Deep Learning-Based Prediction Model for Gait Recovery after a Spinal Cord Injury.

Predicting gait recovery after a spinal cord injury (SCI) during an acute rehabilitation phase is important for planning rehabilitation strategies. However, few studies have been conducted on this topic to date. In this study, we developed a deep learning-based prediction model for gait recovery after SCI upon discharge from an acute rehabilitation facility. Data were collected from 405 patients with acute SCI admitted to the acute rehabilitation facility of Korea University Anam Hospital between June 2008 and December 2022. The dependent variable was Functional Ambulation Category at the time of discharge (FAC-DC). Seventy-one independent variables were selected from the existing literature: basic information, International Standards for Neurological Classification of SCI scores, neurogenic bladders, initial FAC, and somatosensory-evoked potentials of the lower extremity. Recurrent neural network (RNN), linear regression (LR), Ridge, and Lasso methods were compared for FAC-DC prediction in terms of the root-mean-squared error (RMSE). RNN variable importance, which is the RMSE gap between a complete RNN model and an RNN model excluding a certain variable, was used to evaluate the contribution of this variable. Based on the results of this study, the performance of the RNN was far better than that of LR, Ridge, and Lasso. The respective RMSEs were 0.3738, 2.2831, 1.3161, and 1.0246 for all the participants; 0.3727, 1.7176, 1.3914, and 1.3524 for those with trauma; and 0.3728, 1.7516, 1.1012, and 0.8889 for those without trauma. In terms of RNN variable importance, lower-extremity motor strength (right and left ankle dorsiflexors, right knee extensors, and left long toe extensors) and the neurological level of injury were ranked among the top five across the boards. Therefore, initial FAC was the seventh, third, and ninth most important predictor for all participants, those with trauma, and those without trauma, respectively. In conclusion, this study developed a deep learning-based prediction model with excellent performance for gait recovery after SCI at the time of discharge from an acute rehabilitation facility. This study also demonstrated the strength of deep learning as an explainable artificial intelligence method for identifying the most important predictors.

Measurements of Electrodermal Activity, Tissue Oxygen Saturation, and Visual Analog Scale for Different Cuff Pressures.

The quantification of comfort in binding parts, essential human-machine interfaces (HMI) for the functioning of rehabilitation robots, is necessary to reduce physical strain on the user despite great achievements in their structure and control. This study aims to investigate the physiological impacts of binding parts by measuring electrodermal activity (EDA) and tissue oxygen saturation (StO2). In Experiment 1, EDA was measured from 13 healthy subjects under three different pressure conditions (10, 20, and 30 kPa) for 1 min using a pneumatic cuff on the right thigh. In Experiment 2, EDA and StO2 were measured from 10 healthy subjects for 5 min. To analyze the correlation between EDA parameters and the decrease in StO2, a survey using the visual analog scale (VAS) was conducted to assess the level of discomfort at each pressure. The EDA signal was decomposed into phasic and tonic components, and the EDA parameters were extracted from these two components. RM ANOVA and a post hoc paired t-test were used to determine significant differences in parameters as the pressure increased. The results showed that EDA parameters and the decrease in StO2 significantly increased with the pressure increase. Among the extracted parameters, the decrease in StO2 and the mean SCL proved to be effective indicators. Such analysis outcomes would be highly beneficial for studies focusing on the comfort assessment of the binding parts of rehabilitation robots.

Determination of Gait Events and Temporal Gait Parameters for Persons with a Knee-Ankle-Foot Orthosis.

Gait event detection is essential for controlling an orthosis and assessing the patient's gait. In this study, patients wearing an electromechanical (EM) knee-ankle-foot orthosis (KAFO) with a single IMU embedded in the thigh were subjected to gait event detection. The algorithm detected four essential gait events (initial contact (IC), toe off (TO), opposite initial contact (OIC), and opposite toe off (OTO)) and determined important temporal gait parameters such as stance/swing time, symmetry, and single/double limb support. These gait events were evaluated through gait experiments using four force plates on healthy adults and a hemiplegic patient who wore a one-way clutch KAFO and a pneumatic cylinder KAFO. Results showed that the smallest error in gait event detection was found at IC, and the largest error rate was observed at opposite toe off (OTO) with an error rate of -2.8 ± 1.5% in the patient group. Errors in OTO detection resulted in the largest error in determining the single limb support of the patient with an error of 5.0 ± 1.5%. The present study would be beneficial for the real-time continuous monitoring of gait events and temporal gait parameters for persons with an EM KAFO.

TinyFallNet: A Lightweight Pre-Impact Fall Detection Model.

Falls represent a significant health concern for the elderly. While studies on deep learning-based preimpact fall detection have been conducted to mitigate fall-related injuries, additional efforts are needed for embedding in microcomputer units (MCUs). In this study, ConvLSTM, the state-of-the-art model, was benchmarked, and we attempted to lightweight it by leveraging features from image-classification models VGGNet and ResNet while maintaining performance for wearable airbags. The models were developed and evaluated using data from young subjects in the KFall public dataset based on an inertial measurement unit (IMU), leading to the proposal of TinyFallNet based on ResNet. Despite exhibiting higher accuracy (97.37% < 98.00%) than the benchmarked ConvLSTM, the proposed model requires lower memory (1.58 MB > 0.70 MB). Additionally, data on the elderly from the fall data of the FARSEEING dataset and activities of daily living (ADLs) data of the KFall dataset were analyzed for algorithm validation. This study demonstrated the applicability of image-classification models to preimpact fall detection using IMU and showed that additional tuning for lightweighting is possible due to the different data types. This research is expected to contribute to the lightweighting of deep learning models based on IMU and the development of applications based on IMU data.

Quantification of Comfort for the Development of Binding Parts in a Standing Rehabilitation Robot.

Human-machine interfaces (HMI) refer to the physical interaction between a user and rehabilitation robots. A persisting excessive load leads to soft tissue damage, such as pressure ulcers. Therefore, it is necessary to define a comfortable binding part for a rehabilitation robot with the subject in a standing posture. The purpose of this study was to quantify the comfort at the binding parts of the standing rehabilitation robot. In Experiment 1, cuff pressures of 10-40 kPa were applied to the thigh, shank, and knee of standing subjects, and the interface pressure and pain scale were obtained. In Experiment 2, cuff pressures of 10-20 kPa were applied to the thigh, and the tissue oxygen saturation and the skin temperature were measured. Questionnaire responses regarding comfort during compression were obtained from the subjects using the visual analog scale and the Likert scale. The greatest pain was perceived in the thigh. The musculoskeletal configuration affected the pressure distribution. The interface pressure distribution by the binding part showed higher pressure at the intermuscular septum. Tissue oxygen saturation (StO2) increased to 111.9 ± 6.7% when a cuff pressure of 10 kPa was applied and decreased to 92.2 ± 16.9% for a cuff pressure of 20 kPa. A skin temperature variation greater than 0.2 °C occurred in the compressed leg. These findings would help evaluate and improve the comfort of rehabilitation robots.

sEMG-Based Hand Posture Recognition and Visual Feedback Training for the Forearm Amputee.

sEMG-based gesture recognition is useful for human-computer interactions, especially for technology supporting rehabilitation training and the control of electric prostheses. However, high variability in the sEMG signals of untrained users degrades the performance of gesture recognition algorithms. In this study, the hand posture recognition algorithm and radar plot-based visual feedback training were developed using multichannel sEMG sensors. Ten healthy adults and one bilateral forearm amputee participated by repeating twelve hand postures ten times. The visual feedback training was performed for two days and five days in healthy adults and a forearm amputee, respectively. Artificial neural network classifiers were trained with two types of feature vectors: a single feature vector and a combination of feature vectors. The classification accuracy of the forearm amputee increased significantly after three days of hand posture training. These results indicate that the visual feedback training efficiently improved the performance of sEMG-based hand posture recognition by reducing variability in the sEMG signal. Furthermore, a bilateral forearm amputee was able to participate in the rehabilitation training by using a radar plot, and the radar plot-based visual feedback training would help the amputees to control various electric prostheses.

Fall-from-Height Detection Using Deep Learning Based on IMU Sensor Data for Accident Prevention at Construction Sites.

Workers at construction sites are prone to fall-from-height (FFH) accidents. The severity of injury can be represented by the acceleration peak value. In the study, a risk prediction against FFH was made using IMU sensor data for accident prevention at construction sites. Fifteen general working movements (NF: non-fall), five low-hazard-fall movements, (LF), and five high-hazard-FFH movements (HF) were performed by twenty male subjects and a dummy. An IMU sensor was attached to the T7 position of the subject to measure the three-axis acceleration and angular velocity. The peak acceleration value, calculated from the IMU data, was 4 g or less in general work movements and 9 g or more in FFHs. Regression analysis was performed by applying various deep learning models, including 1D-CNN, 2D-CNN, LSTM, and Conv-LSTM, to the risk prediction, and then comparing them in terms of their mean absolute error (MAE) and mean squared error (MSE). The FFH risk level was estimated based on the predicted peak acceleration. The Conv-LSTM model trained by MAE showed the smallest error (MAE: 1.36 g), and the classification with the predicted peak acceleration showed the best accuracy (97.6%). This study successfully predicted the FFH risk levels and could be helpful to reduce fatal injuries at construction sites.

sEMG-Based Hand Posture Recognition Considering Electrode Shift, Feature Vectors, and Posture Groups.

Surface electromyography (sEMG)-based gesture recognition systems provide the intuitive and accurate recognition of various gestures in human-computer interaction. In this study, an sEMG-based hand posture recognition algorithm was developed, considering three main problems: electrode shift, feature vectors, and posture groups. The sEMG signal was measured using an armband sensor with the electrode shift. An artificial neural network classifier was trained using 21 feature vectors for seven different posture groups. The inter-session and inter-feature Pearson correlation coefficients (PCCs) were calculated. The results indicate that the classification performance improved with the number of training sessions of the electrode shift. The number of sessions necessary for efficient training was four, and the feature vectors with a high inter-session PCC (r > 0.7) exhibited high classification accuracy. Similarities between postures in a posture group decreased the classification accuracy. Our results indicate that the classification accuracy could be improved with the addition of more electrode shift training sessions and that the PCC is useful for selecting the feature vector. Furthermore, hand posture selection was as important as feature vector selection. These findings will help in optimizing the sEMG-based pattern recognition algorithm more easily and quickly.

The Performance of Post-Fall Detection Using the Cross-Dataset: Feature Vectors, Classifiers and Processing Conditions.

In this study, algorithms to detect post-falls were evaluated using the cross-dataset according to feature vectors (time-series and discrete data), classifiers (ANN and SVM), and four different processing conditions (normalization, equalization, increase in the number of training data, and additional training with external data). Three-axis acceleration and angular velocity data were obtained from 30 healthy male subjects by attaching an IMU to the middle of the left and right anterior superior iliac spines (ASIS). Internal and external tests were performed using our lab dataset and SisFall public dataset, respectively. The results showed that ANN and SVM were suitable for the time-series and discrete data, respectively. The classification performance generally decreased, and thus, specific feature vectors from the raw data were necessary when untrained motions were tested using a public dataset. Normalization made SVM and ANN more and less effective, respectively. Equalization increased the sensitivity, even though it did not improve the overall performance. The increase in the number of training data also improved the classification performance. Machine learning was vulnerable to untrained motions, and data of various movements were needed for the training.

Detection of Pre-Impact Falls from Heights Using an Inertial Measurement Unit Sensor.

Many safety accidents can occur in industrial sites. Among them, falls from heights (FFHs) are the most frequent accidents and have the highest fatality rate. Therefore, some existing studies have developed personal wearable airbags to mitigate the damage caused by FFHs. To utilize these airbags effectively, it is essential to detect FFHs before collision with the floor. In this study, an inertial measurement unit (IMU) sensor attached to the seventh thoracic vertebrae (T7) was used to develop an FFH detection algorithm. The vertical angle and vertical velocity were calculated using the inertial data obtained from the IMU sensor. Forty young and healthy males were recruited to perform non-FFH and FFH motions. In addition, experiments using a human mannequin and dynamics simulations were performed to obtain FFH data at heights above 2 m. The developed algorithm achieved 100% FFH detection accuracy and provided sufficient lead time such that the airbags could be inflated completely before collision with the floor.

Enhanced Algorithm for the Detection of Preimpact Fall for Wearable Airbags.

Fall-related injury is a common cause of mortality among the elderly. Hip fractures are especially dangerous and can even be fatal. In this study, a threshold-based preimpact fall detection algorithm was developed for wearable airbags that minimize the impact of falls on the user's body. Acceleration sum vector magnitude (SVM), angular velocity SVM, and vertical angle, calculated using inertial data captured from an inertial measurement unit were used to develop the algorithm. To calculate the vertical angle accurately, a complementary filter with a proportional integral controller was used to minimize integration errors and the effect of external impacts. In total, 30 healthy young men were recruited to simulate 6 types of falls and 14 activities of daily life. The developed algorithm achieved 100% sensitivity, 97.54% specificity, 98.33% accuracy, and an average lead time (i.e., the time between the fall detection and the collision) of 280.25 ± 10.29 ms with our experimental data, whereas it achieved 96.1% sensitivity, 90.5% specificity, and 92.4% accuracy with the SisFall public dataset. This paper demonstrates that the algorithm achieved a high accuracy using our experimental data, which included some highly dynamic motions that had not been tested previously.

Evaluation of Inertial Sensor-Based Pre-Impact Fall Detection Algorithms Using Public Dataset.

In this study, pre-impact fall detection algorithms were developed based on data gathered by a custom-made inertial measurement unit (IMU). Four types of simulated falls were performed by 40 healthy subjects (age: 23.4 ± 4.4 years). The IMU recorded acceleration and angular velocity during all activities. Acceleration, angular velocity, and trunk inclination thresholds were set to 0.9 g, 47.3°/s, and 24.7°, respectively, for a pre-impact fall detection algorithm using vertical angles (VA algorithm); and 0.9 g, 47.3°/s, and 0.19, respectively, for an algorithm using the triangle feature (TF algorithm). The algorithms were validated by the results of a blind test using four types of simulated falls and six types of activities of daily living (ADL). VA and TF algorithms resulted in lead times of 401 ± 46.9 ms and 427 ± 45.9 ms, respectively. Both algorithms were able to detect falls with 100% accuracy. The performance of the algorithms was evaluated using a public dataset. Both algorithms detected every fall in the SisFall dataset with 100% sensitivity). The VA algorithm had a specificity of 78.3%, and TF algorithm had a specificity of 83.9%. The algorithms had higher specificity when interpreting data from elderly subjects. This study showed that algorithms using angles could more accurately detect falls. Public datasets are needed to improve the accuracy of the algorithms.