Kai-Chun Liu

Spinal muscular atrophy (SMA) and Duchenne muscular dystrophy (DMD) are rare, childhood-onset neuromuscular diseases (NMDs) characterized by progressive muscle weakness that frequently leads to loss of ambulation. Although several approved disease-modifying therapies (DMTs)–and others currently under investigation–have shown promising results and significantly altered disease trajectories, the standardized in-clinic assessments commonly used as trial endpoints often fail to capture meaningful real-world mobility improvements reported by treated patients. This gap complicates accurate evaluation of therapeutic efficacy and may contribute to delays in patient access to emerging treatments. Wearable-based gait monitoring offers a practical solution for remotely assessing ambulatory function in individuals with motor impairments. Real-world gait patterns can serve as sensitive functional biomarkers to monitor treatment response and track disease progression. However, most real-world studies on NMDs have focused on volume metrics such as step count, which lack stride-level granularity and cannot capture independent gait domains, or on basic spatiotemporal (SPT) metrics related to pace, rhythm, or variability. Kinetic parameters–such as center-of-pressure (COP)-derived metrics–are critical for assessing dynamic stability and muscle strength deficits in NMDs, yet they remain largely unexplored in real-world settings. Unlike other wearables, instrumented insoles allow for the assessment of both SPT and kinetic parameters, with stride-level granularity. Towards the development of multi-domain functional digital biomarkers for ambulatory NMDs, this study evaluates the construct, concurrent, and known-groups validity of insole-based, stride-level SPT and COP-derived parameters in individuals with DMD and SMA during walking tasks designed to simulate real-world walking. Mean absolute errors (MAEs) relative to gold-standard gait-laboratory equipment were used to assess construct validity, while correlations between insole-based metrics and standardized clinical assessment tests provided proof-of-concept concurrent validity. To establish the added value of COP metrics with respect to known-groups validity, we also conducted a comparative analysis of the ability of insole-based SPT and COP-derived features to discriminate between individuals with DMD or SMA and healthy controls. Results suggest that incorporating COP-derived metrics alongside more conventional SPT measures may help better characterize disease-specific gait signatures and their relationships with functional impairment in ambulatory individuals with DMD and SMA. COP metrics were more informative than SPT measures in DMD, whereas they were largely complementary to SPT measures in SMA, with the anteroposterior range of COP (AP-COP) nonetheless standing out for its associations with lower-limb strength and walking capacity. Overall, this work provides proof-of-concept evidence of the utility of combined SPT and COP gait parameters as candidate functional digital biomarkers in ambulatory NMDs, paving the way for more extensive evaluations in free-living settings.

The rising aging population has increased the importance of fall detection (FD) systems as an assistive technology, where deep learning techniques are widely applied to enhance accuracy. FD systems typically use edge devices (EDs) worn by individuals to collect real-time data, which are transmitted to a cloud center (CC) or processed locally. However, this architecture faces challenges such as a limited ED model size and data transmission latency to the CC. Mobile edge computing (MEC), which allows computations at MEC servers deployed between EDs and CC, has been explored to address these challenges. We propose a multilayer MEC (MLMEC) framework to balance accuracy and latency. The MLMEC splits the architecture into stations, each with a neural network model. If front-end equipment cannot detect falls reliably, data are transmitted to a station with more robust back-end computing. The knowledge distillation (KD) approach was employed to improve front-end detection accuracy by allowing high-power back-end stations to provide additional learning experiences, enhancing precision while reducing latency and processing loads. Simulation results demonstrate that the KD approach improved accuracy by 11.65% on the SisFall dataset and 2.78% on the FallAllD dataset. The MLMEC with KD also reduced the data latency rate by 54.15% on the FallAllD dataset and 46.67% on the SisFall dataset compared to the MLMEC without KD. In summary, the MLMEC FD system exhibits improved accuracy and reduced latency.

While traditional electrocardiogram (ECG) monitoring provides vital clinical information, its electrode-based setup restricts patient movement. To address this limitation, contactless ECG monitoring using frequency-modulated continuous-wave (FMCW) radar and deep learning have been developed. However, such approaches face challenges owing to the limited availability of training data and inherent discrepancies between radar and ECG signals. This paper introduces a novel approach to transforming high-fidelity ECG signals from millimeter-wave (mmWave) radar signals reflecting cardiac mechanical activity. The proposed method uses a cascade framework with a cross-modal autoencoder trained using joint waveforms, spectrograms, and deep feature losses. This strategy enables the model to leverage a pre-trained ECG-to-ECG autoencoder and cardiac event predictor for learning general ECG representations while simultaneously capturing time- and frequency-domain features from limited data. We evaluated the effectiveness of the proposed autoencoder model in terms of the signal quality and cardiac event integrity using ablation studies on data from 20 healthy participants. The model achieved high transformation accuracy with a cross-correlation of 0.914 and average timing errors below 31 ms for critical ECG features. These findings demonstrate the feasibility and effectiveness of the proposed FMCW radar-based contactless ECG monitoring method, particularly in overcoming the limitations imposed by small datasets and domain discrepancies.

Surface electromyography (sEMG) is a widely employed bio-signal that captures human muscle activity via electrodes placed on the skin. Several studies have proposed methods to remove sEMG contaminants, as non-invasive measurements render sEMG susceptible to various contaminants. However, these approaches often rely on heuristic-based optimization and are sensitive to the contaminant type. A more potent, robust, and generalized sEMG denoising approach should be developed for various healthcare and human–computer interaction applications. This paper proposes a novel neural network (NN)-based sEMG denoising method called TrustEMG-Net. It leverages the potent nonlinear mapping capability and data-driven nature of NNs. TrustEMG-Net adopts a denoising autoencoder structure by combining U-Net with a Transformer encoder using a representation-masking approach. The proposed approach is evaluated using the Ninapro sEMG database with five common contamination types and signal-to-noise ratio (SNR) conditions. Compared with existing sEMG denoising methods, TrustEMG-Net achieves exceptional performance across the five evaluation metrics, exhibiting a minimum improvement of 20%. Its superiority is consistent under various conditions, including SNRs ranging from -14 to 2 dB and five contaminant types. An ablation study further proves that the design of TrustEMG-Net contributes to its optimality, providing high-quality sEMG and serving as an effective, robust, and generalized denoising solution for sEMG applications.

In recent decades, the rising aging population has increased the importance of fall detection (FD) systems as assistive technology. To enhance accuracy, deep learning techniques are widely applied to these systems. A typical FD system involves placing edge devices (EDs) on individuals to collect real-time data, which is transmitted to a cloud center (CC) or processed by the ED itself. However, this architecture faces challenges such as limited ED model size and data transmission latency to the CC. Mobile edge computing (MEC) has been explored to overcome these challenges, which allows computations at MEC servers deployed between EDs and the CC. A multi-layer mobile edge computing (MLMEC) framework is proposed to manage the trade-off between accuracy and latency. MLMEC divides the architecture into multiple stations, each equipped with a neural network model. If the front-end equipment lacks deterministic detection capability, data is transmitted to a station with robust back-end computing for detection. The knowledge distillation (KD) approach is employed to improve front-end detection accuracy. The KD approach allows high-power back-end stations to provide additional learning experiences, enhancing precision and reducing latency and processing load. Simulation results demonstrate improved accuracy with the KD approach (11.65% average improvement for the SisFall dataset and 2.78% for the FallAllD dataset). The proposed MLMEC with the KD approach outperforms without the KD approach regarding data latency rate by 54.15% for the FallAllD dataset and 46.67% for the SisFall dataset. In summary, the MLMEC FD system exhibits improved accuracy and reduced latency.

This work aims to explore the utility of wearable inertial measurement units (IMUs) for quantifying movement in Romberg tests and investigate the extent of movement in adults with vestibular hypofunction (VH). A cross-sectional study was conducted at an academic tertiary medical center between March 2021 and April 2022. Adults diagnosed with unilateral vestibular hypofunction (UVH) or bilateral vestibular hypofunction (BVH) were enrolled in the VH group. Healthy controls (HCs) were recruited from community or outpatient clinics. The IMU-based instrumented Romberg and tandem Romberg tests on the floor were applied to both groups. The primary outcomes were kinematic body metrics (maximum acceleration [ACC], mean ACC, root mean square [RMS] of ACC, and mean sway velocity [MV]) along the medio-lateral (ML), cranio-caudal (CC), and antero-posterior (AP) axes. A total of 31 VH participants (mean age, 33.48 [SD 7.68] years; 19 [61%] female) and 31 HCs (mean age, 30.65 [SD 5.89] years; 18 [58%] female) were recruited. During the eyes-closed portion of the Romberg test, VH participants demonstrated significantly higher maximum ACC and increased RMS of ACC in head movement, as well as higher maximum ACC in pelvic movement along the ML axis. In the same test condition, individuals with BVH exhibited notably higher maximum ACC and RMS of ACC along the ML axis in head and pelvic movements compared with HCs. Additionally, BVH participants exhibited markedly increased maximum ACC along the ML axis in head movement during the eyes-open portion of the tandem Romberg test. Conversely, no significant differences were found between UVH participants and HCs in the assessed parameters. The instrumented Romberg and tandem Romberg tests characterized the kinematic differences in head, pelvis, and ankle movement between VH and healthy adults. The findings suggest that these kinematic body metrics can be useful for screening BVH and can provide goals for vestibular rehabilitation.

The diagnostic accuracy of otitis media with effusion (OME) depends on a clinician’s experience and evaluation tools. Various assessment technologies have been applied to support clinical diagnosis, such as digital otoscopy and tympanometry. However, several challenges and issues limit the capabilities and usability of these assessment technologies, including high costs and needing to rely on specialists’ interpretations. In this work, we designed and validated OME detection using a machine learning model and in-ear microphones. Two off-the-shelf microphones were placed in the bilateral ear canals to record the voice when participants pronounced five 3-second sustained vowel sounds. Various signal processing and machine learning techniques were applied to the recordings, and the magnitude spectrograms of the vowel sound recording from in-ear microphones can distinguish ears with OME from healthy ears according to the differences in high-frequency response. Our results using in-ear microphones and ML algorithms had an accuracy of 80.65% in detecting OME, similar to that of typical OME detection approaches. This work demonstrates the potential to provide healthcare practitioners with a simple, safe, and more reliable expert-level diagnostic tool.

Falls are the public health issue for the elderly all over the world since the fall-induced injuries are associated with a large amount of healthcare cost. Falls can cause serious injuries, even leading to death if the elderly suffers a "long-lie". Hence, a reliable fall detection (FD) system is required to provide an emergency alarm for first aid. Due to the advances in wearable device technology and artificial intelligence, some fall detection systems have been developed using machine learning and deep learning methods to analyze the signal collected from accelerometer and gyroscopes. In order to achieve better fall detection performance, an ensemble model that combines a coarse-fine convolutional neural network and gated recurrent unit is proposed in this study. The parallel structure design used in this model restores the different grains of spatial characteristics and capture temporal dependencies for feature representation. This study applies the FallAllD public dataset to validate the reliability of the proposed model, which achieves a recall, precision, and F-score of 92.54%, 96.13%, and 94.26%, respectively. The results demonstrate the reliability of the proposed ensemble model in discriminating falls from daily living activities and its superior performance compared to the state-of-the-art convolutional neural network long short-term memory (CNN-LSTM) for FD.