ViTAL Lab Computational Behavior and Health Analytics

Research Projects

Human Activity Recognition
Smart Hospital with Ubiquitous AI
Telehealth with Fair AI
Health Analytics with XAI

Generative AI and Cross-modality Techniques for Human Activity Recognition

Opportunistic Use of Video Data for Wearable-based Human Activity Recognition

figure image

Building human activity recognition model involves training a machine learning model to classify the given sensor signal input for target activities. The performance of the supervised learning approach largely depends on the scale of the training dataset that covers diverse activity patterns in the wild. However, collecting wearable sensor data from users is expensive, and annotation is time-consuming. Thereby, wearable datasets are typically limited in scale. To overcome the challenges in collecting wearable datasets, We proposed a method that uses a virtually unlimited amount of video data in online repositories to generate a training sensor dataset. The proposed technique uses state-of-the-art computer vision algorithms to track full human motions from human activity video and extract virtual inertial measurement signals from the on-body locations. The collection of virtual sensors from target activity videos is used to train the human activity recognition model, which is deployed in the real world. The proposed technique opens up the opportunity to apply complex analysis models for human activity recognition with the availability of large-scale training data

Large Language Model and Generative Motion Model for Wearable-based Human Activity Recognition

figure image

The development of robust, generalized models in human activity recognition (HAR) has been hindered by the scarcity of large-scale, labeled data sets. Recent work has shown that virtual IMU data extracted from videos using computer vision techniques can lead to substantial performance improvements when training HAR models combined with small portions of real IMU data. Inspired by recent advances in motion synthesis from textual descriptions and connecting Large Language Models (LLMs) to various AI models, we introduce an automated pipeline that first uses ChatGPT to generate diverse textual descriptions of activities. These textual descriptions are then used to generate 3D human motion sequences via a motion synthesis model, T2M-GPT, and later converted to streams of virtual IMU data. We benchmarked our approach on three HAR datasets (RealWorld, PAMAP2, and USC-HAD) and demonstrate that the use of virtual IMU training data generated using our new approach leads to significantly improved HAR model performance compared to only using real IMU data. Our approach contributes to the growing field of cross-modality transfer methods and illustrate how HAR models can be improved through the generation of virtual training data that do not require any manual effort.

Robust Human Activity Recognition Model with Uncertainty Modeling

figure image

Building human activity recognition model involves training a machine learning model to classify the given sensor signal feature for target activities. For the conventional supervised learning approach, discriminative sensor feature representation and accurate activity annotations are necessary to learn the robust human activity recognition model. However, acquiring accurate activity annotation is intractable due to the continuous nature of human activities. Thereby, the activity annotations have uncertainty regarding the correct temporal alignment of activity boundaries and the correctness of the actual label. We explore novel machine learning techniques to handle such uncertainties.

Smart Hospital using Edge Computing and Machine Learning Framework with Multi-modal Ambient, Mobile, and Wearable Sensors

Low-cost, Passive, and Continuous Patient Monitoring System Using Artificial Intelligence, Distributed Multi-view Camera System, and Wearables.

figure image

Neurological disorders, like Alzhemer’s Disease or related dementia (ADRD), can influence an individual’s mobility and social interactions, offering insights into disease progression. Yet, accurately measuring intricate movements and social interactions in real-world settings is challenging, making clinicians rely on retrospective surveys. Addressing this challenge, I’ve introduced a cost-effective, passive, continuous patient-monitoring system capable of quantifying patients’ social behaviors and movements within therapeutic facilities spanning 18,000 sqft. This system integrates multiple edge computing devices, combined with AI, cameras, and wearable sensors, to localize, track, and identify group activities. Specifically, it’s applied to quantify Mild Cognitive Impairment (MCI), a preceding condition of ADRD. With minimal modification to the existing infrastructure, this system can transform standard therapeutic areas into smart environments that can evaluate conditions linked to neurological disorders at a low cost.

Intelligent Telehealth Platforms with Fair Artificial Intelligence and Cloud Computing

Multi-modal Machine Learning System for Assessing Remote Interviews

figure image

This research addresses the challenges in current psychiatric evaluations, which often suffer from subjectivity, bias, and the need for highly skilled professionals. The study explores the use of objective digital biomarkers, derived from behavioral and physiological signals in remote interviews, for psychiatric disorder assessment. These signals include facial and vocal expressions, linguistic content, and cardiovascular modulation. The study used a multimodal approach, evaluating each modality individually and in combination, to assess mental health status. This approach aims to make mental health assessments more objective, remote, and cost-effective. The research also critically examines the fairness of these automated mental health assessment tools across different demographics (race, gender, education level, and age). Despite their growing importance, especially in telehealth, these tools can carry biases that may lead to unfair treatment of certain groups. The study systematically evaluated and discussed the need for reporting and mitigating these biases to build trust and ensure appropriate application in clinical settings. Using a developed multimodal mental health assessment system, the study analyzed how various features (facial expressions, voice acoustics, linguistic expressions, and cardiovascular patterns) affected fairness across demographics. The findings indicated that no single modality was consistently fair for all groups. Although methods to mitigate unfairness improved fairness levels, they also highlighted a trade-off between performance and fairness. This calls for a broader moral discussion and further investigation into the ethics and application of automated mental health assessment tools.

Behavior Analytics for Health Assessments

Explainable Artificial Intelligence for Movement Disorder Assessment with Kinematics Dataset

figure image

Traditional methods for diagnosing and assessing Parkinson’s disease (PD) often involve thorough physical examinations by specialists in movement disorders. These methods, including the use of scales like MDS-UPDRS or patient questionnaires (N-FOG-Q), can be subjective and vary between different raters. Particularly challenging is assessing Freezing of Gait (FOG), a complex aspect of parkinsonism that significantly affects quality of life. PD is also known to manifest in different motor subtypes, mainly tremor dominant (TD) and postural instability/gait difficulty (PIGD), each presenting unique challenges in balance and gait. Previous studies have employed machine learning to analyze kinematic data from body movements during gait assessments, but these often lacked the detail necessary to fully capture the complexity of whole-body movements in PD. To address this, a new method using explainable artificial intelligence models has been developed. These models can objectively and accurately measure PD, mirroring the assessment capabilities of human specialists. They show promise in accurately identifying FOG severity levels and PD subtypes, thereby aiding in the early identification of motor subtypes and predicting disease progression. Moreover, these models can potentially discover new kinematic biomarkers for PD, offering valuable insights for clinical trials and hypothesis generation.

Mental Health Assessment with Remote Interviews

Refer “Multi-modal AI System”

Autism Spectrum Disorder

figure image

Rapid identification and accurate documentation of interfering and high-risk behaviors in ASD, such as aggression, self-injury, disruption, and restricted repetitive behaviors, are important in daily classroom environments for tracking intervention effectiveness and allocating appropriate resources to manage care needs. However, having a staff dedicated solely to observing is costly and uncommon in most educational settings. Recently, multiple research studies have explored developing automated, continuous, and objective tools using machine learning models to quantify behaviors in ASD. However, the majority of the work was conducted under a controlled environment and has not been validated for real-world conditions. In this work, we demonstrate that the latest advances in video-based group activity recognition techniques can quantify behaviors in ASD in real-world activities in classroom environments while preserving privacy. Our explainable model could detect the episode of problem behaviors with a 77% F1-score and capture distinctive behavior features in different types of behaviors in ASD. To the best of our knowledge, this is the first work that shows the promise of objectively quantifying behaviors in ASD in a real-world environment, which is an important step toward the development of a practical tool that can ease the burden of data collection for classroom staff.

Cognitive Impairment

figure image

Mild cognitive impairment (MCI) is characterized by a decline in cognitive functions beyond typical age and education-related expectations. Since, MCI has been linked to reduced social interactions and increased aimless movements, we aimed to automate the capture of these behaviors to enhance longitudinal monitoring. Using a privacy-preserving distributed camera network, we collected movement and social interaction data from groups of individuals with MCI undergoing therapy within a 1700$m^2$ space. We developed movement and social interaction features, which were then used to train a series of machine learning algorithms to distinguish between higher and lower cognitive functioning MCI groups. A Wilcoxon rank-sum test revealed statistically significant differences between high and low-functioning cohorts in features such as linear path length, walking speed, change in direction while walking, entropy of velocity and direction change, and number of group formations in the indoor space. Despite lacking individual identifiers to associate with specific levels of MCI, a machine learning approach using the most significant features provided a 71\% accuracy.
We provide evidence to show that a privacy-preserving low-cost camera network using edge computing framework has the potential to distinguish between different levels of cognitive impairment from the movements and social interactions captured during group activities.