Sensors

BackgroundParkinsons Disease (PD) is often reduced to its most visible motor symptoms, yet it is a systemic neurodegenerative disorder with a highly heterogeneous presentation. While cardinal motor signs such as bradykinesia, rigidity, and tremor arise from the loss of dopaminergic neurons in the substantia nigra, they typically manifest only after substantial neurodegeneration (approximately 50-70% loss) has already occurred, inevitably leading to delayed detection [1] PD significantly impacts non-motor and fine-motor domains as well that are frequently overlooked. Research indicates that hypokinetic dysarthria (voice impairment) affects approximately 89% of PD patients, often as an early prodromal sign [18]. Similarly, micrographia (handwriting impairment) is observed in up to 63% of cases, while non-motor symptoms such as hyposmia (loss of smell) and REM sleep behavior disorder occur in over 70% and 40% of patients, respectively--often years before clinical diagnosis [19, 20]. Consequently, diagnostic systems that rely on a single modality fail to capture this complexity, leading to missed detections in patients whose primary symptoms fall outside that specific domain. To address this, we propose a holistic, multimodal AI framework that explicitly targets these diverse pathological vectors--Voice, Gait, Handwriting, and Non-Motor Symptoms--to ensure robust and early detection across the full spectrum of the disease. MethodsWe propose a modular multimodal AI framework that integrates five complementary inputs: voice recordings, signals captured with the help of a smart pen during drawing spiral/meander, hand-drawn spiral/meander images, wearable sensor-driven gait data, and MDS-UPDRS questionnaire-derived symptom scores. Each modality undergoes an independent preprocessing and specialized modeling pipeline. Outputs from these specialized models are combined using a weighted aggregation engine, which allows for customizable contribution of each modality to the final classification. ResultsPreliminary experiments show that the unimodal pipelines achieved high accuracy, with the Random Forest (Voice) achieving 89%, XGBoost (Drawing Signal) up to 93%, and ResNet-18 (Drawing Image) up to 92%. Incorporating the Transformer model for gait data, which achieved 86% accuracy, significantly boosts the detection of subtle motor deficits. The proposed approach is expected to improve the overall diagnostic sensitivity and specificity relative to any unimodal baseline, offering transparent score breakdowns for clinical use. ConclusionThis study validates a comprehensive, multimodal Machine Learning framework designed to capture the holistic nature of clinical Parkinsons Disease. Our results indicate that fine motor control--analyzed through both dynamic handwriting signals and static imagery--serves as a highly discriminative biomarker, offering superior detection of subtle kinematic tremors. Furthermore, the integration of vocal analysis and spatiotemporal gait modeling ensures that the system captures the full spectrum of pathology, distinguishing between phonatory deficits and gross motor irregularities. By synthesizing these diverse clinical indicators, the proposed architecture overcomes the sensitivity limitations of single-modality systems, establishing a robust, non-invasive foundation for objective early screening and longitudinal patient monitoring in real-world settings.

This narrative review examines the theoretical foundations of mental workload, evaluates biometric monitoring methods, addresses ethical and privacy issues, and highlights future directions for longitudinal research. We focus on the integration of wearable sensors, multimodal data, artificial intelligence (AI), and machine learning (ML) frameworks to enhance adaptive task scheduling and safety in cognitively demanding professions. Continuous, real-time monitoring through wearable devices and multidimensional data analysis show promise for identifying and managing cognitive overload before it degrades performance. Despite this potential, significant challenges exist, including data protection, sensor reliability, calibration consistency, information processing, network capabilities, and variability in individuals responses. Physiological and behavioral measures as well as subjective and performance indicators offer valuable insights into the early signs of cognitive strain, suggesting that biometric monitoring could help organizations detect performance decline sooner. Evidence shows that these technologies are feasible in professions that require high precision, rapid decision-making, and sustained attention. However, only sparse longitudinal comparisons exist regarding the effectiveness of different biometric tools in real-world operational contexts, particularly with respect to data security and standardization. Integrating physiological and behavioral data with subjective assessments analyzed through AI and ML may enable early warning signs for overload in both individuals and teams working in high-stress, time-critical settings. Such approaches could inform work-recovery cycles, reduce error rates, and sustain cognitive performance. Further empirical research is necessary to confirm sensor accuracy in applied environments and to validate AI and ML predictions before large-scale deployment in sectors such as air traffic control, public safety, healthcare, and industrial operations. Graphical Abstract O_FIG O_LINKSMALLFIG WIDTH=200 HEIGHT=111 SRC="FIGDIR/small/25334668v1_ufig1.gif" ALT="Figure 1"> View larger version (45K): org.highwire.dtl.DTLVardef@b2c603org.highwire.dtl.DTLVardef@e61f87org.highwire.dtl.DTLVardef@1fee11org.highwire.dtl.DTLVardef@46d485_HPS_FORMAT_FIGEXP M_FIG C_FIG

The increasing use of accelerometers as digital health technologies in clinical trials and clinical care is driving the need for data processing to meet medical standards. The aim of this study was to create and test a modular pipeline for the pre-processing of high-resolution accelerometry that assures the quality, transparency and traceability of digital measures from sensor-level data. The objective is for the pipeline to be a foundational layer in the development, implementation and comparison of measures. The study developed the open GENEAcore package to meet the requirements of regulators, verifying the engineering implementation and analytically validating outputs against reference datasets. Early stages included the optimisation of calibration and non-wear detection. Data-driven detection of behavioural transitions was then validated to give direct bout outputs without the need to identify rules for epoch aggregation and interruptions. The utility for measure development was shown by comparing two algorithms for the characterisation of activity intensity in both the epoch and bout paradigms. Non-wear was detected with a balanced accuracy of 92.3% and the commonly used 13mg acceleration standard deviation threshold was empirically validated for the first time. The detection of transitions proved reliable with 99% detected, on average, within 2 seconds of their occurrence to give a mean expected event duration of 68.6s from a log-normal distribution. The different activity intensity algorithms were more than 99% concordant during movement but their outputs diverged in low movement conditions. Importantly, variable duration bouts created 31% higher daily activity durations compared to 1-second epochs. This evaluation of pre-processing steps has confirmed the attention to detail required to create robust and reproducible results for later clinical validation where small changes in an algorithm or its implementation may have clinically meaningful consequences.

The primary objective of this research is to improve the average classification performance for specific movements in patients with cervical spinal cord injury (SCI). The study utilizes a low-frequency multi-class electroencephalography (EEG) dataset obtained from the Institute of Neural Engineering at Graz University of Technology. The research combines convolutional neural network (CNN) and long-short-term memory (LSTM) architectures to uncover strong neural correlations between temporal and spatial aspects of the EEG signals associated with attempted arm and hand movements. To achieve this, three different methods are used to select relevant features, and the proposed models robustness against variations in the data is validated using 10-fold cross-validation (CV). Furthermore, the study explores the potential for subject-specific adaptation in an online paradigm, extending the proof-of-concept for classifying movement attempts. In summary, this research aims to make valuable contributions to the field of neuro-technology by developing EEG-controlled assistive devices using a generalized brain-computer interface (BCI) and deep learning (DL) framework. The focus is on capturing high-level spatiotemporal features and latent dependencies to enhance the performance and usability of EEG-based assistive technologies.

Peripheral oxygen saturation (SpO2) and heart rate (HR) are critical physiological measures that clinicians need to observe to decide on an emergency intervention. These measures are typically determined using a contact-based pulse oximeter. This approach may pose difficulties in many cases, such as with young children, patients with burnt or sensitive skin, cognitive impairments, and those undergoing certain medical procedures or severe illnesses. Remote Photoplethysmography (rPPG) allows for unobtrusive sensing of these vital signs in a variety of settings for health monitoring systems. Several research studies have been conducted to use rPPG for this purpose; however, there is still not a commercially available, clinically validated system that overcomes the concerns highlighted in this paper. We present a state-of-the-art review of rPPG-related research conducted including related processes and techniques, such as regions of interest (ROI) selection, extracting the raw signal, pre-processing data, applying noise reduction algorithms, Fast Fourier transforms (FFT), filtering and extracting these vital signs. Further, we present a detailed, critical evaluation of available rPPG systems. Limitations and future directions have also been identified to aid rPPG researchers in further advancing this field.

Existing clinical assessments for upper-limb motor rehabilitation post-stroke pose limitations as endpoints for clinical trials. This study aims to develop a wearable-based digital biomarker for assessing motor recovery using accelerometer data collected in naturalistic environments. The study analyzed approximately 23,000 hours of data from 215 participants, including subacute and chronic stroke survivors and healthy individuals. A novel analytical approach decomposed continuous accelerometer data into a lower-level unit of motor behaviors called movement segments, from which key features were extracted and aggregated using a linear mixed-effects model to produce a composite biomarker. The resulting digital biomarker demonstrated excellent interpretability, reliability, concurrent validity, discriminant validity, known-group validity, and responsiveness, enabling a nearly 66% reduction in the required sample size for clinical trials compared to traditional measures. These findings highlight its potential as a low-burden, scalable assessment tool for upper-limb motor recovery, with applications in both clinical trials and routine clinical practice.

Freezing of gait (FOG) is a debilitating symptom of Parkinsons disease (PD), characterized by an absence or reduction in forward movement of the legs despite the intention to walk. Detecting FOG during free-living conditions presents significant challenges, particularly when using only inertial measurement unit (IMU) data, as it must be distinguished from voluntary stopping events that also feature reduced forward movement. Influences from stress and anxiety, measurable through galvanic skin response (GSR) and electrocardiogram (ECG), may assist in distinguishing FOG from normal gait and stopping. However, no study has investigated the fusion of IMU, GSR, and ECG for FOG detection. Therefore, this study introduced two methods: a twostep approach that first identified reduced forward movement segments using a Transformer-based model with IMU data, followed by an XGBoost model classifying these segments as FOG or stopping using IMU, GSR, and ECG features; and an end-to-end approach employing a multi-stage temporal convolutional network to directly classify FOG and stopping segments from IMU, GSR, and ECG data. Results showed that the two-step approach with all data modalities achieved an average F1 score of 0.728 and F1@50 of 0.725, while the end-to-end approach scored 0.771 and 0.759, respectively. However, no significant difference was found compared to using only IMU data in both approaches (p-values: 0.466 to 0.887). In conclusion, adding physiological data does not provide a statistically significant benefit in distinguishing between FOG and stopping.

Oral cancer is a significant public-health issue and the existing methods of its detection are not as simple or fast as to be applicable by a wide population, particularly by those living in underserved communities. Our team has a proposed solution to address this problem by involving the concept of Convolutional Neural Networks (CNNs) to classify smartphone images into normal or malignant categorization in real time. We defined the model training to use a set of 1071 smartphone camera photos which were then pre-processed to convert them to HSV, normalize and resample the images. The CNN had an accuracy of 94.29 %, precision of 95.45%, recall/sensitivity of 93.33%, and F1-score of 94.38 after training. The overall predictive performance evaluation was calculated with an area under the receiver operating characteristic curve (AUC) of 0.99 with an average inference time of less than 5 sec so the clinicians or patients can send their images and get results in a short time. In contrast to other available methods, the EfficientNetB0 model is quicker and computationally less demanding, which is more suitable to be used on a mobile platform. The primary drawbacks which were big obstacles at the beginning of the project were the variance in image quality, the absence of annotated data, and changing the dataset to a larger and more diverse one, along with the application of advanced preprocessing enhanced the performance of models. The next step will be to focus on a large-scale clinical validation and additional model improvement. To sum up, the system is an AI-based method of scalable, cheap, and fast front-end screening that has a potential to significantly improve the outcomes of oral-cancer by early identification.

Detection and characterization of abnormalities of movement are important to develop a method for detecting early signs of Parkinsons disease (PD). Most of the current research in detection of characteristic reduction of movements due to PD, known as parkinsonism, requires using a set of invasive sensors in a clinical or controlled environment. Actigraphy has been widely used in medical research as a non-invasive data acquisition method in free-living conditions for long periods of time. The proposed algorithm uses triaxial accelerometer data obtained through actigraphy to detect walking bouts at least 10 seconds long and characterize them using cadence and arm swing. Accurate detection of walking periods is the first step toward the characterization of movement based on gait abnormalities. The algorithm was based on a Walking Score (WS) derived using the value of the auto-correlation function (ACF) for the Resultant acceleration vector. The algorithm achieved a precision of 0.90, recall of 0.77, and F1 score of 0.83 compared to the expert scoring for walking bout detection. We additionally described a method to measure arm swing amplitude.

Gait impairment is a hallmark symptom of Parkinsons disease (PD). However, traditional clinical assessments cannot capture real-world motor fluctuations, as they are sparsely performed. This study was designed to test and validate the use of nearables and passive sensing technologies, including Kinect RGB-D cameras and ultra-wideband (UWB) radar, for continuous, objective assessment of gait fluctuations in PD within a home-like setting. Fifteen PD patients with mild symptoms and fourteen age-and sex-matched healthy controls (HC) performed 4-meter walking tasks in a living lab facility. Patients repeated the task during both "ON" and "OFF" states of their daily medication cycle. Gait features, including stride length, stride time, and gait speed, were extracted from Kinect, radar, and a ground-truth smart floor. Data were analyzed to evaluate inter-sensor agreements and detect group-level differences. Stride time demonstrated the highest agreement between devices (r=0.903), while stride length showed weaker agreement (r=0.779), with Kinect tending to overestimate. Despite lower agreement, stride length from both Kinect and radar successfully distinguished PD OFF from HC (camera q=0.020; radar q=0.005) and radar was able to further differentiate ON and OFF states (q=0.020). Neither device differentiated PD ON from HC, indicating medication reduced observable gait differences. This study demonstrates that passive, contact-free sensing technologies such as depth cameras and UWB radars can effectively monitor gait in PD within naturalistic environments. While some spatial metrics, like stride length, show device discrepancies, both systems reliably capture gait patterns and medication-dependent changes, supporting their use for longitudinal, real-world monitoring of Parkinsons motor symptoms.

BackgroundFalls among elderly residents in assisted living facilities (ALFs) are prevalent, costly, and frequently under-documented. AUGi, a wall-mounted device employing obfuscated computer vision, deep learning, and mobile integration, provides continuous monitoring while maintaining patient privacy. ObjectiveTo evaluate whether the use of AI-assisted detection of falls could improve documentation of falls and lead to prevention of subsequent falls and related injury. MethodsAn ITS design analyzed monthly fall documentation data collected nine months pre-installation and four months post-installation of AUGi in ALFs. The primary outcome was the monthly documented fall rate, with and without injuries. Segmented regression analysis assessed changes in fall documentation trends related to AUGi installation. ResultsSegmented regression revealed no statistically significant immediate change in total falls post-AUGi installation (p = 0.85) nor a significant trend increase over time post-intervention (p = 0.17). Documented falls with injury also showed no significant immediate (p = 0.99) or trend differences (p = 0.73). Falls without injury similarly showed no immediate (p = 0.62) or trend changes (p = 0.82). Injury rate slightly declined (Cohens d = -0.54), though not significantly. The power analysis indicated low statistical power (13%), yet the moderate effect size suggests clinical relevance. DiscussionThe findings highlight pre-installation under-documentation of falls. Although statistical significance was not achieved, increased documentation post-AUGi installation suggests improved surveillance accuracy and potential for enhanced patient safety through more timely interventions. Future research should explore longer-term outcomes, including reduced injury severity and hospitalization, linked to improved fall documentation. ConclusionITS analysis indicates AUGi effectively enhances documentation of falls, suggesting improved patient safety monitoring in ALFs. Surveillance technologies may significantly improve documentation and decrease costs. Future studies could examine cost-benefits as well as potential to reduce documentation burden using ambient surveillance data.

Gait impairments in patients with Parkinsons Disease (PD) and Normal Pressure Hydrocephalus (NPH) are diagnosed with visual clinical assessments. Despite standardized gait tests and clinicians expertise, such approaches can be subjective and challenging due to similar symptoms between the two diseases. Wearable sensors and machine learning (ML) can assist clinicians by offering objective and quantitative assessments of gait impairments that can help distinguishing between PD and NPH. This study consists of a cohort of 12 PD and 11 NPH patients that performed standardized gait tests. Gait was measured by wearable sensors embedded in patients shoes: a three-axis gyroscope, a three-axis accelerometer and eight pressure sensors in each insole. Sensors and computational pipeline to extract gait cycle features were validated and calibrated on 21 healthy subjects. ML approaches were employed to identify changes in gait cycle features between the PD and NPH patients groups. Twenty-seven ML classifiers were compared, leading to select linear support vector machines, resulting in a classification accuracy of 0.70 {+/-} 0.28 and an area under the ROC curve of 0.74 {+/-} 0.39. Combining wearable sensors with ML algorithms trained on gait cycle features from those sensors showed the potential for objective differentiation of gait patterns between PD and NPH patients.

Continuous and automatic monitoring of an individuals physical activity using wearable devices provides valuable insights into their daily habits and patterns. This information can be used to promote healthier lifestyles, prevent chronic diseases, and improve overall well-being. Smart glasses are an emerging technology that can be worn comfortably and continuously. Their wearable nature and hands-free operation make them well suited for long-term monitoring of physical activity and other real-world applications. To this end, we investigated the ability of the multi-sensor OCOsense smart glasses to recognize everyday activities. We evaluated three end-to-end deep learning architectures that showed promising results when working with IMU (accelerometer, gyroscope, and magnetometer) data in the past. The data used in the experiments was collected from 18 participants who performed pre-defined activities while wearing the glasses. The best architecture achieved an F1 score of 0.81, demonstrating its ability to effectively recognize activities, with the most problematic categories being standing vs. sitting.

Stress can adversely impact health, leading to issues like high blood pressure, heart diseases, and a compromised immune system. Consequently, using wearable devices to monitor stress is essential for prompt intervention and effective management. This study investigates the efficacy of wearable devices in the early detection of psychological stress, employing both binary and five-class classification models. Significant correlations were observed between stress levels and physiological signals, including Electrocardiogram (ECG), Electrodermal Activity (EDA), and Respiration (RESP), establishing these modalities as reliable biomarkers for stress detection. Utilizing the publicly available Wearable Stress and Affect Detection (WESAD) dataset, we employed two ensemble methods, Majority Voting (MV) and Weighted Averaging (WA), to integrate these signals, achieving maximum accuracies of 99.96% for binary classification and 99.59% for five-class classification. This integration significantly enhances the accuracy and robustness of the stress detection system. Furthermore, ten different classifiers were evaluated, and hyperparameter optimization and K-fold cross-validation ranging from 3-fold to 10-fold were applied. Both time-domain and frequency-domain features were examined separately. A review of commercially available wearable devices supporting these modalities was also conducted, resulting in recommendations for optimal configurations for practical applications. Our findings highlight the potential of multimodal wearable devices in advancing the early detection and continuous monitoring of psychological stress, with significant implications for future research and the development of improved stress detection systems.

Behavioral sensing using wearables has emerged as a valuable tool for screening of neurodegenerative conditions, including Mild Cognitive Impairment (MCI). Existing research has predominantly focused on using wearables for quantifying walking patterns in individuals with MCI, typically within controlled environments. On the other hand, the human activity recognition community has been actively studying to quantify kitchen activities, which is an instrumental activity of daily living. Previous studies reported deficits in visuospatial navigation in individuals living with MCI, which affects functional independence within the kitchen environment for these populations. This study investigates the use of wrist and eye-tracking wearable sensors to quantify kitchen activities in individuals with MCI. We collected multimodal datasets from 19 older adults (11 with MCI and 8 with normal cognition) while preparing a yogurt bowl. Our multimodal analysis model could classify older adults with MCI from normal cognition with a 74% F1 score. The feature importance analysis showed the association of weaker upper limb motor function and delayed eye movements with cognitive decline, consistent with previous findings in MCI research. This pilot study demonstrates the feasibility of monitoring behavior markers of MCI in daily living settings, which calls for further studies with larger-scale validation in individuals home environments. ACM Reference FormatBonwoo Koo, Ibrahim Bilau, Amy D. Rodriguez, Eunhwa Yang, and Hyeokhyen Kwon. 2025. Quantifying Mild Cognitive Impairments in Older Adults Using Multi-modal Wearable Sensor Data in a Kitchen Environment. In Proceedings of ACM International Symposium on Wearable Computers (ISWC 25). ACM, New York, NY, USA, 9 pages. https://doi.org/XXXXXXX.XXXXXXX

High movement velocities are among the primary risk factors for work-related musculoskeletal disorders (MSDs). Ergonomists have commonly used two methods to calculate angular movement velocities of the upper arms using inertial measurement units (accelerometers and gyroscopes). Generalized velocity is the speed of movement traveled on the unit sphere per unit time. Inclination velocity is the derivative of the postural inclination angle relative to gravity with respect to time. Neither method captures the full extent of upper arm angular velocity. We propose a new method, the gyroscope vector magnitude (GVM), and demonstrate how GVM captures angular velocities around all motion axes and more accurately represents the true angular velocities of the upper arm. We use optical motion capture data to demonstrate that the previous methods for calculating angular velocities capture 89% and 77% relative to our proposed method. We propose GVM as the standard metric for reporting angular arm velocities in future research.

While the need for continuous blood pressure (BP) monitoring in Japan is high, there are no commercially available cuffless devices for personal daily monitoring use. Fingertip-based sensors are a promising alternative as they eliminate the discomfort of repeated cuff inflation. However, their reliability during winter has been a major technical limitation due to cold-induced peripheral vasoconstriction. This study aimed to address this issue by validating a novel fingertip-based continuous BP monitor used by exercising adults during summer and winter. Eleven community-dwelling older adults (mean age, 73.1 {+/-} 8.8 years) were included in this seasonal comparative study. During exercise, we compared a personal fingertip-based continuous monitor (ArteVu) with a standard oscillometric cuff device (Omron) in summer (mean, 26.5{degrees}C) and winter (mean, 7.4{degrees}C). The study also evaluated the device's accuracy during exercise-induced BP fluctuations and seasonal environmental changes. Awareness of the participants regarding BP management was also assessed using questionnaires. There were strong correlations for systolic BP (SBP) between summer and winter (r = 0.93 in summer; r = 0.88 in winter). Although the mean difference for the SBP was higher in winter than in summer (3.1 {+/-} 11.2 mmHg vs. 0.2 {+/-} 9.4 mmHg), the values remained within a clinically acceptable range for personal monitoring. Notably, 72.7% of participants reported that the ease of using the fingertip-based device significantly increased their awareness and motivation for daily BP management. This study confirms the feasibility of cuffless fingertip-based continuous BP monitoring across different seasons, including in winter. By overcoming the seasonal limitations, this device fills a critical gap in the Japanese health-monitoring market. Our findings support the development of smaller and more portable models, representing a shift from traditional "snapshot" cuff measurements to continuous and integrated lifestyle monitoring for older adults.

Parkinsons disease (PD) is one of the most common neurodegenerative disorders of the central nervous system, which predominantly affects patients motor functions, movement, and stability. Monitoring movement in patients with PD is crucial for inferring motor state fluctuations throughout daily life activities, which aids in disease progression analysis and assessing how patients respond to medications over time. In recent years, there has been an increase in the usage of wearable sensors for PD symptom monitoring. In this study, we present a preliminary analysis of smart glasses equipped with IMU sensors to provide objective information on the motor state in patients with PD. Data were collected from seven Parkinsons patients with varying levels of symptom severity. The patients performed the Timed-Up-and-Go (TUG) Test while wearing IMU-equipped glasses. Our analysis indicates that smart glasses can provide information about patients gait that can be used to assess the severity level of the PD as measured by two standardized questionnaires. Furthermore, patient-specific clusters can be easily detected in the sensor data, hinting at the development of personalized models for patient-specific monitoring of symptom progression. Therefore, smart glasses have the potential to be used as an unobtrusive and continuous screening tool for PD patients gait, enhancing medical assessment and treatment. CCS CONCEPTS * Applied computing * Life and medical sciences * Health informatics ACM Reference FormatFirst Authors Name, Initials, and Last Name, Second Authors Name, Initials, and Last Name, and Third Authors Name, Initials, and Last Name. 2022. The Title of the Paper: ACM Conference Proceedings Manuscript Submission Template: This is the subtitle of the paper, this document both explains and embodies the submission format for authors using Word. In Woodstock 18: ACM Symposium on Neural Gaze Detection, June 03-05, 2018, Woodstock, NY. ACM, New York, NY, USA, 10 pages. NOTE: This block will be automatically generated when manuscripts are processed after acceptance.

This article introduces the Emteqs OCOsense smart glasses equipped with a novel non-contact OCO sensor technology for measuring facial muscle activation and expressions based on high resolution tracking of skin movement. We demonstrate that the OCO sensor technology based on optomyography is a sensitive and accurate approach for assessing skin movement in 3 dimensions, providing a means for measuring the facial expressions used to assess emotional valence such as smile, frown, and eyebrow raise. We propose that glasses-based optomyography sensing has the potential to herald a paradigm shift in real-world facial expression monitoring, thus enabling real-time emotional analytics with healthcare and research applications.

Cardiorespiratory fitness (CRF) declines as age increases in elderly. An individualized CRF exercise prescription can maintain the CRF level and delay aging process. Traditional exercise prescriptions are general and lack of individualization. In this paper, a new study based on back-propagation (BP) neural network, is investigated to predict the individualized CRF exercise prescriptions for elderly by correlate variables (age, sex, BMI, VO2max initial value, improvement etc.). The raw data are split to two parts, 90% for training the machine and the remaining 10% for testing the performance. Based on a database with 2078 people, the exercise prescription prediction models MAE, RMSE and R2 are1.5206,1.4383 and 0.9944. 26 female subjects aged 60-79 years are recruited to test the models validity. The VO2maxs expected improvement was set at 10%. Based on the basic information of these elder women, we get personalized exercise prescription (frequency, intensity, time and volume) of each subject. All of them finished their own exercise intervention. The results show that the post VO2max was significantly different from the pre VO2max and improved by 10.1%, and a total of 20 subjects(74.1%) improved within one standard deviation and 25 subjects(92.6%)improved within 1.96 times standard deviations. Our study shows that a high degree of accuracy in exercise suggestions for elderly was achieved by applying the BP neural network model.