Source-linked AI summary
Remote patient monitoring using artificial intelligence: Current state, applications, and challenges
Thanveer Shaik, Xiaohui Tao, Niall Higgins, Lin Li, Raj Gururajan, Xujuan Zhou, U. Rajendra Acharya
TL;DR
Manual and conventional RPM can be constrained by staff workload and invasive, skin-contact monitoring, motivating a review of AI-enabled RPM. The paper synthesizes technologies, applications, impacts, challenges, and trends, finding that AI-enabled architectures support earlier deterioration detection, personalized monitoring, and adaptive behavior learning.
Problem
RPM needs alternatives to workload-dependent manual monitoring and invasive skin-contact approaches while supporting patients across remote, home, and hospital settings.
Method
The paper comprehensively reviews noninvasive RPM architectures, AI applications, enabling technologies, adoption challenges, and emerging trends.
Results
AI-enabled RPM architectures are reported to detect health deterioration early, personalize monitoring, and learn patient behavior patterns adaptively.
Takeaways & Limitations
AI, IoT, and distributed computing technologies broaden RPM across vital-sign, activity, emergency, and chronic-disease monitoring applications.
Takeaways & Limitations
AI adoption remains constrained by explainability, uncertainty in data and models, imbalanced datasets, and broader data-quality challenges.
Abstract
from arXiv · showhide
The adoption of artificial intelligence (AI) in healthcare is growing rapidly. Remote patient monitoring (RPM) is one of the common healthcare applications that assist doctors to monitor patients with chronic or acute illness at remote locations, elderly people in-home care, and even hospitalized patients. The reliability of manual patient monitoring systems depends on staff time management which is dependent on their workload. Conventional patient monitoring involves invasive approaches which require skin contact to monitor health status. This study aims to do a comprehensive review of RPM systems including adopted advanced technologies, AI impact on RPM, challenges and trends in AI-enabled RPM. This review explores the benefits and challenges of patient-centric RPM architectures enabled with Internet of Things wearable devices and sensors using the cloud, fog, edge, and blockchain technologies. The role of AI in RPM ranges from physical activity classification to chronic disease monitoring and vital signs monitoring in emergency settings. This review results show that AI-enabled RPM architectures have transformed healthcare monitoring applications because of their ability to detect early deterioration in patients' health, personalize individual patient health parameter monitoring using federated learning, and learn human behavior patterns using techniques such as reinforcement learning. This review discusses the challenges and trends to adopt AI to RPM systems and implementation issues. The future directions of AI in RPM applications are analyzed based on the challenges and trends
1 | INTRODUCTION
RPM combines noninvasive monitoring technologies with AI to support clinicians across remote, home, and hospital settings. The review examines architectures, applications, AI impacts, and adoption challenges.
- RPM technologies: RPM uses telehealth, wearable devices, contact-based sensors, and IoT methods to monitor vital signs and physiological parameters.Applications include motion recognition supporting clinical judgments and treatment plans.
- RPM applications: Traditional RPM serves rural patients, chronically ill people, and elderly individuals at home, while nonintrusive systems also support hospital and intensive-care monitoring.AI and machine learning can help clinicians visualize health status through vital-sign and activity-recognition data.
- Review scope: The review investigates noninvasive RPM technologies, AI applications for monitoring, AI-enabled early detection and personalization, and challenges to widespread adoption.Its scope includes vital signs, physical activities, emergencies, chronic diseases, and adaptive learning.
- Review scope: The study reviews traditional and deep learning, video monitoring, IoT devices, cloud, edge, fog, blockchain, reinforcement learning, and federated learning in RPM.It also explores challenges and trends in adopting AI-enabled RPM.
- Review organization: The paper organizes its discussion around research methods, advanced RPM architectures, AI applications, AI impacts, adoption challenges, and future work.The architecture coverage includes telehealth, IoT, cloud, fog, edge, and blockchain technologies.
2 | SEARCH STRATEGY AND SELECTION CRITERIA
The review defines research questions about technologies, AI impacts, adoption challenges, and trends in RPM, then applies a structured database search and selection process. Searches used predefined concepts, Boolean combinations, and eligibility limits, with the review process represented by a PRISMA flowchart.
- Search scope: The search targeted journal articles, reviews, and conference papers on AI-supported health-status monitoring using IoT devices or nontouch techniques.The objective covered geographically remote and local monitoring settings.
- Research questions: The review asks which technologies transformed manual hospital monitoring, how AI transformed RPM, what adoption challenges exist, and which AI trends are emerging.These questions address technologies, impacts, challenges, and trends separately.
- Database search: Literature was retrieved from Web of Science, Scopus, Springer, ACM, IEEE Xplore, PubMed, ScienceDirect, and MDPI using title, abstract, and keyword searches.Searches used keywords, Boolean operators, truncation, and wildcards, with results sorted by relevance.
- Search strategy: The search strategy was built by categorizing article-title concepts into five areas and generating synonyms before constructing Boolean combinations.The final search string combined patient, monitoring, remote setting, sensor, and AI or machine-learning terms.
- Selection criteria: Studies without AI or machine learning were excluded when they reported continuous monitoring, and the selected limits covered publications from 2016 to 2021.The review process was documented with a PRISMA flowchart.
3 | REMOTE PATIENT MONITORING ARCHITECTURES
Remote patient monitoring architectures combine telehealth, IoT devices, cloud, fog, edge, and blockchain technologies to support continuous and less invasive care. These systems extend monitoring beyond conventional manual and skin-contact approaches while introducing privacy, interoperability, latency, and access challenges.
- Manual hospital monitoring depends on workload and staff resources, while conventional devices require skin contact to estimate vital signs.
- Telehealth: Telehealth supports remote communication and monitoring, with applications spanning cardiovascular, pulmonary, mental-health, pain-management, blood-pressure, glucose, stroke, dermatological, and ophthalmic care.
- Telehealth: AI-enabled telehealth can estimate heart rate, respiratory rate, oxygen saturation, cough characteristics, and blood pressure from image, video, and sound data.
- Challenges: Telehealth and RPM remain constrained by internet dependence, rural–urban disparities, healthcare-cost increases from misuse, and patient-data security risks.
- IoT-enabled devices: IoT architectures connect wearable devices and sensors to collect patient data for continuous monitoring and clinical decision-making.
- IoT-enabled devices: Noninvasive RFID and Near-field Coherent Sensing can monitor internal mechanical motion and retrieve signals such as blood pressure, heart rate, and respiration rate.
- Cloud, fog, and edge computing: Cloud, fog, and edge architectures distribute storage and computation, while fog nodes reduce response delays by analyzing IoT data closer to its source.
- Blockchain monitoring: Blockchain architectures secure RPM data transactions and support decentralized storage, retrieval, analysis, and sharing across monitoring systems.
4 | AI IN RPM APPLICATIONS
AI in RPM uses traditional machine learning and deep learning to analyze vital signs and classify patients’ physical activities. These methods support detection and prediction tasks across remote monitoring applications.
- Traditional machine learning and deep learning are common AI methods for detecting and predicting vital signs and classifying patients’ physical activities.
4.1 | Vital signs monitoring
AI-based RPM systems monitor vital signs through wearable devices, ECG telemetry, and cloud-connected systems. The reviewed studies use models such as SVM and MLP to classify patient status and cardiac abnormalities.
- Wearable monitoring: Connected smartwatches continuously collect vital signs and send them to an administrator, whose SVM model supports patient-status decisions communicated to doctors.
- ECG monitoring: An SVM ECG telemetry system combines statistical ECG features with heart-rate-variability features to classify cardiac arrhythmia using 10-fold cross-validation.
- ECG monitoring: Portable ECG devices can transmit data to cloud services where the IDAH-ECG detector identifies abnormal heartbeats and informs physicians.
4.2 | Physical activities monitoring
AI-based physical-activity monitoring focuses especially on fall detection using wearable accelerometers, multisensor systems, RFID, and edge architectures. The reviewed studies compare multiple machine-learning and deep-learning models under different sensing and deployment constraints.
- Fall detection: Fall-detection systems use multisensor acceleration data, RFID-derived signal features, and wearable sensing to distinguish falls from daily activities.
- Algorithms: Studies compare SVM, random forest, KNN, naive Bayes, decision trees, adaptive boosting, LSTM, and CNN models for activity and fall recognition.
- Edge architectures: A resource-constrained edge system combines edge, fog, and cloud layers to collect, analyze, and transmit wearable accelerometer data despite latency, power, and connectivity concerns.
- Systems and devices: The reviewed physical-activity monitoring literature includes wearable, sensor, RFID-tag, and smart-device implementations using machine-learning and neural-network methods.
4.3 | Chronic disease monitoring
AI-based chronic disease monitoring applies machine learning to patient data and behavioral signals, including diabetes and mental health conditions.
- Diabetes monitoring uses machine learning classifiers with clinical and external factors to classify patients.
- Mental health applications use machine learning to detect symptoms, identify risk factors, predict progression, and personalize therapies.
- Mental health monitoring can analyze behavioral and clinical information, including demographics, appointments, notes, assessments, and referrals.
4.4 | Emergency monitoring
AI-based emergency monitoring uses clinical records and physiological signals to predict deterioration, cardiac arrest, mortality, and other critical outcomes.
- Machine learning approaches use emergency-department clinical records to predict outcomes and support automated decision-making for sepsis patients.
- AUROC for cardiac-arrest prediction was 0.781 with machine-learning scores versus 0.680 with MEWS.
- AUROC for in-hospital-death prediction was 0.741 with machine-learning scores versus 0.693 with MEWS.
- A machine-learning score ≥60 predicted cardiac arrest with 84.1% sensitivity, 72.3% specificity, and 98.8% negative predictive value.
- Logistic regression and LSTM models were evaluated for near-term mortality prediction in hospitalized patients with cirrhosis using medical-record features.
4.5 | Facial and emotions recognition
Facial and sensor-based emotion-recognition systems extend RPM beyond vital signs by inferring emotional states from face, heartbeat, temperature, respiration, and RFID signals.
- A smart RPM system combines face recognition, heartbeat, and temperature sensors to detect patients’ emotional states and heartbeat levels.
- An RFID-based framework extracts respiration and heartbeat features, then classifies users’ different emotions.
5 | AI IMPACT ON RPM
AI transforms RPM by supporting continuous deterioration detection, personalized monitoring, explainable prediction, and adaptive learning across connected healthcare architectures. The review highlights improved predictive performance, privacy-preserving federated learning, and reinforcement-learning approaches for modeling patient behavior and interventions.
- 5.1 | Early detection of patient deterioration: Traditional intermittent observation and early warning scores limit continuous detection, while AI models can predict deterioration before onset across hospital and home-care settings.
- 5.1 | Early detection of patient deterioration: An LSTM-RNN framework predicted patient deterioration 1 h before onset and performed better than the traditional method.
- 5.1 | Early detection of patient deterioration: A prognostic tool combining predicted vital signs, laboratory results, and vital signs achieved 80% accuracy for early diagnosis of worsening health status.
- 5.1 | Early detection of patient deterioration: Explainable AI early warning systems use temporal prediction and temporal explanations to help clinicians understand model behavior in sepsis, acute kidney injury, and acute lung injury cases.
- 5.2 | Personalized monitoring: Personalized RPM addresses individual treatment variability, while federated learning trains across decentralized patient devices without transferring patient data.
- 5.2 | Personalized monitoring: Reinforcement learning uses reward-driven sequential decisions to learn patient behavior patterns in uncertain environments.
- 5.2 | Personalized monitoring: Adaptive interventions can select notification types and frequencies while identifying opportune moments for delivery.
6 | CHALLENGES AND TRENDS OF AI IN RPM
AI-enabled RPM faces challenges involving interpretability, privacy, uncertainty, signal processing, imbalanced data, and dataset size. The review discusses technical approaches and practical considerations for addressing these barriers.
- Explainability: AI and ML models can outperform humans in complex prediction but often cannot explain their conclusions or causal relationships.Black-box models such as neural networks and SVMs hinder clinical adoption when healthcare professionals cannot assess how outputs were produced.
- Privacy: Deep neural networks may learn features that enable user identification, creating privacy leakage risks in wearable activity-recognition data.A logistic regressor achieved 84.7% user-classification accuracy from CNN features versus 35.2% from raw sensor data.
- Uncertainty: AI adoption involves uncertainty from data acquisition, DNN construction, and modeling results.Measurement noise, real-world variability, and numerous DNN hyperparameters contribute to data and model uncertainty; uncertainty quantification can reduce its effects.
- Signal processing: Noninvasive RPM signal processing remains difficult because environmental noise and signal fluctuations can obscure respiration and heartbeat measurements.Reported approaches include smoothing, filtering, interpolation, Fourier transforms, and motion detection.
- Imbalanced dataset: Imbalanced datasets can bias decisions because conventional algorithms tend to favor majority classes over minority classes.Undersampling, clustering, and oversampling methods such as SMOTE are used to rebalance clinical data.
- Imbalanced dataset: In a review of imbalanced clinical datasets, SMOTEEN combined with KNN achieved the highest accuracy, recall, precision, and F1 score among evaluated techniques.The comparison included six classifiers and seven label-balancing techniques.
- Dataset volume: Training robust RPM models generally requires large, informative datasets with many subjects, although requirements vary by algorithm.Random forests may approach peak performance with relatively few cases, whereas neural-network performance can improve as more data become available.
7 | FUTURE DIRECTION OF AI ON RPM
Future AI-enabled RPM research should address explainability, privacy, uncertainty, and data-quality challenges while extending applications for providers and patients. The review also identifies reinforcement learning and virtual robots as potential directions for monitoring and prediction.
- Explainability: Future RPM work should improve explainability so healthcare professionals can better understand patient status and make informed decisions.SHAP, LIME, and DeepLIFT are among the techniques being adopted to clarify machine-learning and deep-learning results.
- Privacy: Federated learning may address health-data privacy, but evidence is insufficient to confirm that reverse-engineering local parameters cannot expose private data.Future work should strengthen federated-learning privacy and security.
- Uncertainty: Uncertainty quantification can reduce uncertainty during model optimization and decision-making, potentially improving healthcare professionals’ trust in model results.The review distinguishes aleatoric and epistemic uncertainty and discusses probabilistic and nonprobabilistic approaches.
- Data challenges: AI-based RPM research must prioritize input data because signal noise, imbalance, limited labels, and feature-extraction problems affect results.The review describes efficient, clean data as the first and most time-consuming part of AI methodology.
- Reinforcement learning: Reinforcement learning could support human-behavior modeling, social assistive robots, dynamic treatment regimens, and just-in-time adaptive interventions.The review recommends virtual robots for patient monitoring and prediction of unprecedented events because physical robots have caused threats to humans.
8 | CONCLUSION
The survey centers on AI-enabled RPM for vital signs and physical-activity monitoring, emphasizing advanced infrastructures and AI methods for preventive, predictive, and personalized care. Its scope excludes EEG and neurological diseases and does not cover all chronic-disease monitoring research.
- The review examines noninvasive RPM technologies and AI applications for vital signs, physical activities, chronic diseases, and emergencies.
- Federated learning supports patient-centric monitoring while protecting data privacy, and reinforcement learning adaptively learns patient behavior patterns.
- AI-enabled RPM can support preventive, predictive, and personalized patient monitoring while assisting healthcare practitioners.
- The review does not explore all chronic-disease monitoring research.
CONFLICT OF INTEREST
The supplied material reports no conflict of interest, no applicable data sharing, author identifiers, and a list of related works. It does not provide substantive findings for this section.
- All authors declare that they have no conflict of interest in this work.
- Data sharing is not applicable because no datasets were generated or analyzed during the study.
- The document lists ORCID identifiers for the study’s authors.
- The related-work material lists references on IoT, data analytics, Healthcare 4.0, AI models, telehealth, and digital care implementation.