Measuring the daily use of an affected limb after hospital discharge is crucial for hemiparetic stroke rehabilitation. Classifying movements using non-intrusive wearable sensors provides context for arm use and is essential for the development of a home rehabilitation system. However, the movement classification of stroke patients poses unique challenges, including variability and sparsity. To address these challenges, we collected movement data from 15 hemiparetic stroke patients (Stroke group) and 29 non-disabled individuals (ND group). The participants performed two different tasks, the range of motion (14 movements) task and the activities of daily living (56 movements) task, wearing five inertial measurement units in a home setting. We trained a 1D convolutional neural network and evaluated its performance for different training groups: ND-only, Stroke-only, and ND and Stroke jointly. We further compared the model performance with data augmentation from axis rotation and investigated how the performance varied based on the asymmetry of movements. The joint training of ND + Stroke yielded an increased F1-score by a margin of 31.6% and 10.6% compared to ND-only training and Stroke-only training, respectively. Data augmentation further enhanced F1-scores across all conditions by an average of 11.3%. Finally, asymmetric movements decreased the F1-score by 25.9% compared to symmetric movements in the Stroke group, indicating the importance of asymmetry in movement classification.

Conventional methods for monitoring ammonia (NH₃) emissions from livestock farms have several challenges, such as a poor environment for measurement, difficulty in accessing livestock, and problems with long-term measurement. To address these issues, we applied various neural network models for the long-term prediction of NH₃ concentrations from sow farms in this study. Environmental parameters, including temperature, humidity, ventilation rate, and past records of NH₃ concentrations, were given as inputs to the models. These neural network models took the encoder or the feature extracting parts from the representative deep learning models, including Long Short-Term Memory (LSTM), Convolutional Neural Network (CNN), and Transformer, to encode temporal patterns of time series. However, all of these models adopted dense layers for the decoder to format the task of long-term prediction as a regression problem. Due to their regression nature, all models showed a robust performance in predicting long-term NH₃ concentrations at a scale of weeks or even months despite there being a relatively short period of input signals (a few days to a week). Given one week of input, LSTM showed the minimum mean absolute errors (MAE) of 1.83, 1.78, and 1.87 ppm for the prediction of one, two, and three weeks, respectively, whereas Transformer performed best with a MAE of 1.73 ppm for a four-week prediction. In the long-term estimation of spanning months, LSTM showed the minimum MAEs of 1.95 and 1.90 ppm when trained on predicting two and three weeks of windows. At the same condition, Transformer gave the minimum MAEs of 1.87 and 1.83 when trained on predicting one and four weeks of windows. Overall, the neural network models can facilitate the prediction of national-level NH₃ emissions, the development of mitigation strategies for NH₃-derived air pollutants, odor management, and the monitoring of animal-rearing environments. Further, their integration of real-time measurement devices can significantly prolong device longevity and offer substantial cost savings.

The emergence of wearable devices has motivated people to actively log their daily exercise routines using smart apps. However, most current exercise trackers focus on aerobic exercises, and thus provide limited functionality for tracking and analyzing anaerobic workouts involving complex and repetitive movements. To fill this gap, we developed uLift, an adaptive workout tracker that uses only a single wrist-worn accelerometer and has four main functions: workout detection, repetition counting, workout classification, and quality assessment. First, uLift detects a binary workout state from continuous signals using the weighted sum of autocorrelation. Second, repetition counting is conducted by filtering out unwanted peaks. Third, the segments of a workout are used to generate a representative template for workout classification using the distances calculated from dynamic time warping. Finally, to assess workout quality, the form score is computed by evaluating the consistency across repetitions. As uLift does not require a training process, it can easily add new workouts or delete existing ones using an instant adaptation process. For the evaluation of uLift, we collected a multi-joint workout dataset comprising 15 workouts from 35 participants in a gymnasium. To allow for natural and individual variability, we provided the participants with minimum instructions. The dataset was open-sourced to facilitate future research on anaerobic workout analysis. As a result, uLift achieved 93.09% accuracy for workout detection, mean counting error of 0.61, and classification accuracy of 90.06%. The form score was significantly different among the three subgroups of participants, divided by workout experience.

BACKGROUND Measuring the use of an affected limb in a home setting after hospital discharge is crucial for stroke rehabilitation. Classifying movements using non-intrusive wearable sensor data provides a context for arm use, and it is essential for the development of a home rehabilitation system with monitoring and feedback functions. However, classification of the movements of stroke patients poses some challenges, including variability in duration and patterns, as well as data sparsity. OBJECTIVE We aimed to investigate methods for the preprocessing, training, and evaluation of a deep learning model to effectively classify the movements of stroke patients with hemiparesis. To address the abovementioned challenges, we adopted linear interpolation in preprocessing, compared the model performance using different training groups (non-disabled individuals and patients with stroke), tested the effect of data augmentation, and observed how the asymmetry of movements affected the model performance. METHODS Fifteen patients with subacute to chronic stroke (Stroke group) and 29 non-disabled individuals (ND group) participated in two different tasks, a range of motion (ROM; 14 movements) task and activities of daily living (ADL; 56 movements) task, wearing five inertial measurement unit sensors in a home setting. We trained a one-dimensional convolutional neural network (1D-CNN) and evaluated its performance using the F1-score for different training conditions, including ND only, Stroke only, and combined ND+Stroke. In addition, we tested the effect of axis rotation as a data augmentation technique. The model performance was further investigated with different asymmetricities of movement. RESULTS Training with combined ND+Stroke data demonstrated the best results (ROM, F1=0.721; ADL, F1=0.603), followed by Stroke only (ROM, F1=0.676; ADL, F1=0.526) and ND only (ROM, F1=0.553; ADL, F1=0.454) for the original and augmented data. Axis rotation consistently boosted performance in all training conditions (ROM, P=.014; ADL, P<.001). For the Stroke group, bimanual symmetric movements showed significantly better performance (F1=0.757) than bimanual asymmetric (F1=0.541) and unimanual (F1=0.581) movements (P<.001) in the ADL task. CONCLUSIONS Effectively training a deep learning model for the classification of the movements of stroke patients requires adjusting the data length to a fixed size. Despite the distinct and diverse patterns of stroke movements, it is still beneficial to train a model using combined ND+Stroke data. Proper data augmentation, such as axis rotation, further enhances the model performance. Finally, the result on stroke data was differently affected by the asymmetry of movements, with better performance for symmetric than asymmetric movements. These findings may provide the best practices and processes for classifying the movements of stroke patients when training and evaluating deep learning models.

Stroke survivors with hemiparesis require extensive home-based rehabilitation. Deep learning-based classifiers can detect actions and provide feedback based on patient data; however, this is difficult owing to data sparsity and heterogeneity. In this study, we investigate data augmentation and model training strategies to address this problem. Three transformations are tested with varying data volumes to analyze the changes in the classification performance of individual data. Moreover, the impact of transfer learning relative to a pre-trained one-dimensional convolutional neural network (Conv1D) and training with an advanced InceptionTime model are estimated with data augmentation. In Conv1D, the joint training data of non-disabled (ND) participants and double rotationally augmented data of stroke patients is observed to outperform the baseline in terms of F1-score (60.9% vs. 47.3%). Transfer learning pre-trained with ND data exhibits 60.3% accuracy, whereas joint training with InceptionTime exhibits 67.2% accuracy under the same conditions. Our results indicate that rotational augmentation is more effective for individual data with initially lower performance and subset data with smaller numbers of participants than other techniques, suggesting that joint training on rotationally augmented ND and stroke data enhances classification performance, particularly in cases with sparse data and lower initial performance.