Explainable AI for clinical and remote health applications: a survey on tabular and time series data

Artificial Intelligence Review, Journal Year: 2022, Volume and Issue: 56(6), P. 5261 - 5315

Published: Oct. 26, 2022

Nowadays Artificial Intelligence (AI) has become a fundamental component of healthcare applications, both clinical and remote, but the best performing AI systems are often too complex to be self-explaining. Explainable (XAI) techniques defined unveil reasoning behind system's predictions decisions, they even more critical when dealing with sensitive personal health data. It is worth noting that XAI not gathered same attention across different research areas data types, especially in healthcare. In particular, many remote applications based on tabular time series data, respectively, commonly analysed these while computer vision Natural Language Processing (NLP) reference applications. To provide an overview methods most suitable for domain, this paper provides review literature last 5 years, illustrating type generated explanations efforts provided evaluate their relevance quality. Specifically, we identify validation, consistency assessment, objective standardised quality evaluation, human-centered assessment as key features ensure effective end users. Finally, highlight main challenges field well limitations existing methods.

Language: Английский

---

Yoga Meets Intelligent Internet of Things: Recent Challenges and Future Directions

Bioengineering, Journal Year: 2023, Volume and Issue: 10(4), P. 459 - 459

Published: April 9, 2023

The physical and mental health of people can be enhanced through yoga, an excellent form exercise. As part the breathing procedure, yoga involves stretching body organs. guidance monitoring are crucial to ripe full benefits it, as wrong postures possess multiple antagonistic effects, including hazards stroke. detection possible with Intelligent Internet Things (IIoT), which is integration intelligent approaches (machine learning) (IoT). Considering increment in practitioners recent years, IIoT has led successful implementation IIoT-based training systems. This paper provides a comprehensive survey on integrating IIoT. also discusses types procedure for using Additionally, this highlights various applications safety measures, challenges, future directions. latest developments findings its

Language: Английский

---

A multi-modal deep learning approach for stress detection using physiological signals: integrating time and frequency domain features

Frontiers in Physiology, Journal Year: 2025, Volume and Issue: 16

Published: April 1, 2025

This study aims to develop a multimodal deep learning-based stress detection method (MMFD-SD) using intermittently collected physiological signals from wearable devices, including accelerometer data, electrodermal activity (EDA), heart rate (HR), and skin temperature. Given the unique demands high-intensity work environment of nursing profession, measurement in nurses serves as representative case, reflecting levels other high-pressure occupations. We propose learning framework that integrates time-domain frequency-domain features for detection. To enhance model robustness generalization, data augmentation techniques such sliding window jittering are applied. Feature extraction includes statistical derived raw obtained via Fast Fourier Transform (FFT). A customized architecture employs convolutional neural networks (CNNs) process separately, followed by fully connected layers final classification. address class imbalance, Synthetic Minority Over-sampling Technique (SMOTE) is utilized. The trained evaluated on signal dataset with level labels. Experimental results demonstrate MMFD-SD achieves outstanding performance detection, an accuracy 91.00% F1-score 0.91. Compared traditional machine classifiers logistic regression, random forest, XGBoost, proposed significantly improves both robustness. Ablation studies reveal integration plays crucial role enhancing performance. Additionally, sensitivity analysis confirms model's stability adaptability across different hyperparameter settings. provides accurate robust approach integrating features. Designed occupational environments intermittent collection, it effectively addresses real-world monitoring challenges. Future research can explore fusion additional modalities, real-time improvements generalization its practical applicability.

Language: Английский

---

SAMoSA

Proceedings of the ACM on Interactive Mobile Wearable and Ubiquitous Technologies, Journal Year: 2022, Volume and Issue: 6(3), P. 1 - 19

Published: Sept. 6, 2022

Despite advances in audio- and motion-based human activity recognition (HAR) systems, a practical, power-efficient, privacy-sensitive system has remained elusive. State-of-the-art systems often require power-hungry privacy-invasive audio data. This is especially challenging for resource-constrained wearables, such as smartwatches. To counter the need an always-on audio-based classification system, we first make use of power compute-optimized IMUs sampled at 50 Hz to act trigger detecting events. Once detected, multimodal deep learning model that augments motion data with captured on smartwatch. We subsample this rates ≤ 1 kHz, rendering spoken content unintelligible, while also reducing consumption mobile devices. Our achieves accuracy 92.2% across 26 daily activities four indoor environments. findings show subsampling from 16 kHz down concert data, does not result significant drop inference accuracy. analyze speech intelligibility requirements less than demonstrate our proposed approach can improve practicality systems.

Language: Английский

---

Cohort comfort models — Using occupant’s similarity to predict personal thermal preference with less data

Building and Environment, Journal Year: 2022, Volume and Issue: 227, P. 109685 - 109685

Published: Oct. 13, 2022

Language: Английский

---

Subject-Independent Deep Architecture for EEG-Based Motor Imagery Classification

IEEE Transactions on Neural Systems and Rehabilitation Engineering, Journal Year: 2024, Volume and Issue: 32, P. 718 - 727

Published: Jan. 1, 2024

Motor imagery (MI) classification based on electroencephalogram (EEG) is a widely-used technique in non-invasive brain-computer interface (BCI) systems. Since EEG recordings suffer from heterogeneity across subjects and labeled data insufficiency, designing classifier that performs the MI independently subject with limited samples would be desirable. To overcome these limitations, we propose novel subject-independent semi-supervised deep architecture (SSDA). The proposed SSDA consists of two parts: an unsupervised supervised element. training set contains both unlabeled multiple subjects. First, part, known as columnar spatiotemporal auto-encoder (CST-AE), extracts latent features all by maximizing similarity between original reconstructed data. A dimensional scaling approach employed to reduce dimensionality representations while preserving their discriminability. Second, part learns using acquired part. Moreover, employ center loss minimize embedding space distance each point class its center. model optimizes parts network end-to-end fashion. performance evaluated test who were not seen during phase. assess performance, use benchmark EEG-based task datasets. results demonstrate outperforms state-of-the-art methods small number can sufficient for strong performance.

Language: Английский

---

Personalized Stress Detection Using Biosignals from Wearables: A Scoping Review

Sensors, Journal Year: 2024, Volume and Issue: 24(10), P. 3221 - 3221

Published: May 18, 2024

Stress is a natural yet potentially harmful aspect of human life, necessitating effective management, particularly during overwhelming experiences. This paper presents scoping review personalized stress detection models using wearable technology. Employing the PRISMA-ScR framework for rigorous methodological structuring, we systematically analyzed literature from key databases including Scopus, IEEE Xplore, and PubMed. Our focus was on biosignals, AI methodologies, datasets, devices, real-world implementation challenges. The an overview its biological mechanisms, details methodology search, synthesizes findings. It shows that especially EDA PPG, are frequently utilized demonstrate potential reliability in multimodal settings. Evidence trend towards deep learning found, although limited comparison with traditional methods calls further research. Concerns arise regarding representativeness datasets practical challenges deploying technologies, which include issues related to data quality privacy. Future research should aim develop comprehensive explore techniques not only accurate but also computationally efficient user-centric, thereby closing gap between theoretical applications improve effectiveness systems real scenarios.

Language: Английский

---

Review of wearable technologies and machine learning methodologies for systematic detection of mild traumatic brain injuries

Journal of Neural Engineering, Journal Year: 2021, Volume and Issue: 18(4), P. 041006 - 041006

Published: July 30, 2021

Mild traumatic brain injuries (mTBIs) are the most common type of injury. Timely diagnosis mTBI is crucial in making 'go/no-go' decision order to prevent repeated injury, avoid strenuous activities which may prolong recovery, and assure capabilities high-level performance subject. If undiagnosed, lead various short- long-term abnormalities, include, but not limited impaired cognitive function, fatigue, depression, irritability, headaches. Existing screening diagnostic tools detect acute andearly-stagemTBIs have insufficient sensitivity specificity. This results uncertainty clinical decision-making regarding returning activity or requiring further medical treatment. Therefore, it important identify relevant physiological biomarkers that can be integrated into a mutually complementary set provide combination data modalities for improved on-site mTBI. In recent years, processing power, signal fidelity, number recording channels wearable healthcare devices tremendously generated an enormous amount data. During same period, there been incredible advances machine learning methodologies. These achievements enabling clinicians engineers develop implement multiparametric high-precision this review, we first assess challenges mTBI, then consider hardware implementation sensing technologies used related Finally, discuss state art learning-based detection how more diverse list quantitative biomarker features improve current data-driven approaches providing patients timely

Language: Английский

---

Inferring Mood-While-Eating with Smartphone Sensing and Community-Based Model Personalization

ACM Transactions on Computing for Healthcare, Journal Year: 2025, Volume and Issue: unknown

Published: March 11, 2025

The interplay between mood and eating episodes has been extensively researched within the fields of nutrition, psychology, behavioral science, revealing a connection two. Previous studies have relied on questionnaires mobile phone self-reports to investigate relationship eating. In more recent work, sensor data utilized characterize both behavior independently, particularly in context food diaries health applications. However, current literature exhibits several limitations: lack investigation into generalization inference models trained with from various everyday life situations specific contexts like eating; an absence using explore intersection inadequate examination model personalization techniques limited label settings, common challenge (i.e., far fewer negative reports compared positive or neutral reports). this study, we examined two separate datasets different studies: i) Mexico (N \({}_{MEX}\) = 84, 1843 mood-while-eating distribution positive: 51.7%, neutral: 38.6% negative: 9.8%) 2019, ii) eight countries \({}_{MUL}\) 678, 329K reports, including 24K 83%, 14.9%, 2.2%) 2020, which contain passive smartphone sensing self-report data. Our results indicate that generic experience decline performance contexts, such as during eating, highlighting issue sub-context shifts sensing. Moreover, discovered population-level (non-personalized) hybrid (partially personalized) modeling fall short commonly used three-class task (positive, neutral, negative). Additionally, found user-level posed challenges for majority participants due insufficient labels class. To overcome these limitations, implemented novel community-based approach, building set users similar target user. findings demonstrate can be inferred accuracies 63.8% (with F1-score 62.5) MEX dataset 88.3% 85.7) MUL models, surpassing those achieved traditional methods.

Language: Английский

---

Personalization of Affective State Recognition from Physiological Signals: A Review

Communications in computer and information science, Journal Year: 2025, Volume and Issue: unknown, P. 143 - 158

Published: Jan. 1, 2025

Language: Английский

---