A robust deep learning model for fall action detection using healthcare wearable sensors

Health routines via machine learning

Many researchers have extensively studied machine learning-based methods for improving health routines. Researchers from Arvindhan, Sharma & Kumar (2023) developed a sensor-powered machine learning system that examines human health activities using Accelerometer and Gyro to interpret body motions. Multiple tracking algorithms like Random Forest, K-Nearest Neighbor (KNN), Support Vector Machine (SVM), and Logistic Regression function within the system to analyze the walking sleeping, and running of users. The presented graphical results offer beneficial knowledge points for doctors and dietitians to develop practical health treatment strategies. Bar charts and pie charts in the user interface help people gain a better perception of their routines. Hoffmann, Steinhage & Lauterbach (2017) built a technical system using machine learning detection methods to monitor unsolicited health routine changes across extended time intervals. The system uses sensors with a unique capacitance detection system to track human behaviors including sleep behavior and walking methods. A processing scheme obtains behavioral information from sensor data for analysis through a learning algorithm. The algorithm finds irregular conduct in everyday practices to generate early detection alerts about health alterations while needing no input from patients. A healthcare bot developed by Shankhdhar et al. (2021) uses machine learning to detect patient symptoms which results in specific health routine suggestions. The system uses user input analysis to create personalized diet plans and workouts while matching patients with medical professionals. The system uses the Decision Tree algorithm to detect diseases from patient symptoms which improves diagnostic outcomes. The system develops patient interaction through Python libraries that generate a simulated communication environment to boost user activity and healthcare system management. Toapanta et al. (2005) examined whether machine learning techniques could enhance physical exercise outcome predictions for hypertensive patients. Naive Bayes along with Decision Tree and Logistic Regression participated in modeling the prescribed exercises’ effects according to their research. Analysis using the Decision Tree algorithm produced the most optimal results as it reached an accuracy rate of 79% from the data. The predictive model functions as a part of the Food Corporation of India (FCI) Research Project platform to help health practitioners design customized workout plans. Performance measures including accuracy precision and recall allow the system to measure model effectiveness toward enhancing health results for hypertensive people. Machine learning systems based on weak supervision methods served as the focus of research by Banovic et al. (2017) to create behavioral models that detect actions that affect well-being. They have designed their method to use weak labeling from demonstrated routines which minimizes the need for extensive manual data labeling. The system detects aggressive driving behavior through its identification process and then creates opportunities for patients to reflect on their actions while correcting their behavior. Through machine learning approaches researchers can develop practical observations which support health-promoting actions and general wellness enhancement.

Health routine via deep learning

Healthcare received major enhancements through deep learning methods for monitoring and management as these learning approaches generate exact outputs by using flexible process models. User movement observation and exercise statistics collection occurred in a web application built with React through webcam analysis powered by TensorFlow.js according to Salian & Naik (2023). Users achieve exact motion tracking for fitness judgment through webcam-based monitoring provided by the system. Activity monitoring systems utilizing Node.js technology and MongoDB storage systems achieve revolutionary capability due to their ability to provide real-time user experiences. Zargar, Zargar & Mehri (2023) demonstrated Deep Neural Network (DNN) research based on Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) to improve healthcare data prediction accuracy. The system provides medical image processing through Natural Language Processing (NLP) technology and then merges CNNs and RNNs to achieve top speed for clinical data interpretation. The multiple layers within deep learning systems help represent data better and improve generalization to achieve superior predictions for diverse clinical input information. Nisar et al. (2021) joined other researchers to conduct research focusing on deep learning techniques for healthcare applications in their investigation. Medical imaging with supervised CNNs and Deep CNNs segments brain structures during supervised operations but relies on Boltzmann machines and autoencoders to accomplish successful unstructured data processing in unsupervised stages. Healthcare professionals improve their diagnoses by employing these methods to build intelligent AI diagnostic tools that lower healthcare costs at the same time they develop disease identification algorithms. The article examines two key sections: the data management barriers posed by unstructured information first followed by proof of concept research into optimized unsupervised learning framework. The research team led by Kaul, Raju & Tripathy (2022) demonstrated how deep neural networks can be utilized for disease diagnosis and disease treatment methods for cancer, diabetes, and Alzheimer’s. These processing methods help researchers study disease development patterns and create unique treatment programs for patients as well as improve medical care. Deep learning functions through brain-like processing that leads to high efficiency when analyzing medical data thus it stands as a powerful instrument for monitoring complex medical diseases. The system by Fan (2019) creates a deep learning platform for health recommendations that gathers data from wearable devices before investigation. The data enters a deep learning network that generates personalized food and movement plan indexes through its processing. Data mining of large datasets allows the system to create customized recommendations that generate health management solutions from individual user data. Research by Ventrilho dos Santos & Carvalho (2016) demonstrated that clinical diagnosis accuracy highly depends on deep learning models especially when utilizing CNNs with back-propagation algorithms. Such models demonstrate superior effectiveness compared to traditional approaches as well as human interpretation in analyzing medical images and disease recognition tasks. Deep learning exemplifies its ability to enhance medical diagnosis precision through accurate clinical assessments and assists doctors with exact treatment plan generation according to individual patient information. The authors in Shekhar et al. (2024) developed deep learning models including CNN, CNN-LSTM, and Multi-Headed CNN-Bidirectional Long Short-Term Memory (BiLSTM) to recognize human activities through wearable sensors. According to their study which used the UCI Human Activity Recognition (HAR) dataset the Multi-Headed CNN-BiLSTM model achieved superior performance at all levels of accuracy and precision evaluation along with recall and F1-scores. Advanced deep learning architectures demonstrate their capabilities for healthcare smart applications by providing outstanding tracking of physical activities. The summary of advantages and limitations of existing approaches is displayed in Table 1.

Pre-processing

A sixth-order Butterworth low-pass filter executed preprocessing operations on inertial sensor data as a step for noise elimination. A Butterworth filter enables flat passband frequency response because it works well for systems demanding smooth response characteristics. The filter is defined by its order $n$ and cutoff frequency $f_c$ with the transfer function given by:

(1) $$J\left(x\right)={\displaystyle \frac{1}{\sqrt{1+{\left({\displaystyle \frac{x}{w_c}}\right)}^{2n}}}}$$where x is the complex frequency variable, $w_c$ is the cutoff angular frequency, and n is the filter order. In this study, a sixth order butterworth filter (n = 6) was used, with a cutoff frequency $f_c=0.01$ Hz. Given a sample rate of $f_x=2.4$ Hz, the Nyquist frequency was calculated as $f_N={\displaystyle \frac{f_x}{2}=1.2Hz}$, resulting in a normal cutoff frequency of ${\displaystyle \frac{f_c}{f_n}=0.0083}$. This configuration effectively reduced high-frequency noise while preserving the underlying signal characteristics. The results of raw vs filtered data over both UP-fall and UR-fall datasets are shown in Figs. 2A and 2B.

Windowing

The processed IMU sensor data received division into smaller parts using windowing approaches before additional examination. The research applications accessed these merged windows which were created after their overlapping process finished. Application of the Hamming window served to decrease spectral leakage effects while keeping data continuity throughout each divided section. The definition of the Hamming window exists as follows in mathematical format:

(2) $$\mathrm{w}\left(\mathrm{i}\right)=0.54-0.46\cos \left({\displaystyle \frac{2\pi i}{M-1}}\right)$$where $i$ ranges from 0 to M-1, and M is the total number of samples in the window. A window size of 10 s (M = 1,000, assuming a sample rate of 100 Hz) was chosen as shown in Fig. 3. Hamming window filters performed sensor value transition smoothing during the segmentation process at the start and end of each window. The process of semantic segmentation was finalized by placing stacked windows with overlap to produce one seamless dataset for subsequent evaluation.

Feature extraction

The proposed system obtains several essential features from inertial data following data windowing. SSCE coefficients along with MFCC and LPCC and Parseval’s energy and AR coefficients form the set of extracted features. The upcoming subsections present an overview of the extraction process and evaluation outcomes for these features. The comparative analysis of techniques is shown in Table 2.

State space correlation entropy

By applying SSCE, researchers examined both the time-related structure and signal complexity to extract motion-related features. The analysis procedure divides signals into overlapping embedded vectors through a predetermined time delay system. The distance matrix requires two sets of pairwise vector comparisons which produce self-correlation information on the diagonal and vector distance readings between off-diagonal elements. The derived covariance matrix consists of distances that determines correlation probabilities leading to the computation of SSCE according to the following:

(3) $$\mathrm{S}\mathrm{S}\mathrm{C}\mathrm{E}=-\log \left({\displaystyle \frac{1}{N^2}\sum _{K=1}^N\sum _{L=1}^N\mathrm{\Psi }\left(\mathrm{\delta }-\left|\left|{\mathrm{Z}}_{\mathrm{k}}-Z_l\right|\right|\right)}\right)$$where, $Z_k$ and $Z_l$ are the embedded vectors, $\left|\left|Z_k-Z_l\right|\right|$ represents the Euclidean distance between them, $\delta$ is small threshold, N is the total number of embedded vectors, and $\mathrm{\Psi }$ (x) is the Heaviside function. The computation of SSCE values for every data window allowed researchers to establish an average measure that describes motion signal complexity. The data points were organized according to activity categories and bar charts showed the SSCE values distribution throughout different activities of the dataset. This method extracts dynamic features from signals that lead to reliable insights about separating different activity types. Figure 4 represents results obtained from extracting SSCE from both datasets.

Mel frequency cepstral coefficients

The MFCC feature analyzes the spectral bands of inertial sensor data to extract spectral characteristics from the signal effectively. The information extraction occurs through peak identification of periodic components which transforms the signals into data that exists between time and frequency domains. The movement patterns get robust representation from MFCCs because they use sensor data to explain recorded signal changes. The MFCC feature is calculated using a discrete cosine transform, where the log filter bank amplitude $g_m$ is derived, and the total number of filter bank channels Q determines the final feature, as shown in the following equation.

(4) $$\mathrm{M}\mathrm{F}\mathrm{C}\mathrm{C}\mathrm{f}\mathrm{e}\mathrm{a}\mathrm{t}\mathrm{u}\mathrm{r}\mathrm{e}\mathrm{s}=\sum _{m=0}^{Q-1}{g}_m\cdot cos\left[{\displaystyle \frac{\pi }{Q}\left(m+0.5\right)r}\right].$$

Here $g_m$ shows the log filter bank amplitude, Q is the total number of filter bank channels, and r is the coefficient index. Due to their ability to detect frequency alterations in inertial data the MFCCs produce more precise activity pattern recognition. The Fig. 5 shows a visual display of MFCCs which corresponds to the Ankle Angular Velocity data within the sensor measurements.

Linear predictive cepstral coefficients

The inertial signals produce LPCC features through a combination of signal transfer functions and first derivative analyses of specific frequency bands. The calculated coefficients describe signal movement patterns through sequential computations based on linear prediction coefficients b(m). Different equations determine the computation of LPCC coefficients:

(5) $$d_k=b_k+\sum _{i=1}^{k-1}\left(\begin{array}{c}k\\ i\end{array}\right)d_i{b}_{k-i,}1\le k\le p$$

(6) $$d_k=b_k+\sum _{i=1}^{p-1}\left(\begin{array}{c}k\\ i\end{array}\right)d_i{b}_{k-i,}p\le k\le q.$$

Here, $d_k$ represents the LPCC coefficient. $b_k$ is the linear prediction coefficient, p is the prediction order. q is the total number of coefficients, and $(\genfrac{}{}{0ex}{}{k}{i})$ is the binomial coefficient. The cepstral coefficients are calculated per frame in each equation which results in powerful signal dynamics representations across time and frequency domains. LPC coefficients were computed with order 12 for the inertial signals from all their measurement axes in this research. Visual information about coefficient temporal dynamics appeared in time-series plots for features produced through the LPCC algorithm across the X, Y, and Z axes. The visual representations of LPCC data show their effectiveness in detecting movement patterns as they track signal variations continuously across various activities. An output presented in Fig. 6 represents the data extraction results from the dataset after performing LPCC processing.

Parseval’s energy

We determined the signal power strength through Parseval’s energy applied to accelerometer data windows during different activities. The calculated energy using Parseval’s theorem provides a precise measurement of the signal’s total power which exists in both the time domain and frequency domain. According to Parseval’s theorem, the signal energy measured in the time domain equals its energy values in the frequency domain. The formulation of time domain energy matches the definition of frequency domain energy as follows:

(7) $$E_{TimeDomain}\left(x\left(t\right)\right)={\int }_{-\mathrm{\infty }}^{\mathrm{\infty }}{\left|x\left(t\right)\right|}^{2}dt$$

(8) $$E_{FreqDomain}\left(X\left(j\omega \right)\right)={\displaystyle \frac{1}{2\pi }{\int }_{-\mathrm{\infty }}^{\mathrm{\infty }}{\left|\left(X\left(j\omega \right)\right)\right|}^{2}d\omega .}$$

Parseval’s Theorem states that:

(9) $$E_{TimeDomain}=E_{freqDomain}$$where x(t) is the signal in the time domain. $X(j\omega )$ is its counterpart in the frequency domain, t represents time and $\omega$ shows the angular frequency. Parseval energy was then calculated for each window by summing the squared signal values within the window:

(10) $$E_{window}=\sum _{k=1}^N{\left|yk\right|}^2$$where N is the number of samples in the window, and yk is the signal value at sample k. The calculated energies were normalized for visualization purposes, a subset of human activities has been plotted, as illustrated in Fig. 7A. The figure showcases the normalized Parseval’s energy distribution for selected actions, including standing, sitting, laying, and picking up an object over UP-fall dataset. The variations in energy levels across these activities highlight the distinctive characteristics of movement patterns, reinforcing the applicability of Parseval’s energy as a discriminative feature in activity recognition systems.

Auto regression coefficients

The Auto-Regressive (AR) feature functions as a basic technical method that uses temporal relationships to examine repetitive motion patterns from sensor data. The model expresses each signal value through past observations weighted by autoregressive coefficients together with a dynamic error that adjusts to signal variation. A generalized AR model with an adaptive error component appears as follows:

(11) $$AR\left(feature\right)=\sum _{i=1}^N{a}_i{f}_{n-i}+\sum _{j=1}^M{b}_j{ϵ}_{n-j}+\lambda {\int }_0^T{e}^{-\beta \left(T-\tau \right)}f\left(\tau \right)d\tau$$where $f_{n-i}$ represents past signal values, $a_i$ are autoregressive coefficients capturing temporal dependencies, ${ϵ}_{n-j}$ is a noise component modeled as a moving average, $b_j$ are coefficients controlling the influence of past error terms, $\beta$ introduces a continuous memory effect through an exponential decay function, and $\lambda$ is a weighting factor adjusting the contribution of historical patterns. Figure 8 displays the Auto regression plot for all the datasets.

Particle swarm optimization

The population-based optimization algorithm Particle Swarm Optimization (PSO) derives its methods from active swarm psychology which mimics bird flocking and fish schooling behaviors (Batool, Jalal & Kim, 2019). This research uses PSO to optimize the selection of SSCE coefficient features combined with Parseval’s energy features as well as MFCC features LPCC features and AR coefficient features. Every swarm particle updates its position through its individual best experience together with the best positions located by the neighboring swarm members. The optimization process searches for minimum variance in weighted feature combinations that fulfill the requirement of total weight equality to one. This relationship appears as:

(12) $${min}_{w}Var\left(\sum _{i=1}^N{w}_i{x}_i\right)subjectto\sum _{i=1}^N{w}_i=1$$where $w_i$ are the feature weights and $x_i$ are the extracted features. Through repeated updates of its weights, the PSO algorithm identifies an optimal data representation that enhances the performance of human fall action detection. The results of optimization over UP-fall dataset are shown in Fig. 9.

Fall action classification

An LSTM network (Waheed et al., 2021) serves as the choice for fall action classification. LSTM represents a specialized RNN that excels at processing sequential data through its ability to detect longer-term dependencies thus making it appropriate for time-series classification tasks. The research uses UP-Fall and UR-fall datasets as input sources to extract SSCE coefficients together with Parseval’s energy and MFCC LPCC and AR coefficients from sensor signals. The classification performance receives enhancement through the use of Particle Swarm Optimization (PSO) for selecting essential features that promote discrimination. The model includes two stacked LSTM layers containing 64 units and 32 units which follow dropout regularization on fully connected layers to avoid overfitting. The final classification layer applies softmax activation to execute multiclass prediction because it works for the multiple action category prediction task. Standard scaling methods normalize the feature values while categorical results get encoded for preprocessing of both datasets. Training takes place with the Adam optimizer together with the categorical cross-entropy loss that follows this definition:

(13) $$L=-{\displaystyle \frac{1}{n}\sum _{i=1}^n\sum _{j=1}^C{y}_{ij}\log \left({\hat{y}}_{ij}\right)}$$where $y_{ij}$ is the true label for class j, and ${\hat{y}}_{ij}$ denotes the predicted probability for that class. The research uses n-fold cross-validation instead of a train-test split for robust evaluation to test the model across multiple partitions of data.

UP-fall dataset

The UP-fall detection dataset provides comprehensive data about identifying eleven fall-related activities (Martínez-Villaseñor et al., 2019) based on three trial executions from each subject. Healthy young people without disabilities participated in the dataset collection by performing eleven fall-related activities through wearable sensors and vision equipment with ambient sensors. The UP-fall detection dataset incorporates postures from six activities of daily living (ADL) which are standing (STD), sitting (STG), walking (WK), lying down (LA), jumping (JU), and picking up objects (PO) alongside five fall-related activities that include falling forward with hands (FH), falling forward with knees (FK), falling backward (FBK), falling sideways (FSD), and sitting on an empty chair (FEC) actions. The fall activities span 10 s yet the ADLs need 60 s of completion except for the jumping and object pickup which run for 30 s and 10 s, respectively.

UR-fall dataset

The UR fall detection dataset was created from data obtained through a combination of two USB-connected Kinect cameras and an IMU device which operated at the waist via Bluetooth connectivity. The ADL events were recorded using only Camera 0 as well as an accelerometer. The sensor data acquisition involved PS Move hardware and the x-IMU system. The study was conducted with five volunteers who performed 70 sequences among 30 fall-induced sequences and 40 sequences of ADL activities within an office setting. Each participant executed multiple forward (FF) and backward (FB) along with lateral falls onto the 2 cm thick carpet space with intended motions at least three times. The x-IMU sensor was operated by the researchers near the pelvic region. Participants performed the ADLs by standing (ST), sitting (SG), squatted, and bending (BG) their bodies after which they picked up objects and rested on the settee. Falls detected by the system demonstrated perfect accuracy but the sensor misidentified quick sitting activities (SG) because it confused them with true falls without additional sensor data. The provided dataset comprises falls that happen when someone is standing (FST) along with falls that take place during the sitting (FS) phase. Each dataset entry includes original accelerometer measurements that accompany depth images from two Kinect cameras as well as RGB image time sequences. A threshold-based fall detection technique served as one of the methods described by the researchers.

HealthLINK falls dataset

The HealthLINK Falls Dataset (Paolini et al., 2019) was gathered at the San Diego State University (SDSU) Neuromechanics and Neuroplasticity Laboratory utilizing sixteen Noraxon Inertial Measurement Unit (IMU) sensors positioned at designated anatomical sites on participating graduate students and faculty members. Every participant donned a virtual reality headset and traversed a linear pathway within a regulated laboratory setting. Falls were externally triggered by the sudden removal of a cloth sheet positioned on the floor in front of a mattress, resulting in fall, stumble, trip, or no-response incidents. The IMUs captured 183 kinematic parameters, encompassing joint orientation angles and limb accelerations, at a sampling frequency of 200 Hz. Subsequent to each trial, the recorded data were analyzed utilizing the Noraxon myoRESEARCH (MR3) program to categorize scenarios as either fall (1) or no fall (0), hence producing a binary classification dataset. This dataset encompasses both induced falls and normal ambulation situations, serving as a valuable resource for the development and assessment of machine learning-based fall detection systems.

Conclusion and future work

This work illustrates that the use of sophisticated preprocessing, varied feature extraction methods, and enhanced sequence modeling can significantly elevate the precision and dependability of fall detection using wearable sensors. The integration of time-domain features (e.g., AR coefficients) and frequency-domain features (e.g., MFCC, LPCC) offered complementary insights into motion patterns, while the PSO-driven feature selection markedly diminished redundancy, allowing the LSTM classifier to concentrate on the most distinguishing information. The outcomes on the UP-Fall and UR-Fall datasets demonstrated consistent improvements above baseline methods, validating the efficacy of this design decision. Nevertheless, the research underscored significant problems, including as sensitivity to sensor positioning, inadequate capture of real-world variability in benchmark datasets, and diminished generalization to novel activity categories. These observations highlight the necessity for adaptive learning methodologies, multimodal data integration, and extensive field testing to reconcile laboratory performance with real-world application in healthcare monitoring systems.

In future research, we will enhance the interpretability of the proposed fall detection system by incorporating explainable AI techniques such as Local Interpretable Model-agnostic Explanations (LIME) and SHapley Additive exPlanations (SHAP). These methodologies will clarify the influence of specific attributes and model decisions, similar to how local explanations have been used in intelligent healthcare systems for heart attack risk assessment to justify urgent treatment recommendations. Inspired by applications reported in “Enhancing Transparency in Smart Farming: Local Explanations for Crop Recommendations Using LIME” and “Role of Explainable AI in Crop Recommendation Technique of Smart Farming”, this integration will transform our LSTM-based framework from a non-transparent entity into a clear and trustworthy decision-support tool for healthcare environments.

We also aim to integrate self-attention mechanisms into the LSTM-based architecture to enable the model to focus on critical temporal segments and key feature interactions within inertial sensor data. This is expected to improve accuracy, robustness, and adaptability, leveraging the proven effectiveness of attention-based models in related domains. Further work will explore multimodal sensor fusion by combining inertial data with optical or depth sensing to improve detection reliability in varied conditions, as well as adaptive and online learning strategies to allow real-time parameter updates in response to changes in sensor placement, user movement patterns, or environmental factors. Additionally, we will focus on reducing model complexity for deployment on low-power edge devices, ensuring applicability in wearable and IoT-based healthcare systems. Comprehensive field testing with diverse populations, including older adults and individuals with mobility impairments, will be conducted to evaluate robustness, usability, and user acceptance in real-world scenarios.