A comprehensive health assessment approach using ensemble deep learning model for remote patient monitoring with IoT

Abstract

The goal of this research is to create an ensemble deep learning model for Internet of Things (IoT) applications that specifically target remote patient monitoring (RPM) by integrating long short-term memory (LSTM) networks and convolutional neural networks (CNN). The work tackles important RPM concerns such early health issue diagnosis and accurate real-time physiological data collection and analysis using wearable IoT devices. By assessing important health factors like heart rate, blood pressure, pulse, temperature, activity level, weight management, respiration rate, medication adherence, sleep patterns, and oxygen levels, the suggested Remote Patient Monitor Model (RPMM) attains a noteworthy accuracy of 97.23%. The model's capacity to identify spatial and temporal relationships in health data is improved by novel techniques such as the use of CNN for spatial analysis and feature extraction and LSTM for temporal sequence modeling. Early intervention is made easier by this synergistic approach, which enhances trend identification and anomaly detection in vital signs. A variety of datasets are used to validate the model's robustness, highlighting its efficacy in remote patient care. This study shows how using ensemble models' advantages might improve health monitoring's precision and promptness, which would eventually benefit patients and ease the burden on healthcare systems.

Introduction

The introduction of cutting-edge technology, notably in the areas of deep learning and the Internet of Things (IoT), has caused a paradigm change in the evolution of healthcare. The Remote Patient Monitor Model (RPMM), an ensemble deep learning model created to improve the accuracy and effectiveness of remote patient monitoring (RPM), is one of the notable developments in this field. This extensive study explores the historical background, the pressing need for improved RPM models, and the special ways that RPMM can help meet these demands. In the past, in-person consultations were the main method of healthcare monitoring, which made it difficult to collect ongoing, real-time patient data. The introduction of RPM signaled a revolutionary stage in medical technology by allowing doctors to remotely check on patients' vital signs and other health indicators. RPM originated in the late twentieth century as efforts in telemedicine started looking into ways to reduce the distance between patients and medical professionals. Prioritizing telephone consultations and rudimentary monitoring during the initial implementations paved the ground for later, more advanced technologies.

Traditional monitoring methods in modern healthcare frequently fall short of fully addressing the intricacies included in patient data, especially when addressing temporal fluctuations and individual patient characteristics. These drawbacks highlight the need for creative solutions that can get past these obstacles and offer better patient health monitoring and management.

Our suggested approach fills a major research void in the area of temporal data analysis. The complex temporal dynamics found in patient physiological signals, such as heart rate variability, blood pressure changes, or respiration patterns, are frequently missed by conventional monitoring techniques. Due to this constraint, important information that is stored in the temporal evolution of patient data may go unnoticed, which might impair the accuracy of diagnostic and prognostic judgments.

Moreover, current methods for remote patient monitoring frequently fail to adjust to the unique needs of each patient, which results in less than ideal outcomes in settings for tailored healthcare. Individual patients have various physiological profiles and react differently to therapy interventions; therefore, adaptive monitoring systems that can learn and adjust over time to each patient's unique characteristics are required.

Our suggested LSTM-CNN ensemble model fills in these knowledge gaps by providing a dependable and adaptable remote patient monitoring system. Our approach successfully captures the geographical patterns and temporal relationships found in patient data by combining CNN and LSTM architectures, allowing for more precise and individualized patient health monitoring. By offering a thorough framework for continuous health monitoring that can adjust to the unique variations of each patient and capture intricate temporal dynamics, this novel approach closes a significant gap in the body of research and eventually improves patient outcomes and clinical decision-making.

In contrast to conventional statistical techniques such as decision trees or linear regression, the ensemble approach of RPMM enables a more sophisticated comprehension of the relations among many health factors. Furthermore, RPMM's inclusion of both architectures allows it to capture spatial and temporal correlations simultaneously, improving its ability to detect small changes in vital signs and health trends. This is in contrast to certain other deep learning models that may only use CNN or LSTM independently. For instance, algorithms that only use LSTM may find it difficult to extract spatial features from sensor data, whereas algorithms that just use CNN may find it difficult to efficiently capture temporal sequences. Moreover, while some RPM algorithms might concentrate on a small number of health parameters or not be able to integrate data in real-time, RPMM's thorough assessment of important health metrics and its easy integration with wearables that are IoT-enabled guarantee a comprehensive approach to remote patient monitoring.

When compared to other methods currently used for remote patient monitoring, the LSTM + CNN model presented in the study offers a number of improvements and benefits. Primarily, it is likely that the model has a meticulously developed architecture that best combines CNN and LSTM layers. Through this integration, the model is able to more accurately and perceptively measure health state over time by capturing both temporal dynamics and spatial correlations in physiological data. Some existing LSTM + CNN approaches, on the other hand, might use less optimal architectures, which could make it more difficult for them to identify intricate patterns in medical data. Because of its sophisticated architecture and training methods, the suggested model is also anticipated to exhibit improved performance measures, such as accuracy, precision, recall, and F1-score values. Ensuring dependable and practical insights from remote patient monitoring data is contingent upon this performance enhancement. Furthermore, the proposed LSTM + CNN model places a strong emphasis on real-time monitoring capabilities, which allow for quick responses to anomalies and continuous patient care. It differs from other methods that might not place as much emphasis on real-time monitoring or that do not have the infrastructure needed for a smooth interaction with Internet of Things devices. Crucially, the suggested model exhibits resilience in several demographic categories, guaranteeing uniform outcomes for a range of patient cohorts. Its application in many healthcare contexts is contingent upon its adaptability and generalization. All things considered, the LSTM + CNN model described in this work provides a thorough and sophisticated method of remote patient monitoring, utilizing cutting-edge methods to enhance patient care and healthcare results.

Healthcare entered a new phase in the twenty-first century with the introduction of IoT, which made it possible to seamlessly integrate wearables, sensors, and other linked technology. This connectivity made it easier to gather a wide range of health data, which set the stage for more sophisticated and all-encompassing RPM systems. At the same time, deep learning algorithms became popular across a range of industries due to their capacity to identify intricate patterns from large datasets, opening up new opportunities for healthcare analytics. There is an increasing need for models that can collect a broad range of health data and deliver precise and fast insights as the need for remote patient monitoring grows. The intricacy of health data, the requirement for real-time analysis, and the capacity to identify small patterns suggestive of possible health problems are common obstacles faced by existing RPM models. The dynamic and diverse nature of patient data may make it difficult for traditional models to adjust, requiring the investigation of more complex solutions.

Furthermore, the need for strong RPM models that can provide tailored insights and early intervention techniques is highlighted by the rise in chronic diseases, an aging population, and the global push for customized medicine. The need for creative solutions is highlighted by the shortcomings of traditional models in terms of offering a comprehensive picture of a patient's health status and in identifying complex patterns in vital sign data. The Remote Patient Monitor Model (RPMM) is introduced as an ensemble deep learning architecture in response to the shortcomings of the current models. By fusing temporal sequence modeling, spatial analysis, and feature extraction with Long Short-Term Memory (LSTM) networks, the RPMM leverages the advantages of both architectures. RPMM is unique in that it is specifically designed for Internet of Things (IoT) devices that are utilized for remote patient monitoring, with an emphasis on attaining an impressive accuracy rate of 97.23%.

The decision to use an ensemble model was made after it was realized that combining CNN and LSTM could improve the capacity of the model to represent temporal and spatial interdependence in health data. This two-pronged strategy guarantees a deeper comprehension of the complex patterns included in vital sign data. Interestingly, the model is trained to assess vital signs such as blood pressure, temperature, activity level, weight management, respiratory rate, oxygen saturation, heart rate, blood pressure, pulse, and sleep patterns. With an astounding accuracy of 97.23%, the RPMM outperformed a number of other models. In the medical field, where prompt and precise interventions are critical, this high accuracy is essential. By assessing a wide range of health-related factors, RPMM offers a more comprehensive picture of the patient's condition. RPMM differs from models that concentrate on a small number of vital signs in that it takes a holistic approach. RPMM records real-time physiological data by connecting with IoT-enabled wearable devices in a smooth manner. This real-time component provides a dynamic and current picture of the patient's health and is crucial for prompt actions and ongoing monitoring.

The RPMM, which provides a paradigm-shifting approach to remote patient monitoring, is evidence of the convergence of deep learning and IoT in the healthcare industry. With its exceptional accuracy, thorough parameter analysis, and real-time data integration, RPMM stands out as a significant development in the field of healthcare analytics' quest for accuracy and effectiveness. As the healthcare industry develops, RPMM acts as a lighthouse, pointing the path toward a day when remote patient monitoring is not only precise and prompt but also proactive and tailored to the needs of the individual. The main contribution of this study is

• 1. To improve patient outcomes through faster, more precise health parameter monitoring. Through the use of IoT technologies and deep learning algorithms, RPMM is able to identify even the smallest changes in a patient's condition, allowing for early intervention and individualized care plans that are catered to each patient's needs.

• 2. To enable remote monitoring, lessen the need for frequent hospital visits, and maximize resource utilization in order to streamline the delivery of healthcare services. Healthcare providers can efficiently allocate resources, prioritize cases based on severity, and monitor multiple patients at once thanks to RPMM's real-time data integration capabilities.

• 3. To detect possible health problems before they become serious in order to change the focus of healthcare from reactive to proactive.

The rest of this study is organized as follows: The previous and the most related research works are presented in Section “Related works”. Section “Dataset description” explains the proposed model in detail. The results are discussed in Section “System design”. The proposed work is concluded with future work in Section “Experimental results”.

Related works

Deep learning algorithms have become extremely useful technologies with applications in many different fields, including medicine. We examine important papers that use deep learning in remote patient monitoring (RPM) and related fields in this review of the literature. The surveyed publications provide information on smart healthcare systems, disease diagnostics, optimization methods, and algorithmic reviews.

The fundamental understanding of deep learning algorithms and architectures provided by Shrestha and Mahmood¹ paves the way for further study in the subject. Expanding on this, Ranganathan² explores how preprocessing methods can improve deep learning model performance, especially for image processing jobs related to medical diagnosis. By contrasting different optimization methods, Soydaner³ makes a significant contribution to the optimization of deep learning algorithms and provides insightful information about algorithm selection and performance optimization tactics. Islam et al.⁴ show the effectiveness of AI-driven methods in medical image analysis by conducting a comprehensive review and meta-analysis of deep learning algorithms for diagnosing diabetic retinopathy, which moves into healthcare applications.

Using deep learning methods, Mendonca et al.⁵ provide a lightweight intelligent intrusion detection system designed for the industrial Internet of Things (IIoT) context. The integrity of data and operations is critical in IIoT installations, where strong security measures are needed. This system satisfies this demand. Similar to this, Zhang et al.⁶ investigate how deep learning can be used for defect diagnosis and emphasize how this could lead to predictive maintenance in industrial settings. The aforementioned research highlight the adaptability of deep learning in tackling intricate healthcare problems.

Rajyalakshmi and Lakshmanna⁷ explore the integration of deep learning algorithms with smart city infrastructure and IoT technologies, going beyond specific medical applications and opening up new avenues for creative approaches to urban healthcare delivery. Guo et al.⁸ demonstrate the potential of deep learning to improve public health measures, particularly during pandemics like COVID-19, by utilizing it for infectious disease prediction.

Farias et al.⁹ provide a thorough analysis of remote patient monitoring systems in the context of RPM, highlighting both the advantages and disadvantages of each. The use of mobile health (mHealth) applications for managing chronic diseases is examined by El-Rashidy et al.¹⁰, who highlight how technology may empower patients and enhance healthcare results. Taiwo and Ezugwu¹¹ concentrate on smart healthcare support systems, emphasizing the value of virtual healthcare delivery and patient safety.

In order to ensure usability and acceptance among patients and healthcare professionals, Poncette et al.¹² place a high priority on human-centered design while developing RPM systems. Blockchain-based integrity management for remote patient monitoring is proposed by Jamil et al.¹³, solving privacy and data security issues. The significance of data integrity and user-centric design in RPM systems is emphasized by these research.

Novel approaches are employed to improve remote healthcare monitoring by utilizing emerging technologies including big data, cloud analytics, and IoT^14,15,16. These studies show how interdisciplinary methods have the power to transform healthcare delivery and enhance patient outcomes.

The LSTM + CNN model offers a comprehensive approach to remote patient monitoring across a wide range of health parameters, in contrast to studies that concentrate on specific medical conditions like breast cancer detection, COVID-19 diagnosis, or Alzheimer's disease prognosis^17,18,19. Although some studies use CNN or LSTM alone to diagnose diseases, the ensemble model combines the advantages of both architectures to efficiently capture temporal and spatial relationships in physiological data²⁰.

The suggested deep learning model guarantees dependable performance independent of age or other demographic characteristics, in contrast to research that might not adequately handle demographic variability or generalize across populations^21,22,23.The comparison with baseline models demonstrates the proposed deeplearning model's superior performance in terms of recall, accuracy, precision, and F1-score values, underscoring its potential for high-performance patient monitoring and disease diagnosis^21,22,23,24.

In the age of deep learning, Yang, Lv, and Chen's²⁵ thorough overview of ensemble learning approaches highlights the LSTM + CNN model as one particularly creative use in this field. The model combines the best features of both CNN and LSTM architectures to efficiently capture temporal and spatial correlations in physiological data.

Moreover, Aravindhan, Sangeetha, and Kamesh²⁶ emphasize how critical it is to enhance cloud computing IoT communication performance using hybrid frameworks.

The LSTM + CNN model also distinguishes itself by emphasizing real-time monitoring and prompt response based on forecasts. The suggested methodology allows for continuous patient monitoring utilizing IoT-enabled wearable devices, whereas some research concentrate on offline diagnosis or retrospective analysis^18,27.

The real-time monitoring capability of the LSTM + CNN model ensures timely intervention to prevent adverse health events, potentially reducing hospital readmissions. Additionally, its high accuracy across age groups demonstrates robustness across demographics, ensuring dependable performance independent of age or other characteristics^28,29.

The LSTM + CNN model performs better than baseline models in terms of accuracy, precision, recall, and F1-score values, as demonstrated by the comparison analysis^27,30. The suggested model makes use of the most recent developments in deep learning to obtain state-of-the-art outcomes in remote patient monitoring,^27,29.

Satheeshkumar and Sathiyaprasad³¹ use machine learning and metaheuristic optimization algorithms-based feature extraction and classification strategies to study medical data analysis. The significance of algorithmic optimization and effective data representation in healthcare applications is emphasized by their study. In their co-learning model for medical big data analysis, Rao et al.³² demonstrate the efficiency of Long Short-Term Memory (LSTM) networks in processing sequential data by combining optimization methods such as Whale Optimization.

Ahmed et al.³³ highlight the importance of deep learning in personalized healthcare interventions by introducing a heterogeneous network embedded medicine recommendation system based on LSTM in the field of medicine recommendation systems. Ganaie et al.³⁴ present a thorough analysis of ensemble deep learning techniques, shedding light on their efficacy and usefulness in the field of medicine.

Zhang et al.³⁵ evaluate different ensemble learning algorithms used in remote sensing data analysis and further study ensemble learning approaches in remote sensing applications. In order to further the development of ensemble approaches in machine learning, Savargiv et al.³⁶ provide a novel ensemble learning strategy based on learning automata. Furthermore, Zhou et al.³⁷ present an ensemble learning technique that combines group decision-making with deep neural networks, demonstrating the potential of hybrid techniques in enhancing classification performance. Edeh et al.'s³⁸ artificial intelligence-based ensemble learning model predicts the presence of hepatitis C, one of the healthcare-specific applications of ensemble learning.

Almulihi et al.³⁹ emphasize the value of ensemble approaches in clinical decision support systems by proposing an ensemble learning strategy for early cardiac disease prediction based on a hybrid deep learning model.Younas et al.⁴⁰ provide a deep ensemble learning method for colorectal polyp classification in the context of illness classification, illustrating the usefulness of optimal network parameters in enhancing classification accuracy. Although Ait Nasser and Akhloufi⁴¹ concentrate on the classification of chest disorders using chest X-ray pictures and deep ensemble learning techniques, Aboulmira et al.⁴² investigate ensemble learning methods for skin lesions classification.Research by Khalaf et al.⁴³ and Sangeetha et al.⁴⁴ presents innovative methods for lung cancer classification and anomaly detection in distributed cloud IoT networks, respectively, demonstrating the ongoing development and innovation in deep learning and ensemble methods for healthcare applications.

Many research use conventional machine learning techniques, including Random Forests and Support Vector Machines (SVM), which may find it difficult to represent the intricate temporal dynamics and correlations found in human activities. Furthermore, these methods are less flexible to different datasets and activity kinds since they frequently call for manually designed feature engineering.

Moreover, even though deep learning models have demonstrated potential in HAR tasks, there exist certain constraints associated with them. Obtaining large amounts of labeled data for training can be difficult or expensive, particularly for specialized activities or uncommon events. This is a typical difficulty. Furthermore, interpretability is typically lacking in deep learning models, which makes it challenging to comprehend the underlying mechanisms influencing model predictions—a critical skill for applications in assistive technology and healthcare.

Notwithstanding these drawbacks, new developments in HAR research have brought fresh methods and strategies to deal with these issues. Modern techniques reduce the requirement for manual feature extraction by automatically learning discriminative features from raw sensor data using deep learning architectures like recurrent neural networks (RNNs) and convolutional neural networks (CNNs). Furthermore, it has been suggested that attention mechanisms and graph neural networks can better capture long-range dependencies and spatial–temporal linkages in activity sequences.

Abdel-Basset et al.⁴⁵ present ST-DeepHAR, a deep learning model created especially for Internet of Health Things (IoHT) applications that require Human Activity Recognition (HAR). In order to depict the complex dynamics of human actions over time, this model most likely stresses the inclusion of spatiotemporal information. A thorough analysis of human activity recognition techniques is given by Islam et al.⁴⁶, with a special emphasis on the application of convolutional neural networks (CNNs). They probably reviewed a wide range of CNN architectures, datasets, difficulties, and prospects for the future of the area. Gu et al.⁴⁷ explore deep learning methods for HAR, encompassing a broad spectrum of models, including CNNs and Recurrent Neural Networks (RNNs), as well as their developments and applications. The novel "Attend and Discriminate" method for HAR proposed by Abedin et al.⁴⁸ makes use of wearable sensors and most likely attention mechanisms to distinguish between various activities.

A novel feature extraction method was presented by Supriya, Siuly, Wang, and Zhang⁴⁹ in a complex network framework with the goal of improving automated seizure identification. Their method, which highlights important developments in the fields of applied acoustics and medical diagnostics, focuses on utilizing the characteristics of complex networks to increase the precision and dependability of seizure detection using acoustic signals.

A blockchain and federated learning architecture for the classification of physiological signals was created in 2023 by Sun, Wu, Xu, and Zhang⁵⁰. Continuous learning is incorporated into this framework, enabling the model to adjust and get better over time as new data becomes available. The system addresses privacy issues, guarantees data confidentiality and integrity, and strengthens the robustness of physiological signal classification in healthcare by incorporating blockchain technology.

A federated learning system that is both scalable and transportable was introduced by Sun and Wu⁵¹ with the express purpose of identifying sensor data related to healthcare. Their system is built to handle a wide range of datasets from different sources without sacrificing scalability or performance. With the help of this federated learning strategy, several institutions can work together on model training without exchanging private information, protecting patient privacy and data security while increasing classification accuracy.

To accomplish effective biomedical picture segmentation, Zhang, Zhang, Zhu, and Xu⁵² developed a unique approach incorporating limited multi-scale dense connections. Their approach, which was presented at the 2020 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), aims to increase image segmentation accuracy by using dense connections at various scales. This method improves the model's capacity to identify minute details in biomedical images.

To sum up, the reviewed literature shows how deep learning and ensemble methods are becoming more and more important for illness detection, recommendation systems, and medical data analysis. The creation of interpretable ensemble models, the incorporation of multimodal data sources, and the investigation of collaborative learning strategies for decentralized healthcare analytics are potential avenues for future research.

Dataset description

The Vital DB dataset includes a broad range of vital physiological indicators that are important for clinical analysis and remote patient monitoring. A thorough collection of high-fidelity multi-parameter vital sign data from surgical patients is provided by the VitalDB dataset. Research and development in a variety of healthcare applications, especially those that center on clinical decision support systems and remote patient monitoring, will greatly benefit from this dataset. Numerous physiological data, including heart rate, blood pressure, respiratory rate, oxygen saturation, and temperature, are included in every vital signs recording. These metrics are essential for continuous monitoring throughout surgical procedures and postoperative care since they are important markers of a patient's health status. Due to the richness and granularity of the dataset, we were able to examine correlations, trends, and patterns in physiological data through in-depth examinations of patient circumstances. Furthermore, the inclusion of actual data from surgical settings increases the dataset's relevance and practicality in real-world clinical settings. The information is utilized to create and verify forecast models, evaluate patient results, and enhance the standard of treatment.

These indicators enable thorough evaluations and provide insightful information about the patient's health. The frequency of heart contractions, or heart rate, is expressed in beats per minute (bpm), which reflects how well the heart pumps blood throughout the body. One possible value is 75 bpm. Mean arterial pressure (MAP), diastolic pressure (measured in between heartbeats), and systolic pressure (during heartbeats) make up blood pressure, which is expressed in millimeters of mercury (mmHg). A blood pressure reading might be, for example, 120 mmHg for the diastolic portion and 93 mmHg for the systolic portion. The frequency of breathing cycles is indicated by respiratory rate, which is expressed in breaths per minute and reflects the effectiveness of oxygen exchange and lung ventilation. 16 breaths per minute could be an example of a value. The percentage of hemoglobin saturated with oxygen in arterial blood is known as oxygen saturation. It acts as a vital sign of tissue oxygenation and respiratory function. A possible example value is 98%. Table 1 shows the dataset description.

Table 1 Parameter description.

End-tidal Carbon Dioxide (EtCO2) is the amount of carbon dioxide in the air at the end of a breath exhalation, expressed in millimeters of mercury (mmHg). It provides information on proper ventilation and the elimination of carbon dioxide. A 40 mmHg value is one example. An electrocardiogram (ECG) is a recording of the electrical activity of the heart that shows anomalies in conduction, rhythm, and heart rate. A 70 bpm heart rate in a sinus rhythm could be indicated by an example value. The body's temperature, expressed in degrees Celsius, reflects metabolic activity and represents the thermal condition of the individual. It helps in diagnosing hypothermia or fever. A temperature of 37.0 degrees Celsius could be an example.

As a result of each variable's perfect correlation with itself, the values in the correlation matrix along the diagonal (from top-left to bottom-right) are all 1. Because the correlation between variable A and variable B is equal to that between variable B and variable A, the values below and above the diagonal are identical. The interval of the correlation coefficient is − 1 to 1. The correlation matrix is often used to interpret correlations between various physiological indicators.

A summary of several physiological parameters is given in Table 1, together with information on sample values within normal ranges, significance in healthcare monitoring, techniques for monitoring, and frequency of data. These variables are frequently tracked in clinical settings and are crucial for determining a person's general health state.

As shown in Table 2, there is a significant positive relationship between heart rate and oxygen saturation, diastolic blood pressure, and systolic blood pressure. Heart rate and oxygen saturation possess a relative negative correlation with respiratory rate. The relationship between temperature and end-tidal carbon dioxide is moderately negative. Understanding the relationships between the physiological indicators is made possible by these correlations, which enable clinicians to make decisions and comprehend patients' health status.

Table 2 Sample dataset.

Figure 1 shows how a patient's physiological parameters have changed over time, either individually or in a group of patients. It displays, for example, changes in temperature, end-tidal carbon dioxide, heart rate, blood pressure, respiratory rate, and oxygen saturation during the course of the recording period. Healthcare providers can monitor patients' conditions and identify any significant changes that could need attention with the help of this visualization.

Figure 1

Psychological parameters over time. An examination of how various psychological factors or metrics change or evolve over a period. This could involve tracking parameters such as mood, stress levels, cognitive function, or any other psychological variables across different time points to understand patterns or trends over time.

System design

Data preprocessing

To make sure the physiological data is appropriate for training the model, the first step is to preprocess it. This incorporates handling missing values, feature normalization, and data segmentation into fixed-length time-series sequences. The mentioned algorithm's data preparation step involves a multimodal strategy for selecting and enhancing the unprocessed physiological data streams that are taken from Internet of Things (IoT) wearables that are used for remote patient monitoring (RPM). Before the data is used in the ensemble LSTM-CNN model, it undergoes a number of complex procedures designed to improve its quality, consistency, and relevance.To begin with, the raw data is carefully cleaned and filtered to remove any anomalies, artifacts, or erroneous readings that could compromise the dataset's integrity. In order to identify and correct aberrant data points, sophisticated signal processing techniques are applied, such as noise reduction algorithms and outlier recognition approaches. The preprocessed data is then subjected to feature extraction, which condenses pertinent data about the vital health parameters—heart rate, blood pressure, respiration rate, oxygen saturation, and temperature—into a clear and understandable graphic. In order to identify significant elements that capture the underlying physiological dynamics, sophisticated signal processing methods and mathematical transformations have to be applied. In order to further aid in the identification of temporal relationships and patterns within the data, the preprocessed data is further divided into discrete time intervals or epochs via temporal segmentation. By dividing the data into consecutive segments with a set length, the segmentation procedure allows the model to record and examine temporal sequences of physiological fluctuations and events. Furthermore, the preprocessed data could go through standardization or normalization to guarantee scale homogeneity and consistency across various features and parameters. Rescaling the data to a common range or distribution is part of the normalization process, which facilitates the elimination of any biases or disparities that can result from different measurement units or magnitudes.

To further mitigate the curse of dimensionality and improve computing performance, the preprocessed data is additionally subjected to dimensionality reduction techniques like principal component analysis (PCA) or feature selection algorithms. The input data is streamlined for the ensemble LSTM-CNN model by selecting and keeping the most useful features and removing redundant or unnecessary ones. The preprocessed data is eventually carefully selected and improved upon to provide an extensive and superior dataset that forms the basis for training, validating, and assessing the ensemble LSTM-CNN model. By carefully utilizing advanced data pretreatment methods, the algorithm makes sure that the input data is ready for reliable and accurate inference when it comes to remote patient monitoring.

Collect real-time physiological data from loT-enabled wearable devices, denoted as

$$=\left\{{x}^{(1)},{x}^{(2)},\dots ,{x}^{(m)}\right\},$$

(1)

where ${x}^{(i)}$ represents a sequence of physiological measurements over time for patient $i$.

Pre-process the data by handling missing values: ${x}_{t}^{(i)}=0$ for missing values at time $t$.Normalize features:

$${x}_{t}^{(i)}=\frac{{x}_{t}^{(i)}-\mu }{\sigma },$$

(2)

where $\mu$ is the mean and $\sigma$ is the standard deviation of the feature. Segment data into time-series sequences of fixed length $T$.

The steps involved prior to and following preprocessing are described in detail in Tables 3 and 4, respectively. They show how raw patient data is gathered, cleaned, features are extracted, and temporal segmentation is done to get the data ready for analysis. In order to guarantee data accuracy and integrity, Table 5 lists quality control metrics like signal-to-noise ratio and outlier detection rate. To effectively train and evaluate models, Table 6 presents the data partitioning scheme, which splits the data into subsets for validation, testing, and training. An overview of the preprocessed data is given in Table 7, which includes the mean, standard deviation, minimum, maximum, and quartiles for important characteristics such as blood pressure, heart rate, and breathing rate. In order to guarantee data cleanliness and dependability, Table 8 records the cleaning procedures carried out on the data, such as noise removal, outlier detection, and missing value imputation. The temporal segmentation details are shown in Table 9, which also includes the intervals or epochs into which the data is split for analysis. In order to facilitate additional analysis and interpretation, Table 10 lists data transformation operations such as logarithmic scaling that were applied to the data. The process of preparing, validating, and analyzing patient data in a remote monitoring context is illustrated in these tables collectively, guaranteeing the data's quality, dependability, and appropriateness for medical decision-making.

Table 3 Before preprocessing.

Table 4 After preprocessing.

Table 5 Quality control metrics.

Table 6 Data partitioning scheme.

Table 7 Preprocessed data overview.

Table 8 Data Cleaning Log.

Table 9 Temporal Segmentation details.

Table 10 Data transformation log.

LSTM-CNN model

The physiological data—which includes vital signs like heart rate, blood pressure, temperature, and activity levels—is gathered from remote patient monitoring devices and fed into the Ensemble LSTM-CNN Model for Remote Patient Monitoring (RPMM). The gathered data is next subjected to preprocessing, which includes handling missing values, normalizing features to guarantee consistency, and segmenting the data into time-series sequences with set durations, like 30-min intervals. Next, the preprocessed input is fed into the Long Short-Term Memory (LSTM) network and the Convolutional Neural Network (CNN), which are the two key components. Using convolutional and max-pooling procedures, the CNN's layers—which consist of 64 and 128 filters, respectively—extract spatial patterns from the data. In the background, the two-layer, 64-unit LSTM network learns sequential patterns over time by capturing temporal dependencies in the data.

As shown in Fig. 2, a thorough representation of the patient data is created by merging temporal and spatial characteristics from the outputs of the CNN and LSTM using ensemble fusion. Optimization methods including backpropagation using an Adam optimizer, 0.001 learning rate, and 32-epoch batch size category cross-entropy loss are used to train the model. The model's performance is assessed throughout training, and it achieves 97.5% accuracy on the testing set. After training, the model is used to enable timely actions based on predictions, constantly data streams are analyzed from IoT-enabled devices, and monitored in real-time. To maximize performance, hyperparameters are also adjusted iteratively, and the model architecture is modified in response to input. Subsequently, the model's effect on patient outcomes is evaluated, and potential future research and development paths are investigated to improve healthcare delivery even further.

Figure 2

Ensemble Architecture. Combines the outputs of the Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) models. This architecture merges spatial and temporal features extracted from physiological data to enhance the accuracy and robustness of the remote patient monitoring system.

Algorithm

Use backpropagation and optimization algorithms (e.g., Adam) to minimize the loss.

The proposed algorithm presents a thorough method of using deep learning techniques, namely Long Short-Term Memory (LSTM) networks and Convolutional Neural Networks (CNNs), for remote patient monitoring (RPM). The first step is data preprocessing, which involves gathering and analyzing physiological data in real time from wearables with Internet of Things capabilities. The characteristics are standardized, missing values are handled, and the data is divided into time-series sequences with set lengths.

Using CNNs for spatial feature extraction and LSTM networks for temporal sequence modeling are the next steps. CNNs are used to extract spatial properties from the data using max-pooling layers for downsampling after convolutional layers. In contrast, temporal dependencies in the sequential data are captured by LSTM networks, which helps the model understand long-term patterns and dynamics seen in patient health data.

Subsequently, the technique concatenates the features of the output representations obtained from the CNN and LSTM models to create an ensemble representation. This integration offers a comprehensive picture of the underlying health conditions by combining the temporal and spatial data that was retrieved by both models.

Lastly, for classification and prediction tasks, the ensemble representation is fed into a fully connected layer and then softmax activation. With the help of this stage, the model can now categorize patient health data and forecast outcomes using the features it has learned.

Convolutional neural networks (CNNs) and long short-term memory (LSTM) models are the two types of neural network models that are built and trained using the configurations and parameters listed in these tables. The CNN model's architecture is described in Table 11, along with the layer types, filter sizes, number of filters, and activation functions that are employed. It discusses the two convolutional layers that come before the max-pooling layers, which are standard parts of CNNs used for dimensionality reduction and feature extraction. The LSTM model's configuration is shown in Table 12, which also includes information on the activation function used in each cell, the number of units (or cells) in each LSTM layer, and the dropout rate used to prevent overfitting. The training parameters (optimizer (Adam), learning rate, loss function (categorical cross-entropy), number of epochs, and batch size) that are shared by both models are listed in Table 13. These parameters control how the models learn from the data and adjust their parameters to minimize the loss function during the training process.

Table 11 Convolutional neural network (CNN).

Table 12 LSTM model configuration.

Table 13 Training parameters.

Experimental results

Performance analysis

The study involves various evaluation metrics to assess the performance of the proposed ensemble deep learning model for remote patient monitoring.

Accuracy is calculated by using the formula,

$$\text{Accuracy}=\frac{\text{ Number of Correct Predictions }}{\text{ Total Number of Predictions}}$$

(7)

Measures the proportion of correct predictions among all predictions made by the model.

Precision is calculated by using the formula,

$$\text{Precision}=\frac{\text{ True Positives }}{\text{ True Positives }+\text{ False Positives}}$$

(8)

Measures the proportion of correctly predicted positive cases among all cases predicted as positive by the model.

Recall is calculated by using the formula,

$$\text{Recall}=\frac{\text{ True Positives }}{\text{ True Positives + False Negatives}}$$

(9)

Measures the proportion of correctly predicted positive cases among all actual positive cases.

F1-score is calculated by using the formula,

$$F1-\text{score}=2\times \frac{\text{ Precision }\times \text{ Recall }}{\text{ Precision }+\text{ Recall}}$$

(10)

Harmonic mean of precision and recall, providing a balance between the two metrics.

These evaluation metrics collectively offer comprehensive insights into the model's performance across different dimensions, aiding in the assessment of its effectiveness in remote patient monitoring applications.

Table 14 presents the accuracy and performance metrics of the system, exhibiting balanced precision, recall, and F1-score values along with high accuracy of 97.5%. Real-time monitoring performance metrics are shown in Table 15, which includes response time, abnormality detection accuracy, and quick action to guarantee timely intervention. The robustness of the model across various demographic groups is shown in Table 16, which also shows consistently high accuracy and performance metrics across age groups. Significant gains in healthcare outcomes are shown in Table 17, which includes a 20% decrease in hospital readmissions, 92% early detection rates, and high patient satisfaction ratings. Performance metrics are broken down by health parameters in Table 18, which displays excellent accuracy and precision across a range of vital signs and activity levels.

Table 14 Accuracy and performance metrics.

Table 15 Real-time monitoring performance.

Table 16 Model robustness and generalization.

Table 17 Impact on healthcare outcomes.

Table 18 Performance metrics by health parameter.

In order to ensure accurate and dependable monitoring across a variety of scenarios and populations, Tables 19 and 20 provide further detail regarding real-time monitoring performance and model robustness by action type and patient demographics, respectively. Table 21 evaluates the effect on healthcare outcomes for individual conditions and shows significant decreases in readmissions to hospitals as well as high rates of early detection for respiratory, diabetes, and cardiovascular diseases. Table 22 displays the developed model's superior accuracy, precision, recall, and F1-score values when compared to baseline models. Figures 3 and 4 give a graphical depiction of the efficacy and dependability of the system by visualizing performance metrics and real-time monitoring performance over time. Model robustness by patient demographics and comparison with baseline models are shown in Figs. 5 and 6, which provide information about the model's superiority over conventional methods and its ability to generalize. Additional model accuracy visualizations across health parameters and demographic groups are shown in Figs. 7, 8, 9 and 10 along with real-time performance trend tracking over time.

Table 19 Real-time monitoring performance by action type.

Table 20 Model robustness by patient demographics.

Table 21 Impact on healthcare outcomes by condition.

Table 22 Comparative analysis with baseline models.

Figure 3

Performance metrics by health parameter. Various performance metrics, such as accuracy, precision, recall, and F1-score, specifically analyzed across different health parameters. This analysis allows for an understanding of how well the model performs in identifying and classifying different health conditions based on specific physiological measurements.

Figure 4

Real-Time monitoring performance over time. The performance of a monitoring system, such as the RPMM (Remote Patient Monitor Model), evolves or changes over a period of time. This visualization typically shows metrics related to the system's monitoring capabilities, such as response time, abnormality detection accuracy, or prompt action accuracy, plotted against time intervals or timestamps.

Figure 5

Model Robustness by Patient Demographics. A representation or analysis of how resilient or stable a model is across different demographic groups of patients. It typically involves evaluating the performance metrics, such as accuracy, precision, recall, or F1-score, of the model across various demographic categories such as age groups or gender.

Figure 6

Comparative Analysis with Baseline Models. A comparison between the performance of a developed model and that of baseline models, which typically represent simpler or traditional approaches in the field. This analysis helps assess the effectiveness and superiority of the developed model by evaluating its performance metrics, such as accuracy, precision, recall, or F1-score, in comparison to baseline models.

Figure 7

Comparative Analysis of Model Accuracy. Indicates a comparison of the accuracy metrics of different models. This analysis involves assessing how accurate each model is in making predictions or classifications compared to each other.

Figure 8

Accuracy across Health Parameters and Demographic Groups. An examination of how accurate the model's predictions are across various health parameters and demographic categories. This analysis aims to understand if the model's accuracy varies based on different health conditions or patient demographics.

Figure 9

Real-Time Monitoring Performance over time. Indicates an assessment of how the system's monitoring performance evolves or remains consistent over a period. It likely involves tracking key performance metrics or indicators, such as response time, abnormality detection accuracy, or other relevant measures, to understand how effectively the system operates over time in real-world scenarios.

Figure 10

Impact on Healthcare Outcomes by condition. An analysis of how the proposed healthcare system affects specific medical conditions. It includes insights into the system's effectiveness in reducing hospital readmissions, detecting health issues early, and improving patient satisfaction for various conditions such as cardiovascular disease, diabetes, and respiratory disorders.

Finally, experimental observations are shown in Table 23, which includes vital sign measurements and activity levels for specific patients. This helps illustrate how well the system functions in real-world situations.

Table 23 Experimental observation.

Therefore, the remote patient monitoring system exhibits remarkable efficacy, resilience, and influence on healthcare results. It offers precise and prompt monitoring, which enhances patient care, lowers readmissions to hospitals, and raises patient satisfaction. Its superiority over conventional methods is further supported by comparative analysis, which makes it an invaluable tool for healthcare practitioners managing remote patients.

Complexity analysis

The computational complexity of the suggested approach is carefully examined. This analysis takes into account several parameters, including memory needs, computational resources (e.g., CPU, GPU), and model training and inference times.

• Experimental Setup:

  • Programming Framework: TensorFlow 2.0

  • Hardware:

    • CPU: Intel Core i7-8700 K (3.7 GHz, 6 cores)

    • GPU: NVIDIA GeForce GTX 1080 Ti (11 GB VRAM)

• Time complexity analysis:

  • ${T}_{\text{train }}=$ End time—Start time

  • ${T}_{\text{inference }}=\frac{{T}_{\text{total }}}{{N}_{\text{samples}}}$ where ${N}_{\text{samples}}$ is the number of samples.

  • Let ${T}_{\text{train }}^{\text{ensemble}}$ be the training time of the ensemble model.

  • Let ${T}_{\text{train }}^{\text{baseline}}$ be the training time of a baseline model.

  • Let ${T}_{\text{inference }}^{\text{ensemble}}$ be the average inference time per sample of the ensemble model.

  • Let ${T}_{\text{inference }}^{\text{baseline}}$ be the average inference time per sample of a baseline model.

  • Values:${T}_{\text{train }}^{\text{ensemble }}=10$ hours, ${T}_{\text{train }}^{\text{baseline }}=15$ hours.

  • Values: ${T}_{\text{inference }}^{\text{ensemble }}=0.05$ seconds, ${T}_{\text{inference }}^{\text{baseline }}=0.08$ seconds.

  • Ensemble Model: ${T}_{\text{train }}^{\text{ensemble }}<{T}_{\text{train }}^{\text{baseline}}$, indicating faster training.

  • Ensemble Model: ${T}_{\text{inference }}^{\text{ensemble }}<{T}_{\text{inference, }}^{\text{baseline}}$, indicating faster inference.

• Space Complexity Analysis:

  • Monitor ${M}_{\text{train}}$ and ${M}_{\text{inference}}$ using system monitoring tools or built-in memory profiling functions.

  • Let ${M}^{\text{ensemble}}$ be the peak memory usage during training or inference of the ensemble model.

  • Let ${M}^{\text{baseline}}$ be the peak memory usage during training or inference of a baseline model.

  • Values: ${M}^{\text{ensemble }}=8 \, \text{GB},{M}^{\text{baseline }}=10 \, \text{GB}$.

  • Ensemble Model: ${M}^{\text{ensemble }}<{\text{M}}^{\text{baseline}}$, indicating lower memory usage.

• Comparative Analysis.

The time complexity in seconds for the training and inference phases of several models, such as Random Forest, Support Vector Machine (SVM), standalone Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and their ensemble (LSTM + CNN), is shown in Table 24. For the identical models, Table 25 lists the space complexity in megabytes for both training and inference.

Table 24 Time complexity (values in seconds).

Table 25 Space Complexity (values in megabytes).

Table 26, however, offers a comparison of several metrics between solo CNN, standalone LSTM models, and ensemble approaches. The metrics include peak memory consumption in gigabytes, inference time per sample in seconds, and training time in hours.

Table 26 Comparative analysis.

When compared to the baseline models, the comparative analysis shows that the ensemble technique gives a significant gain in terms of both time and space complexities. In particular, the ensemble approach uses less memory and achieves faster training and inference times. The ensemble approach is a strong option for real-world applications with strict computational restrictions because of these computational benefits.

Ablation study

An ablation study was carried out in the framework of the proposed Ensemble LSTM-CNN Model for human activity recognition to examine the relative significance of the CNN and LSTM components as well as their combined ensemble architecture.

First, the reference model for comparison was the baseline model, which had CNN and LSTM layers in an ensemble arrangement. Then, three different ablation variations were looked at: one in which the LSTM layer was the only one removed, one in which the CNN layer was the only one removed, and a third in which the LSTM and CNN layers were both eliminated, leaving only the ensemble component. A validation dataset was used to assess each version, and performance metrics like accuracy, precision, recall, and F1-score were calculated for comparison.

The ablation study's findings provided fascinating new information on the relative contributions of the CNN and LSTM parts to the overall performance of the model. The significance of the synergistic integration of both architectures is demonstrated by the baseline Ensemble LSTM-CNN Model's consistent superior performance over all ablated variations across all performance parameters. It's interesting to note that performance fell more when the LSTM layer was removed than when the CNN layer was removed, indicating that the temporal data that the LSTM component was able to capture was important for activity detection tasks.

Moreover, the Ensemble Ablation variation, in which the CNN and LSTM layers were eliminated, performed quite poorly out of all the variations, highlighting the importance of the combined architecture in obtaining better outcomes. These results demonstrate how well-suited LSTM and CNN architectures are for capturing temporal and spatial data, respectively, and how well their ensemble fusion performs for reliable human activity recognition.

Table 27 shows a performance evaluation of several models in terms of accuracy, precision, recall, and F1-score. The models include a baseline model and several ablation models. The performance metrics for each model are displayed as percentages.

Table 27 Performance analysis.

In contrast, Table 28 presents a comparison of the variations in performance between the ablation models and the baseline model. When compared to the baseline model, it displays the percentage differences in accuracy, precision, recall, and F1-score. These tables offer insights into the efficacy and areas for improvement of model architectural alterations or variations by assessing how they affect performance measures.

Table 28 Comparison table.

Case study

• Problem statement:

  Remote patient monitoring (RPM) has become an essential tool for ongoing health surveillance in modern healthcare, particularly for patients with chronic diseases or those receiving postoperative treatment. Unfortunately, current RPM systems frequently struggle to analyze physiological data correctly and identify anomalies quickly, which could cause intervention delays and worsen patient outcomes.

• Proposed Solution:

  A group of researchers created an Ensemble LSTM-CNN Model for Remote Patient Monitoring (RPMM) in order to overcome these difficulties. The strengths of Long Short-Term Memory (LSTM) networks for temporal sequence modeling and Convolutional Neural Networks (CNN) for spatial analysis are combined in this novel model. With the use of both temporal and geographical elements taken from physiological data obtained from Internet of Things wearables, the RPMM seeks to deliver precise and fast insights into the health status of its patients.

• Methodology

  • Data Collection: Patients' IoT wearables are used to gather physiological data, such as heart rate, blood pressure, temperature, respiration rate, and activity levels.

  • Preprocessing: Missing value handling, normalization, and segmentation into fixed-length time-series sequences are some of the preprocessing procedures that are applied to the raw data.

  • Complete Model: The preprocessed data is supplied into the Ensemble LSTM-CNN Model, which is made up of temporal sequence modeling LSTM layers and spatial analysis CNN layers.

  • Training and Evaluation: Categorical cross-entropy loss and backpropagation using an Adam optimizer are two optimization techniques used to train the model. To guarantee accuracy and generalizability, performance is assessed using testing and validation datasets.

• Experimental Results:

  • Accuracy and Performance Metrics: The RPMM generates balanced precision, recall, and F1-score values, resulting in an accuracy of 97.5%.

  • Real-Time Monitoring Performance: The model shows a response time of 2.5 s, a high 95% abnormality detection rate, and a 98% accuracy rate for timely action.

  • Model Robustness: The model performs between 95.8% and 97.2%, maintaining a consistently high accuracy across various demographic groupings.

  • Impact on Healthcare Outcomes: The RPMM improves healthcare outcomes significantly, resulting in a 20% decrease in readmissions to hospitals, 92% early identification of health problems, and excellent patient satisfaction ratings.

• Conclusion and Future Directions:

  The case study demonstrates the Ensemble LSTM-CNN Model's efficacy in improving remote patient monitoring. Through precise real-time physiological data analysis, the RPMM facilitates prompt interventions and enhances patient outcomes. Subsequent investigations may concentrate on extending the model to encompass supplementary health metrics and investigating its comprehensibility for therapeutic implementation.

• Significance for Medical Professionals:

  Healthcare professionals can use the RPMM to improve the capabilities of remote patient monitoring, which will improve patient care, lower the number of readmissions to hospitals, and increase patient satisfaction. The model is a useful tool for managing patients with chronic diseases or those who need ongoing monitoring after surgery because of its accuracy and resilience.

Conclusion

The proposed model RPMM is an ensemble deep learning marvel that seamlessly combines CNN with LSTM in this ground-breaking investigation of healthcare technology. The RPMM system, which is specifically made for Internet of Things (IoT) devices, has been shown to have an outstanding accuracy rate of 97.23% when tested over a range of vital health parameters, such as heart rate, blood pressure, pulse, temperature, activity level, weight management, respiratory rate, medication adherence, sleep patterns, and oxygen levels. The model works well with Internet of Things wearables by utilizing CNN's spatial analysis and feature extraction capabilities in conjunction with LSTM's skillful temporal sequence modeling to collect physiological data in real time. Early diagnosis of cannabis is made easier by RPMM outstanding effectiveness in spotting trends and irregularities in vital sign data. This research represents a major advancement in the development of deep learning applications for the Internet of Things in healthcare, as well as a demonstration of the potential of ensemble models to improve timeliness and accuracy in health monitoring. Future research could focus on improving RPMM by adding more health parameters, examining its interpretability for clinical use, and looking into real-world data integration for ongoing model updates. This would guarantee RPMM flexibility and effectiveness in a range of healthcare scenarios and further advance RPM systems.

Future enhancement

A number of drawbacks in the suggested Ensemble LSTM-CNN Model could make it less useful and applicable in actual healthcare environments. First and foremost, there is a big problem with the model's limited generalization to different patient groups with different medical profiles and demographics. Future studies should concentrate on utilizing domain adaptation tactics and data augmentation approaches to improve the model's resilience across diverse patient cohorts in an effort to lessen this constraint. Another significant concern is the inherent lack of interpretability and explainability in deep learning models, such as the Ensemble LSTM-CNN. To overcome this obstacle, interpretability strategies like attention mechanisms or post-hoc explanation approaches like SHAP must be integrated in order to clarify the variables affecting model predictions and help physicians comprehend and have confidence in the model's judgments.

Moreover, the implementation of the suggested model in extensive healthcare systems may give rise to scalability concerns, which could hinder its ability to monitor real-time data. It is crucial to maximize model inference speed and computational efficiency using distributed computing frameworks and hardware acceleration in order to get past this obstacle. Furthermore, maintaining patient confidence and regulatory compliance in remote patient monitoring systems depends critically on data privacy and security. Safeguards against data breaches and privacy violations must include robust encryption, anonymization, and compliance with legal standards like GDPR and HIPAA.

Furthermore, even though the suggested approach shows potential in experimental settings, there are a lot of obstacles to overcome before it can be validated and used in actual clinical situations. Thorough clinical validation investigations, encompassing randomized controlled trials and usability evaluations, are essential for evaluating the model's practical efficacy, clinical relevance, and influence on patient results. Fostering acceptance and adoption of the suggested method within clinical practice requires early interaction with clinicians, collaboration with healthcare institutions and stakeholders, and resolving their views and concerns.