fractures, trauma, motor and neurological damage, or even death [ 2, 3, 4 ]. For this reason, it is crucial the design and deployment of user-friendly technologies to detect falls.

In recent years, solutions such as the Personal Emergency Response System (PERS) have been proposed [ 5 ]. PERS is a manual system whereby a person, after falling to the ground, must press a warning button (usually in a pendant or bracelet), and an emergency team is immediately dispatched to provide assistance. However, this system might not be a good solution in some Elderly Person

Smartwatch

Fall Detection System (FDS)

Sensors Data (Time Signals) 3-Axis Accelerometer 3-Axis Gyroscope

Feature Extraction and Classification

Threshold Approach Machine Learning Approach

Output

Fall Activities of Daily

Life (ADLs)

Emergency Call cases, e.g., if the person has fainted or lost consciousness due to the fall and can not press the emergency button.

To overcome the limitations of PERS, a wide variety of Fall Detection Systems (FDS) have been proposed in the last decade, providing automatic and user-friendly solutions for elderly people [ 3, 6 ]. Most FDS are based on wearable devices [ 7 ], such as belts or bracelets with accelerometer sensors [ 8, 9, 10 ], image-based devices, such as indoor surveillance cameras [ 11, 12 ], or smartphones [ 13, 14, 15 ], among many others.

This paper presents CareFall, an automatic FDS based on wearable devices and Artificial Intelligence (AI) methods. Fig. 1 provides a graphical representation of CareFall. CareFall considers a scenario where the smartwatch is positioned on the wrist, acquiring information related to its inertial sensors, such as the 3-axis accelerometer and gyroscope [ 16, 17 ], or heart rate monitor [ 18, 19 ]. Once the information is acquired by the smartwatch, the time signals (accelerometer and gyroscope signals) are used for feature extraction and classification. Two diferent approaches are considered: i) threshold-based, and ii) machine learning-based. In case the FDS detects a fall, it automatically warns the emergency services.

2. Methods

CareFall considers two of the most popular methods for fall detection in the literature [ 20 ]. They are fed to the 3-axis time signals of accelerometer and gyroscope sensors. The sampling frequency of the smartwatch is between 20-25Hz. For a simple and real-time analysis, we consider separate time windows of 1 minute.

1. Threshold-based: it is one of the simplest and least computationally expensive solutions to detect a fall. It is based on the extraction of additional time signals from the original accelerometer and gyroscope ones such as the Signal Magnitude Vector (SMV), the Fall Index (FI), and the Absolute Vertical Direction (AVD), among others [ 21, 22 ]. After that, a specific threshold is defined for each time sequence. In case the instant value of the time sequence surpasses the threshold, the output of the system would be fall. It is important to highlight that, in case of considering several time signals in the analysis (e.g., SMV, FI, and AVD), the final output of the system would be based on the majority voting of all the time signals considered. 2. Machine Learning-based: this approach automatically learns the discriminative patterns for the task using data. From the original 6 time signals (3-axis accelerometer and gyroscope) and 2 additional time signals (SMV of the accelerometer and gyroscope), we extract the following 11 global features per time window (1 minute) related to statistical information: Mean, Variance, Median, Delta, Standard Deviation, Maximum Value, Minimum Value, 25th Percentile, 75th Percentile, Power Spectral Density (PSD), and Power Spectral Entropy (PSE). In total, we obtain a feature vector with 44 global features related to the accelerometer information and 44 global features related to the gyroscope. Once we have the feature vector with the 88 global features, we train machine learning classifiers for the task of fall detection. The most widely used algorithms are K-Nearest Neighbor (KNN) [ 23 ], Support Vector Machine (SVM) [ 24 ], Gradient Boosting (GB) [ 25 ], Random Forest (RF) [ 25 ], and Artificial Neural Network (ANN) [ 26 ], among others.

3. Experimental Setup

Two popular public databases are considered in the experimental framework of the paper: Erciyes Univesity [ 23 ] and UMAFall [ 27 ]. Table 1 shows the most relevant information from these databases: i) the number of Activities of Daily Life (ADLs) such as walking, sitting, lying down, etc., and simulated falls (forward, backward, sideways, etc.); ii) participant information (number, gender, height, weight, and age range); iii) type of time signals captured (accelerometer and gyroscope); iv) sensor position; and v) the sampling rate. The main criteria for selecting these databases were the position of the sensor (wrist), the sampling rate of the sensors (20-25Hz), and the variability in the type of activities and falls.

Regarding the experimental protocol, both databases are divided into development (80% of participants) and final evaluation (20% remaining participants) datasets. As a result, diferent subjects are considered for the training and final evaluation of CareFall. Regarding metrics, we consider three popular metrics in the literature: Sensitivity (SE), Specificity (SP), and Accuracy. SE refers to the probability of detecting a fall, SP to the probability of detecting a non-fall (i.e., ADLs), and accuracy to the overall system performance.

4. Experimental Results

Table 2 (top) shows the results for the Erciyes University database over the final evaluation set. The results presented correspond to the best configuration of each fall detection approach. The results obtained in general (accuracy) with the threshold-based approach are significantly worse compared with the machine learning approach (77.3% vs. 98.4%), resulting in a higher number of false positives (no falls detected as falls). This trend can be observed by looking at the specificity (68.4% vs. 96.7%). Nevertheless, it is interesting to remark that the Threshold approach outperforms the Machine Learning approach in terms of sensitivity (100% vs. 98.9%), showing to be a simple but eficient approach to detecting falls. In addition, analysing the Machine Learning approach, we can see how the combination of accelerometer information (44 global features) and gyroscope information (44 global features) achieves the best results.

Finally, we can also see in Table 2 (bottom) the results achieved for the public UMAFall database. Similar conclusions are obtained, although better results are achieved on Erciyes database. This can be produced due to the quality of the device and the acquisition process. This seems to indicate that combining accelerometer and gyroscope information is a good practice for the fall detection task.

5. Acknowledgments

This work has been supported by projects: INTER-ACTION (PID2021-126521OBI00 MICINN/FEDER), HumanCAIC (TED2021-131787B-I00 MICINN), and Cartronic Group.

References

[1] W. H. Organization , Falls, 2017 . URL: https://www.who.int/en/news-room/fact-sheets/ detail/falls.

[2] A. F. Ambrose , G. Paul, J. M. Hausdorf , Risk Factors for Falls Among Older Adults: A Review of the Literature , Maturitas 75 ( 2013 ) 51 - 61 . doi:https://doi.org/10.1016/ j.maturitas. 2013 . 02 .009.

[3] Y. S. Delahoz , M. A. Labrador , Survey on Fall Detection and Fall Prevention Using Wearable and External Sensors, Sensors 14 ( 2014 ) 19806 - 19842 . doi:https://doi.org/10.3390/ s141019806.

[4] Z. Zhang , C. Conly , V. Athitsos , A Survey on Vision-Based Fall Detection, in: Proceeding of the 8th ACM International Conference on Pervasive Technologies Related to Assistive Environments (PETRA '15), Association for Computing Machinery , New York, NY, USA, 2015 .

[5] A. Bourke , J. O'Brien , G. Lyons, Evaluation of a Threshold-Based Tri-Axial Accelerometer Fall Detection Algorithm , Gait & Posture 26 ( 2007 ) 194 - 199 . doi:https://doi.org/10. 1016/j.gaitpost. 2006 . 09 .012.

[6] P. Rashidi , A. Mihailidis , A Survey on Ambient-Assisted Living Tools for Older Adults , IEEE Journal of Biomedical and Health Informatics 17 ( 2013 ) 579 - 590 . doi:https://doi. org/10.1109/JBHI. 2012 . 2234129 .

[7] S. Romero-Tapiador , B. Lacruz-Pleguezuelos , R. Tolosana , G. Freixer, R. Daza , C. M. Fernández-Díaz , E. Aguilar-Aguilar , J. Fernández-Cabezas , S. Cruz Gil , S. Molina-Arranz , M. C. Crespo , T. Laguna-Lobo , L. J. Marcos-Zambrano , R. Vera-Rodriguez , J. Fierrez , A . Ramírez de Molina, J. Ortega-Garcia , I. Espinosa-Salinas , A. Morales , E. Carrillo de Santa Pau, AI4FoodDB: A Database for Personalized e-Health Nutrition and Lifestyle through Wearable Devices and Artificial Intelligence , Database: The Journal of Biological Databases and Curation ( 2023 ).

[8] A. Sucerquia , J. D. López , J. F. Vargas-Bonilla , SisFall: A Fall and Movement Dataset, Sensors 17 ( 2017 ) 1 - 14 . doi:https://doi.org/10.3390/s17010198.

[9] L. Martínez-Villaseñor , H. Ponce , J. Brieva , E. Moya-Albor , J. Núñez-Martínez , C. PeñafortAsturiano, UP-Fall Detection Dataset: A Multimodal Approach , Sensors 19 ( 2019 ) 1 - 28 . doi:https://doi.org/10.3390/s19091988.

[10] B. Kwolek , M. Kepski , Human Fall Detection on Embedded Platform Using Depth Maps and Wireless Accelerometer , Computer Methods and Programs in Biomedicine 117 ( 2014 ) 489 - 501 . doi:https://doi.org/10.1016/j.cmpb. 2014 . 09 .005.

[11] Y. M. Galvão , J. Ferreira , V. A. Albuquerque , P. Barros , B. J. Fernandes , A Multimodal Approach Using Deep Learning for Fall Detection , Expert Systems with Applications 168 ( 2021 ) 114226 . doi:https://doi.org/10.1016/j.eswa. 2020 . 114226 .

[12] G. Mastorakis , T. Ellis , D. Makris , Fall Detection Without People: A Simulation Approach Tackling Video Data Scarcity , Expert Systems with Applications 112 ( 2018 ) 125 - 137 . doi:https://doi.org/10.1016/j.eswa. 2018 . 06 .019.

[13] M. A. Guvensan , A. O. Kansiz , N. C. Camgoz , H. I. Turkmen , A. G. Yavuz , M. E. Karsligil , An Energy-Eficient Multi-Tier Architecture for Fall Detection on Smartphones , Sensors 17 ( 2017 ) 1 - 21 . doi:https://doi.org/10.3390/s17071487.

[14] S. Mellone , C. Tacconi , L. Schwickert , J. Klenk , C. Becker , L. Chiari, Smartphone-Based Solutions for Fall Detection and Prevention: the FARSEEING Approach , Zeitschrift für Gerontologie und Geriatrie 45 ( 2012 ) 722 - 727 . doi:https://doi.org/10.1007/ s00391-012-0404-5.

[15] L.-J. Kau , C.-S. Chen , A Smart Phone-Based Pocket Fall Accident Detection , Positioning, and Rescue System, IEEE Journal of Biomedical and Health Informatics 19 ( 2015 ) 44 - 56 . doi:https://doi.org/10.1109/JBHI. 2014 . 2328593 .

[16] P. Delgado-Santos , R. Tolosana , R. Guest , F. Deravi , R. Vera-Rodriguez , Exploring Transformers for Behavioural Biometrics: A Case Study in Gait Recognition, Pattern Recognition ( 2023 ). doi:https://doi.org/10.48550/arXiv.2206.01441.

[17] P. Delgado-Santos , R. Tolosana , R. Guest , R. Vera-Rodriguez , F. Deravi , A . Morales, GaitPrivacyON: Privacy-Preserving Mobile Gait Biometrics using Unsupervised Learning , Pattern Recognition Letters 161 ( 2022 ) 30 - 37 . doi:https://doi.org/10.1016/j.patrec. 2022 . 07 .015.

[18] J. Hernandez-Ortega , R. Daza , A. Morales , J. Fierrez , R. Tolosana , Heart Rate Estimation from Face Videos for Student Assessment: Experiments on edBB , in: Proceedings of the IEEE 44th Annual Computers, Software, and Applications Conference (COMPSAC) , IEEE Computer Society , Los Alamitos, CA, USA, 2020 , pp. 172 - 177 .

[19] J. Hernandez-Ortega , R. Tolosana , J. Fierrez , A . Morales, DeepFakesON-Phys: DeepFakes Detection based on Heart Rate Estimation , in: Proceedings of the 35th AAAI Conference on Artificial Intelligence Workshops (AAAIw) , Vancouver, Canada, 2021 .

[20] L. Schwickert , C. Becker , U. Lindemann , C. Maréchal , A. Bourke , L. Chiari , J. L. Helbostad , W. Zijlstra , K. Aminian , C. Todd , S. Bandinelli , J. Klenk, for the FARSEEING Consortium, the FARSEEING Meta Database Consensus Group, Fall Detection with Body-Worn Sensors , Zeitschrift für Gerontologie und Geriatrie 46 ( 2013 ) 706 - 719 . doi:https://doi.org/10. 1007/s00391-013-0559-8.

[21] E. Casilari , M. A. Oviedo-Jiménez , Automatic Fall Detection System Based on the Combined Use of a Smartphone and a Smartwatch , PLOS ONE 10 ( 2015 ) 1 - 11 . doi:https://doi. org/10.1371/journal.pone. 0140929 .

[22] E. Casilari , J.-A. Santoyo-Ramón , J.-M. Cano-García , Analysis of Public Datasets for Wearable Fall Detection Systems, Sensors 17 ( 2017 ). doi:https://doi.org/10.3390/ s17071513.

[23] A. T. Özdemir , B. Barshan , Detecting Falls with Wearable Sensors Using Machine Learning Techniques , Sensors 14 ( 2014 ) 10691 - 10708 . doi:https://doi.org/10.3390/ s140610691.

[24] A. Dinh , D. Teng , L. Chen , Y. Shi , C. McCrosky , J. Basran , V. D. Bello-Hass , Implementation of a Physical Activity Monitoring System for the Elderly People with Built-in Vital Sign and Fall Detection , in: Proceeding of the 6th International Conference on Information Technology: New Generations , Las Vegas, NV , USA, 2009 , pp. 1226 - 1231 .

[25] N. Zurbuchen , A. Wilde , P. Bruegger, A Machine Learning Multi-Class Approach for Fall Detection Systems Based on Wearable Sensors with a Study on Sampling Rates Selection , Sensors 21 ( 2021 ). doi:https://doi.org/10.3390/s21030938.

[26] S. Abbate , M. Avvenuti , F. Bonatesta , G. Cola, P. Corsini , A. Vecchio , A Smartphone-Based Fall Detection System , Pervasive and Mobile Computing 8 ( 2012 ) 883 - 899 . doi:https: //doi.org/10.1016/j.pmcj. 2012 . 08 .003.

[27] E. Casilari , J. A. Santoyo-Ramón , J. M. Cano-García , UMAFall: A Multisensor Dataset for the Research on Automatic Fall Detection , Procedia Computer Science 110 ( 2017 ) 32 - 39 . doi:https://doi.org/10.1016/j.procs. 2017 . 06 .110.