An Internet of Things-Enabled Patient Monitoring Framework Using Hybrid CNN–LSTM Networks for Real-Time Healthcare Diagnosis

In DL methods, mainly the amalgamation of CNNs and Long Short-Term Memory (LSTM) systems has become increasingly important in recent years for healthcare applications, principally for monitoring the patient’s health status and suggesting diagnoses. The key goal of merging these two algorithms is to combine the strengths of individually and every one's strength: CNNs are uncountable at detecting spatial features, whereas LSTMs are better at capturing how data differ over time. Chhowa et al. [1] presented a DL framework to discover approaches for tracking and assessing health conditions by means of collective medical big data obtained from IoT systems. Contempt development of machine learning models in underachieving subdivisions has proven inadequately accurate when dealing with large-scale IoT-generated medical information, often requiring manual feature extraction.

Veeraiah et al. [2] developed an IoT-enabled framework using collaborative DL transfer aimed at the early detection of COVID-19. It should enable instantaneous identification of theoretically high-risk patients and provide corresponding precautionary alerts. The proposed architecture, powered by multiple DL models, significantly aided radiologists in swiftly and accurately diagnosing COVID-19 patients. Additionally, the system integrates a COVID-19-compliant functional identity mechanism compatible with IoT protocols.

Ahmad et al. [3] introduced a non-invasive technique for 24-hour cardiac monitoring, recognized as ambulant electronic cardiography, which continuously tracks the heart’s electrical activity and identifies risk issues for heart diseases. Their smart heart rate observing and monitoring system, driven by a DL based cardiac algorithm, estimates the risk of subsequent cardiac arrest by analysing real-time heart rate patterns, thereby enabling earlier intervention and increasing recovery chances.

Hembram et al. [4] proposed a secure data transfer method within an IoT ecosystem where device-generated information is encoded by means of the PDH-AES algorithm before cloud transmission. Once decrypted, a Deep Learning Multilayer Neural Network (DLMNN) classifier categorises the data into ‘normal’ or ‘abnormal’ classes, providing physicians with a snapshot of a patient’s cardiac health and issuing alerts when anomalies are detected. Their findings showed that DLMNN outperformed earlier diagnostic approaches, and the PDH-AES method achieved a 95.87% success rate while offering faster encryption and decryption compared to traditional AES.

Deepa et al. [5] proposed a Real-Time Face Mask with Healthiness Screening (RTFMHS) system for pandemic situations. The model uses a high-speed, multi-level augmented CNN for real-time face mask detection on IoT nodes.

Elbagoury et al. [6] focused on predicting early signs of stroke—such as paralysis, numbness, vision loss, disorientation, and speech impairment—as well as the risk of sudden death. Scoring the possibility of mobile health (mHealth) applications in hazardous circumstances similar to this kind of submission in remote site intelligence for stroke diagnosis remains immature. This research proposes the combination of a hybrid intelligent analytic device hooked on mHealth systems. This tool would exploit neural networks that have been trained with Clustering Method of Data Handling shortened with Clustering Method of Data Handling (CMDH), as well as Sparse Auto-Encoder-based DL methods, to improve the detection of heart strokes.

Table 1 above represents the different DL models in terms of assistance, indication of performance, and limitations of various models in the aspect of IoT-enabled healthcare monitoring systems in the medical industry. On the other hand, Table 1 looks at the various advantages and disadvantages as well as the pros and cons of various DL systems for broad-minded patient monitoring and diagnosis systems. Islam et al. [7] also proposed a comparative study of different DL algorithms that contain the advantages and disadvantages of different disease detection using IoT devices.

The feasible method founds in originate over the method of merger of IoT devices combined with DL approaches can provocatively grow the health monitoring of patient previously likewise do the diagnostic procedures. Figure 1 demonstrates that physical body-attached sensor data is chiefly composed together from devices similar to ECG images, investigative imaging equipment, temperature sensing strategies, and heartbeat sensors. The procedure involves assembling and uploading this live information based on the construction envisioned at the well-being examination. At the heart of this method lies a Joining Entryway, which acts as a node of dominant and aggregated information, collecting all linked 24-hour care strategies and earlier interactive to a cloud-based infrastructure for extra dispensation. Following its transmission through the joining Entryway, the information is resolutely warehoused, pre-processed, and enhanced in a cloud-based scheme. Formerly, the DL model splits the dataset into training and testing subsections to assist with prototypical expansion and accuracy testing. The algorithm's roadmap for analysing medical information is founded on the amalgamation of CNNs and LSTM systems, which is shown in Figure 2. In this proposed system, the CNNs generate analytical images and detect spatial topographies that are noteworthy for the detection of the patient's diseases. The LSTM, on the additional sideways advent at time-series data and discovers the time patterns in essentials that are repeatedly existence experiential corresponding heart rate and blood pressure. Figure 2 energies into extra feature about the model training pipeline, which comprises stages similar important the input shape, establishing the convolutional layer, flattening the output, and preparing the LSTM input. Later, these portions are assembled to form the full model. The LSTM's memory cells and exchange mechanisms to retain and discard relevant substantial over time is significant portion of this model. This capability to recall information over time is very noteworthy for careful real-time monitoring systems and additional dependable diagnostic calculations. The suggested system is based on the amalgamation of CNN and LSTM layers, which enable it to monitor and analyse data over time, as shown in Figure 2. This technique pledges ongoing health intensive care and fast judgment by analyzing real-time figures from the IoT devices, which leads to earlier health interventions. The integrative technique supplements the medical patient healthcare system's scope to classify embryonic health risks promptly and reply efficiently, leading to healthier medical penalties and a transformative effect on care delivery [7, 12, 13]. These systems are excellent at discovery bizarre belongings and patterns in complex visual data, which assistances discovery illnesses initial happening. CNNs use their convolutional and pooling layers to extract the high-level diagnostic topographies that medical images retain, which makes the model considerably more precise and vigorous [14]. At the similar period, LSTM nets, which remain exposed in Figure 2, are designed to examine consecutive designs in patient data that originate from IoT devices. They are very good at detecting the health propensities that precede bad things happen, while they can consider and recall noteworthy time-related data.

The important health metrics collected with heart rate discrepancy and blood pressure predispositions are continually checked to enable real-time predictive analytics inside the system [15, 16]. The synergetic mixture of CNNs and LSTMs permits the proposed algorithm to process both spatial and temporal data streams, easing complete and precise real-time health valuations [17, 18]. IoT systems allow for incessant, efficient data monitoring and examination, providing secondary healthcare providers with prompt, timely data for care delivery.

Hyperparameter values dictionary in the proposed DL model is presented in Table 2. The hyperparameters were combined with various types of configurations to find the model that will produce excellent results in training and predicting accuracy. The hyperparameters of filter size in the convolutional layer (32, 64, 128), (64, 128, 256), (128, 256, 512) produced the best results, which were (64, 128, 256). The dropout rates of 0.2, 0.4, and 0.8 were tested in the model to reduce overfitting, but the best result at which point the model generalised was 0.4. The batch sizes of 128, 256, and 512 were also tested, and the best for efficient training of the model was 512. The learning rate of 0.1, 0.01, or 0.001 was changed while training the model, where 0.001 provided the most stable convergence points and the fastest accuracy improvement in the model. Also, when testing the configuration of (128, 256), (256, 512), (512, 256) as neurons, the best representation and learning of features was (512, 256) with respect to each of the configurations. In relation to how long the model was trained (20) - 20 epochs were the most efficient for training a model. Finally, the Adam optimizer was the most efficient and converged at the fastest rate. Table 3 offers a comparison of simulation consequences and the accuracy of the deliberate CNN + LSTM model with additional DL procedures like SVM, LSTM, etc. The datasets used in these assessments comprise well-established scientific and physiological cascades such as MIMIC-III [19], PhysioNet [20], the eICU Concerted Database [21], MIMIC X-ray by chest [22], PTB ECG Database [23], and the broader eICU Clinical Database [24]. Presentation contrasts are complete by means of normal calculation metrics—F1-score, recall, precision, and accuracy—in the way of dimension of each model’s efficiency in real-world IoT-driven healthcare submissions. The final goal is to recover patient diagnosis and specialist care through intelligent, data-driven methods allied with the developing landscape of IoT-enabled healthcare.

Table 2. Proposed method hyperparameter values

Table 3. Comparison of different models with the proposed model

For the purpose of the analysis presented in this project, all data were drawn from publicly available sources: namely, the MIMIC-III database, the PhysioNet database, and the eICU database. The MIMIC-III dataset is comprised of large-scale EHR data that were developed using records of patients admitted to the Beth Israel Deaconess Medical Centre’s intensive care unit (ICU) during the years 2001 through 2012. This dataset contains information from 40,000+ patients who were admitted to that hospital's ICUs during that time period. By contrast, the eICU dataset contains records from approximately 139,367 patients who have been admitted to ICUs in hospitals throughout the United States during calendar years 2014 and 2015. Both datasets include extensive clinical information such as demographic information, laboratory values, vital signs, medication-related documentation, procedure documentation, and clinical narrative notes related to patient treatment. Prior to any analysis of simulated versus actual patient care pathways, efforts were taken to improve the quality of the data by removing and/or addressing any issues with missing values or duplicate records and irrelevant attributes. The data have been pre-processed to normalize the formats of individual clinical variables and converted into a standardized feature (also referred to as a feature vector) that is suitable for use as input into machine learning and DL models. Specific preprocessing pipelines have been developed as well to facilitate feature extraction/engineering, outlier detection, the aggregation of related clinical variables, and converting temporal records of patient care into a format that is appropriate for accurate predictions and comparisons of simulated and actual care pathways in the dataset. In a healthcare scenario enabled by the IoT, the assessment metrics demonstrate how well each model container identifies and retains the pathways of patients. The CNN and LSTM hybrid technique performed unusually well on the MIMIC-3 dataset for reaching an F1-score and accuracy of 0.99, and a recall of 0.98, illustrating its prognostic control. The standards exhibit high trustworthiness in monitoring and identifying patient circumstances as rapidly as by means of this collective method. These results demonstrate that the CNN and LSTM model lengthways by extra DL approaches vessel assistance to make more precise and fast therapeutic decisions on an assortment of datasets. The CNN model performed well on the health dataset, achieving 0.93 for F1-score, accuracy, and recall. Formerly these consequences display the model the ration healthy for examining medical images to detect and track diseases.