Machine Learning Methods for Fear Classification Based on Physiological Features

Strategies for Reliable Stress Recognition: A Machine Learning Approach Using Heart Rate Variability Features

Stress recognition, particularly using machine learning (ML) with physiological data such as heart rate variability (HRV), holds promise for mental health interventions. However, limited datasets in affective computing and healthcare research can lead to inaccurate conclusions regarding the ML model performance. This study employed supervised learning algorithms to classify stress and relaxation states using HRV measures. To account for limitations associated with small datasets, robust strategies were implemented based on methodological recommendations for ML with a limited dataset, including data segmentation, feature selection, and model evaluation. Our findings highlight that the random forest model achieved the best performance in distinguishing stress from non-stress states. Notably, it showed higher performance in identifying stress from relaxation (F1-score: 86.3%) compared to neutral states (F1-score: 65.8%). Additionally, the model demonstrated generalizability when tested on independent secondary datasets, showcasing its ability to distinguish between stress and relaxation states. While our performance metrics might be lower than some previous studies, this likely reflects our focus on robust methodologies to enhance the generalizability and interpretability of ML models, which are crucial for real-world applications with limited datasets.

Effective relax acquisition: a novel approach to classify relaxed state in alpha band EEG-based transformation

A relaxed state is essential for effective hypnotherapy, a crucial component of mental health treatments. During hypnotherapy sessions, neurologists rely on the patient's relaxed state to introduce positive suggestions. While EEG is a widely recognized method for detecting human emotions, analyzing EEG data presents challenges due to its multi-channel, multi-band nature, leading to high-dimensional data. Furthermore, determining the onset of relaxation remains challenging for neurologists. This paper presents the Effective Relax Acquisition (ERA) method designed to identify the beginning of a relaxed state. ERA employs sub-band sampling within the Alpha band for the frequency domain and segments the data into four-period groups for the time domain analysis. Data enhancement strategies include using Window Length (WL) and Overlapping Shifting Windows (OSW) scenarios. Dimensionality reduction is achieved through Principal Component Analysis (PCA) by prioritizing the most significant eigenvector values. Our experimental results indicate that the relaxed state is predominantly observable in the high Alpha sub-band, particularly within the fourth period group. The ERA demonstrates high accuracy with a WL of 3 s and OSW of 0.25 s using the KNN classifier (90.63%). These findings validate the effectiveness of ERA in accurately identifying relaxed states while managing the complexity of EEG data.

Racial and ethnic differences in predictors of participation in an intergenerational social connectedness intervention for older adults

BACKGROUND

Social connectedness is a key determinant of health and interventions have been developed to prevent social isolation in older adults. However, these interventions have historically had a low participation rate amongst minority populations. Given the sustained isolation caused by the COVID-19 pandemic, it is even more important to understand what factors are associated with an individual's decision to participate in a social intervention. To achieve this, we used machine learning techniques to model the racial and ethnic differences in participation in social connectedness interventions.

METHODS

Data were obtained from a social connectedness intervention that paired college students with Houston-area community-dwelling older adults (> 65 yo) enrolled in Medicare Advantage plans. Eligible participants were contacted telephonically and asked to complete the 3-item UCLA Loneliness Scale. We used the following machine-learning methods to identify significant predictors of participation in the program: k-nearest neighbors, logistic regression, decision tree, gradient-boosted decision tree, and random forest.

RESULTS

The gradient-boosted decision tree models yielded the best parameters for all race/ethnicity groups (96.1% test accuracy, 0.739 AUROC). Among non-Hispanic White older adults, key features of the predictive model included Functional Comorbidity Index (FCI) score, Medicare prescription risk score, Medicare risk score, and depression and anxiety indicators within the FCI. Among non-Hispanic Black older adults, key features included disability, Medicare prescription risk score, FCI and Medicare risk scores. Among Hispanic older adults, key features included depression, FCI and Medicare risk scores.

CONCLUSIONS

These findings offer a substantial opportunity for the design of interventions that maximize engagement among minority groups at greater risk for adverse health outcomes.

Wearable Devices and Explainable Unsupervised Learning for COVID-19 Detection and Monitoring

Despite the declining COVID-19 cases, global healthcare systems still face significant challenges due to ongoing infections, especially among fully vaccinated individuals, including adolescents and young adults (AYA). To tackle this issue, cost-effective alternatives utilizing technologies like Artificial Intelligence (AI) and wearable devices have emerged for disease screening, diagnosis, and monitoring. However, many AI solutions in this context heavily rely on supervised learning techniques, which pose challenges such as human labeling reliability and time-consuming data annotation. In this study, we propose an innovative unsupervised framework that leverages smartwatch data to detect and monitor COVID-19 infections. We utilize longitudinal data, including heart rate (HR), heart rate variability (HRV), and physical activity measured via step count, collected through the continuous monitoring of volunteers. Our goal is to offer effective and affordable solutions for COVID-19 detection and monitoring. Our unsupervised framework employs interpretable clusters of normal and abnormal measures, facilitating disease progression detection. Additionally, we enhance result interpretation by leveraging the language model Davinci GPT-3 to gain deeper insights into the underlying data patterns and relationships. Our results demonstrate the effectiveness of unsupervised learning, achieving a Silhouette score of 0.55. Furthermore, validation using supervised learning techniques yields high accuracy (0.884 ± 0.005), precision (0.80 ± 0.112), and recall (0.817 ± 0.037). These promising findings indicate the potential of unsupervised techniques for identifying inflammatory markers, contributing to the development of efficient and reliable COVID-19 detection and monitoring methods. Our study shows the capabilities of AI and wearables, reflecting the pursuit of low-cost, accessible solutions for addressing health challenges related to inflammatory diseases, thereby opening new avenues for scalable and widely applicable health monitoring solutions.

Multi-Input CNN-LSTM deep learning model for fear level classification based on EEG and peripheral physiological signals

Objective and accurate classification of fear levels is a socially important task that contributes to developing treatments for Anxiety Disorder, Obsessive-compulsive Disorder, Post-Traumatic Stress Disorder (PTSD), and Phobia. This study examines a deep learning model to automatically estimate human fear levels with high accuracy using multichannel EEG signals and multimodal peripheral physiological signals in the DEAP dataset. The Multi-Input CNN-LSTM classification model combining Convolutional Neural Network (CNN) and Long Sort-Term Memory (LSTM) estimated four fear levels with an accuracy of 98.79% and an F1 score of 99.01% in a 10-fold cross-validation. This study contributes to the following; (1) to present the possibility of recognizing fear emotion with high accuracy using a deep learning model from physiological signals without arbitrary feature extraction or feature selection, (2) to investigate effective deep learning model structures for high-accuracy fear recognition and to propose Multi-Input CNN-LSTM, and (3) to examine the model's tolerance to individual differences in physiological signals and the possibility of improving accuracy through additional learning.

Quantitative identification of ventral/dorsal nerves through intraoperative neurophysiological monitoring by supervised machine learning

Objective

This study aimed to investigate the electro-neurophysiological characteristics of the ventral and dorsal nerves at the L2 segment in a quantitative manner.

Methods

Medical records of consecutive patients who underwent single-level approach selective dorsal rhizotomy (SDR) from June 2019 to January 2022 were retrospectively reviewed. Intraoperative electro-neurophysiological data were analyzed.

Results

A total of 74 males and 27 females were included in the current study with a mean age of 6.2 years old. Quadriceps and adductors were two main muscle groups innervated by L2 nerve roots in both ventral and dorsal nerve roots. Dorsal roots have a higher threshold than that of the ventral ones, and muscles that first reached 200 µV innervated by dorsal roots have longer latency and smaller compound muscle action potential (CMAP) than those of the ventral ones. Supervised machine learning can efficiently distinguish ventral/dorsal roots using threshold + latency or threshold + CMAP as predictors.

Conclusion

Electro-neurophysiological parameters could be used to efficiently differentiate ventral/dorsal fibers during SDR.

Evaluating Predictors of Participation in Telephone-Based Social-Connectedness Interventions for Older Adults: A Dual Machine-Learning and Regression Approach

Social isolation is a well-documented contributor to poor mental and physical health, and interventions promoting social connectedness have been associated with various health benefits. This study examined predictors of participation in a telephone-based social connectedness intervention for socially isolated older adults. Data were obtained from a social-connectedness intervention that paired college students with Houston-area, community-dwelling adults aged 65 years and older and enrolled in Medicare Advantage plans. We combined machine learning and regression techniques to identify significant predictors of program participation. The following machine-learning methods were implemented: (1) k-nearest neighbors, (2) decision tree and ensembles of decision trees, (3) gradient-boosted decision tree, and (4) random forest. The primary outcome was a binary flag indicating participation in the telephone-based social-connectedness intervention. The most predictive variables in the ML models, with scores corresponding to the 90th percentile or greater, were included in the regression analysis. The predictive ability of each model showed high discriminative power, with test accuracies greater than 95%. Our findings suggest that telephone-based social-connectedness interventions appeal to individuals with disabilities, depression, arthritis, and higher risk scores. scores. Recognizing features that predict participation in social-connectedness programs is the first step to increasing reach and fostering patient engagement.

Analysis of Marine-Pilot Biometric Data Recordings during Port-Approach Using a Full-Mission Simulator

The purpose of this study is to analyse data from the marine pilots’ bio-sensor readings to determine how experience affects their biometrical response during the port approach. The experiences play a significant role in the participant’s decision-making process and correlate with the repetitions. Through the repetitions of the experimental task, the participants gain experience, which correlates with the biometrical response, e.g., heart rate, electrodermal activity, etc. After exposing the two experience-distinct groups of participants to the same simulated port-approaching task, their collected biometric data is analysed and discussed. The results show that biometrical readings of the less experienced participants typically vary compared to that of the experienced participants, who take the simulated task more seriously. The study also yields insight into the workload process, involving disturbing factors during the task.