A temporal convolutional network approach for gait-based fall risk prediction to support safe sport participation in older adults

Abstract

The risk of falling represents a significant barrier preventing many older adults from engaging in mass sports and physical activity. Objective assessment of the fall risk using wearable technologies constitutes an essential support in safe sport participation through early detection of gait instability. This study proposes a hybrid Temporal Convolutional Network (TCN)-based framework for gait-based fall risk identification based on lower back-mounted inertial measurement unit (IMU) sensor data acquired during one-minute laboratory tests and three-day free-living recording. The proposed methodology integrates data-driven temporal modeling of raw inertial signals with clinically interpretable handcrafted gait features. Temporal dependencies are modeled using dilated convolutions, while subject-level predictions are obtained through statistical aggregation of window-level representations. The framework is evaluated using subject-wise cross-validation and demonstrates consistent discrimination between fallers and non-fallers.

The proposed methodology:

● Processes waveform-level gait dynamics and captures detailed gait dynamics from raw accelerometer signals from short accelerometer recordings using the TCN.

● Aggregates window-level embeddings at the subject level using statistical descriptors.

● Incorporates clinically interpretable gait features that describe spatiotemporal characteristics, inter-axis coordination, and time- and frequency-domain properties of walking.

Graphical abstract

Specifications table

Subject area	Mathematics and Statistics
More specific subject area	Hybrid Deep Learning-based Gait Analysis for Fall Risk Prediction
Name of your method	Hybrid Temporal Convolutional Network–Based Gait Analysis Pipeline for Fall Risk Prediction
Name and reference of the original method	Temporal Convolutional Networks (TCNs) Bai S., Kolter J.Z., Koltun V., An empirical evaluation of generic convolutional and recurrent networks for sequence modeling, arXiv:1803.01271, 2018.
Resource availability	Dataset: https://www.physionet.org/content/ltmm/1.0.0/

Background

Age-related physiological changes, including reduced muscle mass, joint stiffness, decreased mobility and fatigue response, present a barrier to active sports participation for middle-aged and elderly people. Health-related factors have been consistently reported as the major contributors to limited engagement in physical activity among this age population, bringing the need for continuous monitoring approaches [1,2]. Moreover, research suggests that the elderly experience the fear of falling, which is strongly associated with avoidance behavior and ultimately leads to decreased physical activity levels [3]. This consequently results in muscle weakness and balance impairments.

Wearable technology sensors provide strong support for promoting personalized physical activity by enabling continuous, objective assessment of movement patterns. In addition to monitoring physical activity levels and intensity, they enable objective quantification of gait and locomotor behavior during both sports and daily activities. There has been a growing interest in using machine learning-based approaches for wearable smart sensors for personalized healthcare [4,5]. Prior studies have employed machine learning models for the gait symmetry evaluation [6], gait-based biometric authentication [7], and activity recognition [[8], [9], [10]].

Furthermore, the application of machine learning techniques to gait analysis provides non-invasive risk assessment, serving not only as a diagnostic label but as a tool for injury prevention [[11], [12], [13]]. Early assessment of the sport-related risks may support in creating a personalized activity plan that matches the individual’s functional capacity and addresses the specific deficits in the individual’s routine (mobility, stability, gait symmetry, etc.). Essentially, wearable sensors can effectively provide quantitative data for fall risk assessment. A single-activity assessment framework was previously proposed in which 168 gait features, derived from both time- and frequency-domain analyses, were extracted from a one-minute walking trial [14]. In that work, a ReliefF-based feature selection method was combined with a multilayer perceptron classifier, yielding a reported classification accuracy of 82.2 % for fall risk prediction [14]. Fall risk prediction among older adults was further investigated through the analysis of spatiotemporal gait parameters and gait variability [15]. The results indicated that step width variability, particularly under increased gait speed conditions, exhibited strong discriminative capability.

Collectively, these approaches have demonstrated a strong potential for scalable, sensor-based fall-risk screening. Therefore, this paper aims to show a reproducible methodology for the ML-based gait analysis to predict falls in older adults.

Method details

Dataset

The proposed methodology utilizes an open-source PhysioNet Long-Term Movement Monitoring (LTMM) dataset, which comprises continuous 3-day recordings and standardized 1-minute laboratory walking trials of 71 elder community residents [16,17]. The health selection criteria included no prior diagnosis of gait or balance disorders and cognitive function within the normal limits, as assessed by standard screening tools (e.g., MMSE > 24). The inertial sensor data were collected through the 3D accelerometer mounted on the lower back (at the L5 level) of participants during both the 3-day and 1-minute tests. The sensor readings provide tri-axial accelerometer signals in vertical (V), mediolateral (ML), and anterior-posterior (AP) axes. Additionally, signals contain gyroscope recordings in roll, pitch, and yaw directions.

In addition to wearable sensor recordings, participants underwent clinical functional performance assessment tests including the Dynamic Gait Index (DGI), the Berg Balance Scale (BBS), the Timed Up and Go test (TUG), the Four Square Step Test (FSST), the Mini-Mental State Examination (MMSE), and the Activities-specific Balance Confidence scale (ABC). Demographic information and fall history were recorded alongside these assessments. The fall status labels (controls and fallers) were assigned in accordance with the documented fall record of the past year. A person was classified as a faller if they reported a history of falling at least 2 times within the past 12 months, whereas all others were categorized otherwise.

Data pre-processing

The dataset for static features was analyzed to include only common subjects within 3-day and 1-minute tests, reducing the overall number of subjects to 63. The dataset exhibited a relatively balanced distribution between fallers and controls (34 and 29, respectively), allowing evaluation without additional resampling. The gender distribution was comparable between fallers and non-fallers, with 12 males and 22 females in the control (non-fallers) group and 11 males and 18 females in the fallers group. The mean age was 78 with an age range of 65–87.

Prior to the feature extraction, the raw inertial data were pre-processed to suppress noise and non-physiological artifacts. First, a median filter with a window length of three samples was applied independently to each accelerometer axis to attenuate impulsive noise while preserving step-related signal peaks. Subsequently, the quasi-static gravitational component was removed using a second-order Butterworth high-pass filter with a cutoff frequency of 0.5 Hz. Zero-phase filtering was employed to avoid phase distortion.

Following gravity removal, the acceleration signals were decomposed into two physiologically relevant frequency bands using fourth-order Butterworth band-pass filters. A step-frequency band (0.5–3 Hz) was extracted to facilitate reliable step detection and cadence estimation, while a broader gait-frequency band (0.5–15 Hz) was used for the computation of time-domain, frequency-domain, and gait stability features. Step events were identified as the local maxima in the vertical acceleration signal [18]. To avoid false detection, the peaks on the V graph were required to exceed an adaptive prominence threshold and to be separated by at least 0.3 s.

To mitigate inter-subject variability arising from differences in walking intensity, body mass, and sensor placement, each accelerometer axis was normalized by its maximum absolute value. The resulting normalized signals provided standardized representations of gait dynamics suitable for subsequent step detection, feature extraction, and machine learning analysis. The whole pre-processing and feature extraction pipeline is shown in Fig. 1.

Fig. 1.

Overview of the gait signal processing and feature extraction pipeline.

Feature extraction

From each set of trials, a comprehensive set of features was extracted from the pre-processed signals and categorized into two complementary perspectives, according to biomechanical and signal-level interpretations.

Spatiotemporal gait parameters (SP) are most associated with the functional status and have been linked to factors such as fear of falling, cognitive decline, and disease progression, making them representative in the scope of this study [19,20]. For instance, swing-time variability and stride-length variability are reported as robust markers of instability and fall-prone walking patterns among other quantitative gait markers [21]. In our methodology, the spatiotemporal parameters were derived from the temporal structure of the detected gait events. These features characterize the timing and rhythm of the heel strike event and include step time, stride time, gait cycle duration, cadence, step and stride regularity, etc. To quantify the consistency and stability, variability measures were computed across steps and strides. Additionally, high-order statistical descriptors were used to characterize the distributional properties of gait features: standard deviation, coefficient of variation, skewness, kurtosis, etc.

Inter-axis coordination features were computed to quantify the coupling between vertical, mediolateral, and anteroposterior trunk accelerations during gait. Pearson correlation coefficients and covariance measures were used to capture directional synchronisation and shared variability across axes. These features provide insights into multi-directional trunk dynamics and impaired motor coordination associated with the risk of falling.

Time-domain features describe temporal properties of walking and comprise stride regularity, signal energy, zero-crossing rate, and linear trend computed independently for each axis. Frequency-domain features were derived from the power spectral density of gait-band signals to quantify the distribution of movement energy across frequencies. They include spectral entropy, spectral flatness, spectral crest factor, dominant harmonic amplitude, and prominence. These features describe the rhythmicity and spectral complexity of gait.

Handcrafted features were computed separately from one-minute laboratory test and three-day free-living recordings and subsequently concatenated at the subject level.

Model architecture

This methodology utilizes a hybrid architecture in which Temporal Convolutional Networks (TCN) is used to learn high-level representations from raw IMU signals, which are then concatenated with the handcrafted gait features and processed by a Linear Head with sigmoid activation function for binary classification of fallers versus non-fallers [22]. Each input to TCN consisted of fixed-length temporal windows extracted from raw one-minute inertial measurements. The dynamic input comprised six synchronized channels, including three linear accelerations (in ML, V, and AP axes) along with the three angular velocities in roll, pitch, and yaw directions. A 4th-order Butterworth band-pass filter with cutoff frequencies 0.5 - 10 Hz was applied to all channels to remove noise while preserving dominant gait features.

Signals were sampled at 100 Hz and segmented into the overlapping windows of fixed durations. Each window was represented as a tensor:

$$X_k\in R^{C\times T},$$

where $C=6$ represents the number of channels and T is the number of temporal samples. Each window corresponds to a contiguous sequence:

$$X_k=\left[x\left(t_k\right),x\left(t_k+1\right),...,x\left(t_k+T-1\right)\right]$$ (1)

where $t_k$ denotes the starting time index of the k-th index. Windows of 15 s (1500 samples) are extracted with a stride of 1 s. If a window extends beyond the available signal length, it is padded by repeating the final sample to preserve window dimensionality.

The dynamic signals were then concatenated with the set of subject-level handcrafted features and demographic data to capture the subject-specific characteristics. Dataset splitting was performed at the subject level prior to window construction. Feature selection was performed within the cross-validation loop using the Mutual Information (MI) method, with the 5 highest-ranked features selected and passed to the classifier as part of an integrated machine-learning pipeline. Dynamic and static features were then separately normalized. The dynamic data, all training windows are flattened across time and subjects to fit a StandardScaler, which is subsequently applied to test and validation sets.

Temporal convolutional network structure

The TCN architecture consisted of a stack of one-dimensional convolutional blocks with exponentially increasing dilation factors, enabling the network to capture temporal dependencies over multiple time scales without relying on recurrent connections. To capture temporal dependencies across the raw measurements, dilated one-dimensional convolutions were incorporated. A dilated convolution is defined as:

$$y\left(t\right)=\sum _{k=0}^{K-1}{w}_{k}x\left(t-dk\right),$$ (2)

where $x\left(t\right)$ is the input signal, $w_k$ are the convolutional filter weights, $K$ is the kernel size, and $d$ is the dilation factor.

Dilation factors increased exponentially across successive layers, $d=2^i,$ allowing the receptive field to grow efficiently with the network depth. Residual connections were implemented using identity mappings or $1\times 1$ convolutions when input and output channel dimensions differed. The output of the $i$-th block is given by:

$$y^{\left(l\right)}=F\left(x^{\left(l\right)}\right)+x^{\left(l\right)},$$ (3)

Where $x^{\left(l\right)}$ is the block input, and $F(·)$is the nonlinear transformation implemented by the dilated convolution.

The temporal receptive field (RF) determines the range of input samples that produce each output activation [23]. Given a network with a kernel size of $k$, $L$ convolutional layers, and dilation factors $\left\{d_1,d_2,\dots ,d_L\right\}$, the receptive field is:

$$R=1+\sum _{l=1}^L\left(k-1\right)d_l$$ (4)

This implies exponential growth of the receptive field with depth, allowing the model to capture long-term dependencies. The TCN block consists of a one-dimensional convolutional layer with a kernel size of k and a dilation factor of d, a leaky rectified linear unit (ReLU) activation function, a dropout layer for regularization, and a residual connection for stable gradient propagation (Fig. 2).

Fig. 2.

Proposed Temporal convolutional network architecture with dilated causal convolutions and subject-level feature fusion for fall risk classification.

Subject-level temporal aggregation

The gait recordings produce a set of overlapping temporal windows. Let

$$\left\{X_1,X_2,\dots ,X_{N_{\omega }}\right\}$$

represent a $N_{\omega }$ window extracted from an inertial signal of a subject. Each window $X_i\in R^{C\times T}$ is independently processed by TCN to obtain a latent embedding:

$$h_i=f_{TCN}\left(X_i\right),h_i\in R^D$$ (5)

Where $D$ is the dimensionality of the final TCN feature space. The window-level embeddings are then statistically aggregated to capture both the central tendency and variability of gait patterns across time. Specifically, the following statistics are computed across the window embeddings:

$$Meanembedding:\mu =\frac{1}{N}\sum _{i=1}^{N_{\omega }}{h}_i$$ (6)

$$Standarddeviation:\sigma =\sqrt{\frac{1}{N_{\omega }}\sum _{i=1}^{N_{\omega }}{\left(h_i-\mu \right)}^2}$$ (7)

$$Maximumactivation:h_{max}=ma{x}_{i=1,\dots ,N_{\omega }}{h}_i$$ (8)

$$Skewness:Skew\left(h\right)=\frac{1}{N_{\omega }}\sum _{i=1}^{N_{\omega }}\frac{{\left(h_i-\mu \right)}^3}{{\sigma }^3+\epsilon }$$ (9)

The temporal stability of gait patterns is explicitly modeled by computing the cosine similarity between consecutive window embeddings [24]:

$$cos\left(h_i,h_{i+1}\right)=\frac{h_i·h_{i+1}}{\left|\left|h_i\left|\left|\right|\right|h_{i+1}\right|\right|}$$ (10)

This metric quantifies the directional alignment of embeddings over time, with values closer to 1 indicating higher temporal consistency. We summarize the temporal evolution by computing the mean and standard deviation of these cosine similarity values across all adjacent windows. Finally, the subject-level dynamic representation is constructed by concatenating these temporal features (mean and standard deviation), along with additional descriptors such as maximum, skewness, and overall cosine similarity.

Classification head

The output of the last TCN layer produces a latent feature map with $C_{TCN}$channels (which is determined by the last convolutional block), which is further concatenated with the static feature vector and passed to a linear classification head. Mean pooling was applied along the temporal dimension to obtain a fixed-length representation. The fused feature vector is passed through a fully connected layer to produce a single logit representing fall risk. Dropout is applied before the classification layer to reduce overfitting. The model outputs a subject-level probability via a sigmoid activation during evaluation.

Training

The model was trained using the Adam optimizer with a fixed learning rate of $3\times {10}^{-4}$ and a weight decay of 0.05. To prevent data leakage, data splitting was performed at the record (subject) level prior to window generation. Specifically, the list of gait recordings was first partitioned into training, validation, and test sets, and temporal windows were subsequently constructed independently within each split. A lightweight data augmentation was applied by adding Gaussian noise to dynamic signals and random amplitude sampling. Class imbalance was addressed using a weighted binary cross-entropy loss, where the positive class weight was computed from the training fold distribution.

Method validation

Model validation was performed at the subject level using a 5-fold subject-wise cross-validation. Subjects are split such that there are no individual appears in more than one fold. Classification performance was quantified using accuracy:

$$Accuracy=\frac{1}{N}\sum _{i=1}^{N}1\left\{\widehat{y_i}=y_i\right\}$$ (11)

where $N$ is the number of samples, $y_i$ is the ground-truth label, $\widehat{y_i}$ is the predicted label, and $1(·)$ is the indicator function. To assess the discriminative ability of the model independent of the fixed decision threshold, the area under the receiver operating characteristic curve (ROC-AUC) was also computed. ROC–AUC represents the probability that a randomly selected faller receives a higher predicted score than a randomly selected non-faller:

$$AUC=P\left(s\left(x^{+}\right)>s\left(x^{-}\right)\right),$$ (12)

where $s\left(x\right)$ denotes a classifier score, $x^{+}$ is a positive (faller), and $x^{-}$ is a negative(faller). An optimal classification threshold is selected using Youden’s index on the ROC curve [25].

Model optimization was performed using a weighted binary cross-entropy loss with logits. The model outputs a single real-valued logit $z$ for each subject, which represents the unnormalized likelihood of belonging to the faller class. A positive class weighting strategy was applied, defined as:

$$\omega =\frac{N_{non-fallers}}{N_{fallers}},$$ (13)

where $N_{non-fallers}$ and $N_{fallers}$ denote the number of subjects in each class within the training set.

The resulting weighted loss for a single subject is:

$$L\left(y,z\right)=-\omega ·ylog\left(\sigma \left(z\right)\right)-\left(1-z\right)log\left(1-\sigma \left(z\right)\right),$$ (14)

where $y\in \left\{0,1\right\}$ is a ground-truth label and $\sigma (·)$ is a sigmoid activation [26]. The sigmoid operation is applied only during loss computation and inference, while the network itself outputs raw logits to ensure numerical stability during optimization. During inference, predicted class labels were obtained by applying a sigmoid activation to the model output. Table 1 summarizes the classification performance obtained across the five folds of subject-wise cross-validation.

Table 1.

Fold-wise classification performance obtained using subject-wise cross-validation.

N Fold	Accuracy	AUC
1	0.846	0.929
2	0.538	0.525
3	0.725	0.692
4	0.667	0.686
5	0.833	0.914

Five-fold subject-wise cross-validation resulted in an average accuracy of 0.722 and a ROC-AUC of 0.749.

Subject-wise analysis was performed to identify subjects in low-performing test splits, using gait-temporal feature z-scores (|z| ≥ 2.5) to detect extreme outliers. No extreme gait-temporal outliers (|z| ≥ 2.5) were observed among low-performing subjects, suggesting that misclassification was not driven by obvious signal artifacts or extreme gait abnormalities but rather inter-subject gait heterogeneity. These results indicate that inherent gait variability, occurring both within and across individuals, plays a critical role in influencing predictive model accuracy as well as clinical outcome assessments, as supported by previous studies [15,27].

Limitations

The proposed methodology has several limitations. First, the dataset size is relatively small, which may limit the generalizability of the findings to a broader population. Further study with a larger and more diverse population in the dataset is required. In addition, in this work, we used one-minute recordings as the input to the TCN block. This test reflects laboratory recordings collected under controlled conditions and provides standardized gait measurements for a short time period, which cannot fully capture long-term gait patterns, resulting in limited intra-subject variability. Future work should focus on incorporating multi-day free-living gait data not only in static feature extraction but as a dynamic input to the model to better characterize habitual walking behavior and account for environmental and task-related factors influencing fall risk.

Ethics statements

This study does not include any collection of data.

CRediT author statement

Almira Askhatova: Conceptualization, Methodology, Software. Ulan Sharipov: Validity tests, Data curation, Software. Sultan Kasenov: Visualization, Investigation, Software. Khanat Kassenov: Supervision. Zhaxat Kenzhin: Reviewing and Editing. Tursynzada Kuangaliyeva: Reviewing and Editing. Dinara Turzhanova: Reviewing and Editing, Prasant Jamwal: Conceptualization, Methodology, Supervision.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Data availability

Data will be made available on request.