1. Introduction

2. Problem Statement

3. Dataset

4. Methodology

Figure 1. Overview of our proposed pipeline for fall detection

Figure 1. Overview of our proposed pipeline for fall detection

5. Experiments

5.1. Baselines

This work investigated the performance and efficiency of various ML models, including both traditional and deep models. In this section, we delve deeper into the details of these models.

5.1.1. Deep Models

In this study, we investigated the performance of deep Models by implementing three well-known architectures. By deploying these architectures, the study aimed to comprehensively assess the efficacy of standard deep learning models in the context of fall detection:

• 1) MLP: We built a fully connected network using a combination of dense layers to classify whether a fall occurred in incoming time windows or not.
• 2) CNN + MLP: We implemented a two-segment architecture within this architecture. In the first segment, a combination of convolution and dropout layers extract efficient features from the input time window. In the second segment, we exploited fully connected layers for classification.
• 3) LSTM + MLP: Since working on time-series data, we also investigated the efficiency of utilizing LSTM layers for feature extraction. We leveraged the output of LSTM models as an input for a fully connected network, which classifies the incoming time windows.

5.1.2. Traditional Models

While implementing traditional ML models, a crucial pre-processing step involves reshaping the dimension of samples within the training and evaluation datasets, as they are initially incompatible with 3D inputs. We transform the sample dimension from $(n,k,d)$ to $(n,k\times d)$ to achieve this. Following this conversion, the shapes of the training and evaluation datasets become $(61036,1800)$ and $(16375,1800)$, respectively. Additionally, recognizing the inherent challenges posed by high-dimensional data for many traditional machine learning models, we address this concern by reducing the datasets’ dimension to 64. This reduction is accomplished by employing a decision tree classifier as a feature selector, ensuring the retention of the most salient features essential for model training and evaluation. After completing these processing steps, we trained the following classifiers and evaluated their performance in fall detection.

5.3. Results

The comprehensive assessment of ML models on the SisFall dataset is presented herein, with a detailed breakdown of results in Table 1, differentiating between fall activities and ADLs. Examination of Table 1 reveals that the MLP and CNN + MLP architectures exhibit superior performance compared to all other baseline models. Key insights derived from the findings are as follows: (1) While Random Forest emerges as the top performer for fall detection among traditional ML models, its efficiency lags behind that of deep models. Notably, while optimal within the traditional ML paradigm, ensemble models confront challenges when confronted with high-dimensional data. (2) Given the proximate results of deep models, particularly the MLP and CNN + MLP architectures, it can be inferred that a proficient representation of the features matrix can be attained with MLP networks, obviating the necessity for intricate structures. (3) In summation, the deep models exhibit commendable performance in fall detection, accurately discerning the temporal windows corresponding to the occurrence of falls.

Table 1. Performance of standard ML models in fall detection

Table 1. Performance of standard ML models in fall detection

ML models	$\mathit{P}{\mathit{r}}_{\mathit{F}\mathit{a}\mathit{l}\mathit{l}}$	$\mathit{P}{\mathit{r}}_{\mathit{A}\mathit{D}\mathit{L}}$	$\mathit{R}{\mathit{e}}_{\mathit{F}\mathit{a}\mathit{l}\mathit{l}}$	$\mathit{R}{\mathit{e}}_{\mathit{A}\mathit{D}\mathit{L}}$	$\mathit{F}{1}_{\mathit{F}\mathit{a}\mathit{l}\mathit{l}}$	$\mathit{F}{1}_{\mathit{A}\mathit{D}\mathit{L}}$
Logistic Regression	37.19%	99.16%	82.67%	93.67%	51.31%	96.34%
Decision Tree	55.79%	97.98%	55.63%	98.00%	55.71%	97.99%
Random Forest	91.37%	98.13%	58.16%	99.75%	71.08%	98.93%
KNN	95.62%	97.75%	49.29%	99.89%	65.05%	98.81%
MLP	66.59%	99.44%	87.88%	98.00%	75.77%	98.71%
CNN + MLP	75.21%	98.85%	74.78%	98.88%	75.00%	98.87%
LSTM + MLP	56.45%	99.10%	80.70%	97.17%	66.43%	98.13%

6. Conclusion

References

1. Lockhart, T.E. An integrated approach towards identifying age-related mechanisms of slip initiated falls. Journal of Electromyography and Kinesiology 2008. [Google Scholar] [CrossRef] [PubMed]
2. Organization, W.H. World Report on Ageing and Health; World Health Organization, 2015. [Google Scholar]
3. Fleming, J.; Brayne, C.; collaboration, C. Inability to get up after falling, subsequent time on floor, and summoning help: prospective cohort study in people over 90. BMJ (Clinical research ed.) 2008. [Google Scholar] [CrossRef] [PubMed]
4. Huang, C.N.; Chiang, C.Y.; Chen, G.C.; Hsu, S.; Chu, W.C.; Chan, C.T. Fall Detection System for Healthcare Quality Improvement in Residential Care Facilities. Journal of Medical and Biological Engineering 2010. [Google Scholar] [CrossRef]
5. Alshammari, S.A.; Alhassan, A.M.; Aldawsari, M.A.; Bazuhair, F.O.; Alotaibi, F.K.; Aldakhil, A.A.; Abdulfattah, F.W. Falls among elderly and its relation with their health problems and surrounding environmental factors in Riyadh. Journal of family and community medicine 2018. [Google Scholar] [CrossRef] [PubMed]
6. Palestra, G.; Rebiai, M.; Courtial, E.; Giokas, K.; Koutsouris, D. A Fall Prevention System for the Elderly: Preliminary Results. IEEE 30th International Symposium on Computer-Based Medical Systems (CBMS), 2017.
7. Stevens, J.A.; Corso, P.S.; Finkelstein, E.A.; Miller, T.R. The costs of fatal and non-fatal falls among older adults. Injury Prevention 2006. [Google Scholar] [CrossRef] [PubMed]
8. Tongskulroongruang, T.; Wiphunawat, P.; Jutharee, W.; Kaewmahanin, W.; Rassameecharoenchai, T.; Jennawasin, T.; Kaewkamnerdpong, B. Comparative Study on Fall Detection using Machine Learning Approaches. 19th International Conference on Electrical Engineering/Electronics, Computer, Telecommunications and Information Technology (ECTI-CON), 2022.
9. Pandya, B.; Pourabdollah, A.; Lotfi, A. Comparative Analysis of Real-Time Fall Detection Using Fuzzy Logic Web Services and Machine Learning. Technologies 2020. [Google Scholar] [CrossRef]
10. Sucerquia, A.; López, J.D.; Vargas-Bonilla, J.F. SisFall: A fall and movement dataset. Sensors 2017. [Google Scholar] [CrossRef] [PubMed]
11. Casilari-Pérez, E.; Santoyo-Ramón, J.A.; Cano-García, J.M. UMAFall: A Multisensor Dataset for the Research on Automatic Fall Detection; FNC/MobiSPC, 2017. [Google Scholar]
12. Vavoulas, G.; Pediaditis, M.; Chatzaki, C.; Spanakis, E.G.; Tsiknakis, M. The MobiFall Dataset: Fall Detection and Classification with a Smartphone. International Journal of Monitoring and Surveillance Technologies Research 2014. [Google Scholar] [CrossRef]
13. Zeng, F.; Gao, W. Early Rumor Detection Using Neural Hawkes Process with a New Benchmark Dataset. In Proceedings of the North American Chapter of the Association for Computational Linguistics; 2022. [Google Scholar]
14. Dablain, D.; Krawczyk, B.; Chawla, N.V. DeepSMOTE: Fusing Deep Learning and SMOTE for Imbalanced Data. IEEE Transactions on Neural Networks and Learning Systems 2023. [Google Scholar] [CrossRef] [PubMed]