IoT-based bed and ventilator management system during the COVID-19 pandemic

Abstract

The COVID-19 outbreak put a significant pressure on limited healthcare resources. The specific number of people that may be affected in the near future is difficult to determine. We can therefore deduce that the corona virus pandemic’s healthcare requirements surpassed available capacity. The Internet of Things (IoT) has emerged an crucial concept for the advancement of information and communication technology. Since IoT devices are used in various medical fields like real-time tracking, patient data management, and healthcare management. Patients can be tracked using a variety of tiny-powered and lightweight wireless sensor nodes which use the body sensor network (BSN) technology, one of the key technologies of IoT advances in healthcare. This gives clinicians and patients more options in contemporary healthcare management. This study report focuses on the conditions for vacating beds available for COVID-19 patients. The patient’s health condition is recognized and categorised as positive or negative in terms of the Coronavirus disease (COVID-19) using IoT sensors. The proposed model presented in this paper uses the ARIMA model and Transformer model to train a dataset with the aim of providing enhanced prediction. The physical implementation of these models is expected to accelerate the process of patient admission and the provision of emergency services, as the predicted patient influx data will be made available to the healthcare system in advance. This predictive capability of the proposed model contributes to the efficient management of healthcare resources. The research findings indicate that the proposed models demonstrate high accuracy, as evident by its low mean squared error (MSE), root mean squared error (RMSE), and mean absolute error (MAE).

Introduction

Because of the potential for SARS coronavirus to spread, it is critical to separate healthcare services for COVID and non-covid patients. Treating both groups in the same facility could result in the infection spreading rapidly. However, this approach requires extensive testing with many individuals and the availability of hospitals with critical care infrastructure that can be turned into COVID facilities. The high dependency unit (HDU) and intensive care unit (ICU) should have more beds in the COVID facilities. As a result, ventilators, vital monitors, pulse oximeters, infusion pumps, non-invasive ventilators, and a continual supply of oxygen are in high demand. Aside from treating moderate to severe symptomatic patients, another important consideration is the infrastructure to reduce the burden on already overburdened healthcare facilities caused by asymptomatic or mildly infected cases, which account for nearly 80% of infected patients. All over the world, life expectancy has grown steadily and has been affected by this disease outbreak. In the last few decades, the COVID-19 pandemic has resulted in a significant increase in critically ill older people and other patients¹. Ensuring a good standard of life for older people has become a major social problem. Innovative healthcare innovations and resources are being enabled by the rapid proliferation of information and communication technologies, which promise to address the ex-listing issues. IoT has evolved as one of the most influential connectivity concepts of the twenty-first century and industry 4.0². Because of its interaction and computational capacities, all artifacts in our everyday lives are part of the web in the IoT world. IoT broadens and expands the definition of the Internet³. IoT enables seamless interactions between various types of devices, such as medical sensors, video surveillance, and household appliances⁴. As a result, IoT has become more productive in various sectors, including health services or healthcare systems.

IoT in healthcare refers to a variety of low-cost sensors (wearable, implanted, and environmental) that enable older people to access current medical healthcare services from anywhere, at any time⁵. Furthermore, it significantly increases the quality of life of elderly patients and other patients who need urgent treatment. One of the most imperative technologies is the IoT-oriented, modern health system in the body sensor network (BSN)⁶. It is essentially a series of low- and lightweight wireless sensor networks that control the functions of the human body and the environment. BSN nodes are used to gather confidential information (critical to life) and can function in extreme conditions in the healthcare system. BSN has made patient tracking possible by improving its use in healthcare applications. Several wireless health studies and initiatives have recently been suggested ensuring a continuous patient tracking process for the health monitoring system, in-ambulatory, in-clinical and open-air environment⁷. The body sensor network allows contact with the base station between invasive procedures. An on-body network of sensor systems, on the other hand, enables the communication between the wearable and the non-invasive sensors⁸. When a BSN Care server collects data from the local processing unit from an individual (which uses several bio-sensors), the BSN server then feeds and analyses the BSN data into the database⁹. Depending on the level of anomalies, it could then communicate with the family members, the local doctor, or even an emergency department of a regional health facility¹⁰. When a person (not necessarily a patient) wears multiple biosensors on their body, the BSN server receives regular updates via local processing units from these sensors¹¹. These data are used to find the severity in the patients and reflect the requirements of early-stage treatment of the patients; this leads to the need for the availability of beds and ventilators in that particular health care location.

Literature review

This section discusses the various literatures reviewed in the same field. The three parameters has been considered for the literature review analysis, and the same has been discussed in the Table 1, and these are High dimentional data, demographic information, and hospital admission management.As, in the context of COVID-19 healthcare management, high-dimensional data, demographic information, and hospital admission management play crucial roles in optimizing patient care and resource allocation. High-dimensional data refers to complex datasets with numerous features, such as medical history, symptoms, lab results, imaging scans, and real-time monitoring from IoT healthcare devices. This data helps in predicting disease severity, identifying high-risk patients, and improving treatment strategies using AI and machine learning. Demographic information, including age, gender, ethnicity, socioeconomic status, and pre-existing conditions, is essential for understanding vulnerable populations, tracking infection trends, and prioritizing vaccination efforts, which also adds the weitage into the dataset for getting better results. It enables healthcare providers and policymakers to design targeted interventions for different demographic groups. Hospital admission management, on the other hand, involves tracking bed and ventilator availability, ICU occupancy, and patient discharge planning. Efficient management of hospital resources during COVID-19 ensures that healthcare systems are not overwhelmed, improves patient flow, and enhances overall care delivery. AI-driven forecasting models can leverage these factors to predict patient surges, optimize resource allocation, and support data-driven decision-making in pandemic response efforts and also for deciding on many more healthcare-related queries. Table 1 presents the parameters selected for performing the literature review. Various machine learning and artificial intelligence techniques have been extensively utilized to analyze vast amounts of COVID-19-related data, enabling accurate predictions of disease evolution, patient risk assessment, and healthcare resource optimization. These advanced computational methods focuses on high-dimensional datasets, incorporating medical history, laboratory-tests findings, imaging results, and real-time patient monitoring to enhance diagnostic accuracy and treatment strategies. Additionally, AI-driven models assist in forecasting infection trends, identifying potential outbreaks, and optimizing hospital admission management to prevent healthcare system overload. A comprehensive discussion and detailed analysis of these methodologies, along with their impact on COVID-19 healthcare management, have been presented and systematically highlighted in Table 2.

Table 1.

Parameters selected for performing the literature review.

Parameter notation	Parameter name
P1	High dimentional data
P2	Demographic information
P3	Hospital admission management

Table 2.

Literature review on COVID-19 prediction algorithms.

Author	Year	P1	P2	P3	Pros	Cons
Liew et al.12	2025	Y	Y	Y	Reducing the number of features lowers computational costs, improves model efficiency, reduces overfitting, and enhances interpretability	To achieve minimal false positive rates
Xiang et al.13	2025	Y	Y	Y	Correlated the Diabetic disease with the COVID-19	The study’s small number of positive participants and lack of multicenter data limit its external validation and generalizability
Park et al.14	2025	Y	Y	N	The analysis of 1031 cases from two healthcare centers, incorporating 178 variables on demographics, clinical data, and medications were proposed	The modest sample size and low delirium prevalence hindered external and temporal validation, emphasizing the need for larger studies
Farahat et al.15	2024	N	N	Y	CT imaging will assess the three levels of respiratory support received by the patient	Improving lesion segmentation is needed, as the current method depends on radiologist-selected regions of interest
Elaraby et al.16	2024	N	N	Y	X-ray images are denoised using a median filter, and features from Speeded Up Robust Features (SURF), Grey-Level Co-Occurrence Matrix (GLCM), Histogram of Oriented Gradients (HoG), Segmentation-Based Fractal Texture Analysis (SFTA), and Local Binary Pattern (LBP) are extracted and fused for better diagnosis	The study uses chest X-ray images from 1124 patients (403 COVID-19, 721 healthy) to train a model for classification, though the data is imbalanced with a bias toward healthy cases
Cai et al.17	2024	Y	N	Y	This paper provides new recommendations on required blood test data for suspected COVID-19 patients	The challenge is handling imbalanced data and enhancing LSTM for irregular sampling intervals, with large-sample Omicron patient data
Solayman et al18.	2023	N	Y	Y	This work aims to predict COVID-19 infection status and raise awareness of individuals’ circumstances	A key limitation of this study is the use of an open-source dataset with patient data from a specific region
Li et al.19	2023	N	N	Y	Gradient Boosting Decision Tree (GBDT), Logistic Regression (LR), and simplified LR models were developed to predict the COVID-19 mortality	Fivefold cross-validation was used due to limited external data, and the models’ generalizability beyond Chinese patients remains uncertain
Hasan et al.20	2023	N	N	Y	This innovative work significantly reduced COVID-19 cases and deaths in Delhi-NCR	Data analysis from various other states of India has not been considered
Famiglini et al.21	2022	N	Y	Y	The proposed methods, especially the SVM model, showed strong predictive performance using just two demographics and a cost-effective CBC test	The challenge is externally validating the models with diverse hospital data to assess impacts of viral variants, treatments, and vaccines
Liu et al.22	2022	N	N	Y	The machine learning and deep learning in ensuring model generalizability, addressing overfitting and out-of-distribution were targeted	Generalizability in medical image analysis remains challenging due to variations in imaging centers, scanners, and dataset characteristics
Xu et al.23	2021	Y	Y	Y	Nineteen risk factors, including demographics, symptoms, comorbidities, and lab markers, were used to predict Acute Respiratory Distress Syndrome (ARDS) in COVID-19 patients	A challenge is ensuring the careful use of Artificial Intelligence (AI) in healthcare to provide reliable clinical decision support

Y = Yes and N = No.

Liew et al.¹² utilized Random Forest (RF), XGBoost, and AdaBoost on a selection of 14 parameters to enhance predictive modeling. Similary, in another research article by Xiang et al.¹³ applied SHAP (Shapley Additive Explanations) values to quantify the impact of each variable on predictions, with non-normal data reported using the median and interquartile range (IQR), while normal data were expressed as mean ± standard deviation (SD). Likewise, Park et al.¹⁴ developed and validated a Random Forest-based machine learning model to predict delirium in hospitalized COVID-19 patients.

A neural network-based fusion approach was introduced by Farahat et al.¹⁵, where Gibbs energy estimates from three Markov–Gibbs models were combined to optimize patient decision-making. Elaraby et al.¹⁶ designed a cloud-based platform leveraging modern AWS services to ensure high availability, elasticity, coverage, configurability, and reliability. Meanwhile, Cai et al.¹⁷ employed Deep Long Short-Term Memory (LSTM) models to capture dynamic changes in disease severity, clinical outcomes, and Spike antibody levels. In an effort to enhance interpretability, Solayman et al.¹⁸ incorporated Explainable Artificial Intelligence (AI) using the LIME (Local Interpretable Model-agnostic Explanations) framework to provide transparency in prediction results.

For classification tasks, Li et al.¹⁹ developed a Gradient Boosting Decision Tree (GBDT) and Logistic Regression (LR) model for binary classification. Hasan et al.²⁰ focused on forecasting COVID-19 trends using data acquisition techniques and the ARIMA model to predict active cases, positive case rates, and death rates. Additionally, Famiglini et al.²¹ demonstrated the effectiveness of the Support Vector Machine (SVM) model, which exhibited strong performance in discrimination, calibration, and utility.

To address segmentation, diagnosis, and prediction challenges, Liu et al.²² employed widely used Convolutional Neural Network (CNN) models, including ResNet, DenseNet, VGG, Inception, EfficientNet, U-Net, V-Net, and U-Net++. Lastly, Xu et al.²³ assessed model performance using tenfold cross-validation and external tests, where datasets were divided into 9 parts for training and 1 part for validation, ensuring robust evaluation.

These studies collectively contribute to the growing body of research aimed at utilizing AI-driven methodologies for improving COVID-19 diagnosis, prediction, and patient management.

Motivation and novelty

The coronavirus disease (COVID-19) outbreak has spread globally, posing a significant threat to global healthcare organizations. Due to this, the healthcare systems in India are facing lots of problems. As every work is done using a manual process. The Invention of IoT can prove to be a complete factor for the Indian healthcare organization. This motivated us to come up with a proposed scheme. The scheme monitors the patient’s current health recovery process without the physical presence of the human being. The scheme manages the total count of available beds and ventilators in the healthcare system. Also, it calculates how many patients can be accommodated in the healthcare organizations for better services, helping the elderly patients get managed treatment in the hospital.

Research contribution

The role of disease prevention and control is more difficult as the COVID-19 epidemic has spread across the world. In this paper, we present a technological framework for reacting to the COVID-19 epidemic from a health informative perspective to include theoretical and practical references to other countries on how health information technology can respond to the COVID-19 epidemic for various public health emergencies and disasters. IoT is used to identify the availability of beds and ventilators in healthcare organizations. Hence, the following are the contributions of this research paper based on the proposed framework IOT-BVM.

The proposed approach smoothly solved the allotment of beds in the healthcare organization. The availability of the beds supports the emergency services if the patient’s information is present in advance, informs the healthcare systems regarding their capacity, and makes the information available on demand.The IoT-based framework and its analysis system will quickly emanate messages to the required person to decide on a time. The performance of the proposed scheme is validated against the mentioned methodologies.

Organization

In this section, the introduction of the research article has been presented, encompassing an in-depth discussion of the literature review, motivation, and contributions. “Methodology” section focuses on the methodology, providing a comprehensive overview. Subsequently, “Proposed framework and its analysis” section elaborates on the proposed architecture, including its flowchart and algorithms, followed by a discussion of the experimental results. In “Conclusion” and “Future work” sections, the conclusion and future directions have been addressed, concluding with the references.

Methodology

IoT-based bright pulse oximeter: A physician can track a patient remotely using an oximeter connected to an Internet of Things (IoT) device. This acts as a real-time warning system that can warn patients if they’re in danger. IoT-based body temperature measurement device (temperature sensor): The unit is small and convenient, and it runs on a 3.3 V battery cell.

We can forecast a time series using the ARIMA model by looking at the past values of the series. A time series is a set of measurements taken at regular intervals. Forecasting is the next step in the process, and it involves predicting the series’ future values. The fundamental business planning, sourcing, and development activities are all guided by forecasting²⁴. Any forecasting errors can affect the entire supply chain or any business context. As a result, it’s essential to get the predictions right to save money and achieve success. A time series can now be divided into two forms of forecasting²⁵. Univariate Time Series Forecasting is when you use only the past values of a time series to estimate future matters. Multi-Variate Time Series Forecasting is when you use predictors other than the series (also known as exogenous variables) to forecast. ARIMA, which stands for “Auto-Regressive Integrated Moving Average,” is a classification method that “explains” a particular time series based on its previous values, i.e., its hick-ups and stagnated. Prediction error, results in an equation that can be used to forecast future information. An ARIMA model is represented by 3 terms: a, b, and c. Where a= the AR term’s sequence, b= the MA term’s sequence and c= the number of differences that must be present in order for the time series to become stationary.

A pure Auto Regressive (AR only) model is one in which X_t is solely determined by its own lags. X_t, in other words, is a feature (or function) of the lags of X′t.

1

where X_t1 is the series’ lag1, β is the lag1 coefficient calculated by the model, and α is the intercept term estimated by the model.

A pure Moving Average (MA only) model, on the other hand, is one in which X_t is solely determined by the lagged forecast errors.

2

The errors of the autoregressive models of the corresponding lags are represented by the error terms. The errors ε_t and ε_t−1 are the results of the Equations (3) and (4).

3

4

AR and MA were the models of the ARIMA. So here’s what an ARIMA model’s equation looks like. An ARIMA model is one in which the time series has been differenced at least once to make it stationary, and the AR and MA terms have been combined. The equation is presented in Equation (5).

5

The second model, which has been utilized for predicting future patient admissions, has been computed and analyzed using the Transformer model. The implementation process has been systematically carried out, ensuring that the model’s parameters and configurations are optimized for accurate predictions. Furthermore, the effectiveness of the Transformer model has been evaluated based on its ability to process sequential data and capture complex patterns within the dataset. It has the following layers:

• Data preparation and embeddings The original Transformer design relies on embeddings to represent discrete tokens. For a time-series application, numerical values are normalized and then projected into a continuous embedding space. A linear transformation is frequently used to map each time-step value into an embedding vector. Positional information, which is critical for sequence modeling, is preserved by adding positional encodings. These encodings impose an ordering on the input data, ensuring that temporal relationships are recognized by the model.

• Encoder module The encoder is composed of multiple layers, each containing a multi-head self-attention mechanism and a position-wise feed-forward network. Within each encoder layer, self-attention computes the relevance of each time step to every other time step in the input sequence. This operation is facilitated by query, key, and value projections, which are combined to form attention scores. Multi-head attention extends this process by allowing multiple sets of projections to capture different types of temporal dependencies simultaneously. After the attention sublayer, a feed-forward network is applied to each embedding vector independently. Residual connections and normalization are included to stabilize training and retain information from earlier layers.

• Decoder module (Adapted for Regression) In a standard Transformer, the decoder generates outputs step by step for language modeling tasks. For time-series regression, the decoder can be configured to receive a shifted version of the target sequence (e.g., previous time steps or placeholder values) to predict subsequent values. The decoder follows a structure similar to the encoder but incorporates an additional cross-attention mechanism. This mechanism uses the encoder outputs as keys and values while the decoder states serve as queries, allowing the model to focus on relevant information from the encoded sequence. The final output is often produced by a linear layer that maps the decoder’s hidden states to numeric predictions rather than discrete tokens.

• Training objective Since the goal involves predicting continuous patient counts, a regression loss function such as mean squared error or mean absolute error is employed. Backpropagation is used to adjust the parameters within the encoder-decoder blocks, including attention weights, feed-forward layers, and embedding transformations. Gradient-based optimization algorithms, such as Adam, are commonly utilized to achieve stable convergence.

• Temporal attention and long-range dependencies A primary advantage of this architecture lies in its ability to capture long-range dependencies without relying solely on recurrent connections. By computing attention scores across all time steps simultaneously, the model can learn complex temporal relationships in patient count data, which may include seasonal trends or abrupt changes. Multi-head attention further refines these relationships by enabling the model to attend to different aspects of the sequence in parallel.

• Inference and forecast generation Once the model has been trained, forecast values are produced by feeding the most recent observed data (or previously predicted outputs) into the encoder-decoder structure. The decoder iteratively predicts subsequent time steps until the desired forecast horizon is reached. In practice, teacher forcing or scheduled sampling can be applied during training to balance stability and predictive accuracy. The transformer model architecture has been discussed in the result section.

Proposed model and algorithms

The world’s current COVID-19 pandemic requires a patient safety policy that places a greater focus on epidemiology, particularly identifying effective population-based developmental and learning programs and addressing the consequences²⁶. The healthcare facilities and programs are still in the early stages of growth, with skill shortages, sickness absence, insufficient facilities, and low-quality care among the challenges^27,28. Even though the government’s determination and the National Health Mission, sufficient and sustainable healthcare remains a mirage²⁹. The rural healthcare system is beset by a persistent shortage of medical practitioners, which negatively affects the overall accessibility of care for rural residents³⁰. In certain places, the healthcare infrastructure is insufficient or unprepared to deal with the pandemic like COVID-19 transmission^31,32. To deal with such issues^33,34 and to solve the problem, the proposed scheme has been implemented in this research paper.

The utilization of IoT devices significantly contributed to patient monitoring during the COVID-19 pandemic. Smart sensors were employed to continuously measure vital health parameters such as body temperature, oxygen levels, heart rate, and respiratory function. Data collected through wearable devices and remote monitoring systems were transmitted to healthcare professionals, enabling timely medical responses. By facilitating remote assessments, these technologies helped reduce direct patient contact, thereby lowering the risk of virus transmission and their will not be any direct contacts with the patients. Cloud-based platforms were used for data storage and analysis, supporting predictive insights and early detection of health deterioration. AI-driven algorithms were integrated into IoT systems to identify irregularities and notify medical personnel, ensuring swift intervention. Additionally, automated medication dispensers and intelligent inhalers were implemented to enhance treatment adherence. IoT-enabled tracking mechanisms further assisted in contact tracing and monitoring quarantine compliance. As a result, hospital resources were effectively managed, and patient care was significantly enhanced throughout the pandemic³⁵.

Figure 1 depicts the architecture of the IoT-based healthcare management system for the benefit of the patients and as well as for the hospital management system. Here the patient’s health is being monitored using the body area sensor network. For the sake of simplicity, only parameters are being considered here, i.e., the patients’ body temperature and oxygen level through the IoT devices. This information is then passed to the Local Process Unit (LPU). After that, the data is passed via the Internet to the various stakeholders of the healthcare management systems, such as such emergency ward, physician, and care center of the BSN, where the parameters related to the patient’s details are maintained and observed for any credulity of the patients.

Fig. 1.

Architecture of the proposed IOT-BVM scheme.

The illustrated architecture is composed of several layers, each intended to manage patients who use IoT devices during the COVID-19 period. In the first layer, body sensors are placed on individuals to collect vital signs. The readings are then received and filtered by a local processing unit. Next, the data are transferred via the internet to a central repository, where they are stored in a database. Critical cases are flagged, and positive results are tracked automatically. This information is then shared with healthcare services, such as emergency departments, physicians, and care centers, to enable ongoing monitoring.

One notable advantage of this layered design is its ability to provide real-time updates on patient health, which supports faster intervention and more informed clinical decisions. In addition, the need for in-person visits can be reduced, lowering the potential for virus transmission. The architecture can also be scaled up if the number of patients increases, offering flexibility in times of high demand. However, certain drawbacks must be acknowledged. A reliable internet connection is required for seamless data flow, and any disruption can affect timely updates. Moreover, patient privacy must be protected through robust data security measures³⁶. Costs related to installing and maintaining sensors, databases, and other infrastructure can also be considerable. Despite these challenges, the layered approach continues to offer valuable support for remote patient care and broader public health initiatives. Fig. 1 shows the architecture of the proposed IOT-BVM scheme.

The working of the proposed scheme, IOT-BVM, is expressed in flowcharts 1 and 2. Figure 2 depicts the working mechanism of flowchart 1 and the operations of the process that is being undergone inside the hospital management. Initially, the details of the patients are identified who are presently admitted to the hospital. The patients are being monitored using the IoT-based environment, and it is being checked whether or not the temperature and oxygen level are in the correct range. If it’s not proper, continuous monitoring will be done.

Fig. 2.

Flowchart of the proposed system.

If the monitored information results in the details that indicate that the patients are being recovered and if they have completed the 14 days of the isolation period, then they are to be in the process of getting discharged from the hospital, and its information must be passed to the stakeholders of the hospital system. The outcome of this algorithm is 1 the total number of available beds and is denoted as “V”, The algorithm 1 represents herein.

The time complexity of the algorithm 1 can be analyzed by looking at the steps involved in the procedure. Let n represent the number of hospitalized patients. In Step 3, retrieving the details of the patients is a constant-time operation, O(1), since it’s just an input data retrieval. In Step 4, fetching the IoT-based body temperature (T) and oxygen level readings (O) for each patient requires iterating through each patient’s data. This operation takes O(n) time, where nn is the number of patients, because it involves fetching and storing data for all patients. Step 5 consists of checking each patient’s body temperature and oxygen level. The comparison of these values is a constant-time operation, O(1), for each patient. If a patient meets the specified conditions (temperature < 99°F and oxygen level > 96%), the algorithm either discharges the patient or continue’s monitoring based on whether the isolation period has passed. This decision-making process also takes constant time, O(1), per patient. If the conditions are not met, an alarm is generated and the hospital management is notified, which is again an O(1) operation. Since these steps are executed for each patient, Step 5 has a time complexity of O(n). Finally, Step 6 involves repeating the evaluation process until all patients are discharged. Since each patient is processed only once, the overall time complexity of this loop remains O(n). Thus, considering all steps together, the total time complexity of the algorithm is O(n), where nn is the number of hospitalized patients. This indicates that the algorithm’s execution time increases linearly with the number of patients, making it a highly efficient solution for managing patient data in real time.

The system performance of this algorithm can be evaluated based on its efficiency, responsiveness, and scalability. In terms of efficiency, the algorithm performs well because it operates in linear time, O(n). Each step of the procedure involves essential operations, such as retrieving data, comparing patient conditions, and triggering the necessary actions based on those comparisons. As the number of patients increases, the system’s performance remains proportional to the data size, making it well-suited for managing larger datasets without excessive delays. Regarding responsiveness, the algorithm is designed to make decisions promptly. It evaluates each patient’s condition and, based on the results, either discharges the patient, continues monitoring, or raises an alarm. This ensures that the hospital management system can respond quickly to any changes in a patient’s condition, improving patient care and hospital workflow. The ability to react immediately is crucial for real-time healthcare applications, where fast decision-making can directly impact patient outcomes.

In terms of scalability, the algorithm is capable of handling an increasing number of patients without significant performance degradation. Since the time complexity is linear, the system scales efficiently with the number of patients. For very large datasets, the processing time will increase proportionally, but the algorithm will still function adequately as long as the hospital’s infrastructure is equipped to manage the additional data. Finally, in terms of resource utilization, the algorithm relies mainly on data retrieval and condition checks, both of which are low-complexity operations. This results in minimal resource consumption, making the system efficient in terms of CPU and memory usage. However, as the dataset grows, additional infrastructure might be needed to handle the increased data processing load. Hence, the algorithm offers efficient performance with linear time complexity, ensuring that it can handle the increasing scale of patient data without significant delays. Its quick decision-making process enhances responsiveness, and its scalability makes it suitable for use in larger hospitals or healthcare facilities. Overall, the system is designed to perform effectively in real-time environments, balancing efficiency, scalability, and resource usage.

Algorithm 1.

The procedure inside the hospital.

Flowchart 2 is the subpart of the proposed scheme and is represented in Fig. 3. Figure 3 depicts the scenario where the comparison of the available beds (V) and the positive cases and its prediction of the Confirmed cases (A) are taken into consideration in that particular region. Then this is compared to get the idea that the vacant beds are available in enough away so that the patients can be admitted and will be given good services by the healthcare systems. In case the available beds are not present, then the nearby hospital vacancies beds data is to be identified so that the patients can be accommodated in the hospitals; the same is depicted in Algorithm 2.

Fig. 3.

Flowchart depicting the process inside the hospital and the confirmed cases around the same region.

Algorithm 2.

The procedure inside the hospital and the positive cases around the same region.

The outcome of the mentioned approach solves the problem of managing the patients in this pandemic situation. The services can be improved as the data will be available in advance, and the hospital management will have enough time to arrange the physicians, medical devices, staff, and vacant beds. This will also aid the healthcare systems to be prepared in advance with the availability of medical equipment. The following section discusses the performance evaluations of the proposed scheme.

The time complexity of the algorithm 2 can be analyzed by evaluating the operations involved in each step. Let n represent the number of critical COVID-19 patients and mm represent the number of available vacant beds in the region. In Step 3, retrieving the details of the COVID-19 patients for a specified region is an O(1) operation, assuming the data is readily available. Step 4 involves assigning the number of critical patients (denoted as A) from the retrieved data, which is a simple assignment and thus takes constant time, O(1). In Step 5, the algorithm compares the number of critical patients A with the available vacant beds VV. This comparison is a constant-time operation, O(1). If the number of critical patients is less than the available beds, the algorithm allocates beds by admitting the patients to hospitals in the region, which again takes constant time, O(1). However, if the number of critical patients exceeds the available beds, the algorithm needs to identify hospitals with vacant beds, which requires scanning the list of hospitals. This step takes O(m) time, where mm is the number of hospitals. Once the hospitals with available beds are identified, assigning patients to these hospitals takes O(1) per patient. Finally, Step 6 involves repeating the process until all patients are discharged. Since each patient is evaluated once, this loop runs O(n) times. Given that the algorithm may require scanning all hospitals in the worst case, the overall time complexity of the algorithm is O(n⋅m), where n is the number of critical patients and mm is the number of hospitals. This time complexity suggests that as the number of critical patients and hospitals increases, the algorithm’s execution time will scale accordingly.

The system performance of this algorithm can be assessed in terms of efficiency, responsiveness, and scalability. In terms of efficiency, the algorithm performs well for smaller datasets, as most operations are constant-time, O(1), including retrieving data and assigning patients to hospitals. The primary area where performance may degrade is when identifying hospitals with available beds, which requires scanning the list of hospitals. As the number of hospitals grows, this step can become a bottleneck, increasing the overall time complexity to O(n⋅m). Therefore, while the algorithm is efficient for smaller datasets, its performance may decrease with a larger number of hospitals or patients. In terms of responsiveness, the algorithm ensures quick decision-making when allocating beds to critical patients. If the number of critical patients is less than the number of vacant beds, they can be admitted to the available hospitals in the region, which ensures that patients are admitted promptly. If there are more critical patients than vacant beds, the algorithm identifies nearby hospitals with available beds and directs patients accordingly. This ensures that the system is responsive and that no critical patient is left without attention, making it highly suitable for real-time applications in healthcare settings.

Regarding scalability, the algorithm can handle a growing number of patients, but its scalability is affected by the number of hospitals in the region. As the number of hospitals increases, the algorithm’s performance may suffer due to the need to scan a larger list of hospitals for available beds. This means that while the algorithm is scalable to a degree, it may require optimizations, such as indexing hospitals or clustering them based on location or availability, to improve efficiency when the number of hospitals or patients grows significantly. Lastly, in terms of resource utilization, the algorithm is generally efficient, as it primarily performs basic comparisons and assignments. However, the process of identifying vacant beds in hospitals can be resource-intensive, especially if the list of hospitals is large. As the dataset grows, the system may need more resources to manage the increased computational load. Therefore, the algorithm is suitable for smaller-scale applications, but for larger datasets, it might require infrastructure improvements or optimization techniques to ensure optimal performance. Hence, while the algorithm offers an efficient and responsive solution for allocating beds to critical patients, its time complexity of O(n⋅m) makes it less efficient as both the number of patients and hospitals increases. The system performs well for smaller datasets but may face challenges with scalability and resource usage when handling larger volumes of data. To ensure better performance at scale, optimizations such as data indexing or improved searching algorithms could be considered.

Proposed framework and its analysis

The world’s current COVID-19 Pandemic requires a patient safety policy that places a greater focus on epidemiology, particularly identifying effective population-based developmental and learning programs and addressing the consequences³⁶. The healthcare facilities and programs are still in the early stages of growth, with skill shortages, sickness absence, insufficient facilities, and low-quality care among the challenges. However, the government’s determination and the National Health Mission that adequate and sustainable healthcare remain a mirage. Keeping these things in the mind, the said approach is proposed. In this section, the performance evaluation has been discussed and is as follows. Figures 4 and 5 represents the conventional (X-ray image to see the impact of the infection) methods to identify the criticalness of the patient’s health. But every time doing the X-ray of the target patients is not good for their health. Hence, monitoring the parameters such as oxygen level and temperature is another solution implemented using IoT-based devices and theories. IoT (Internet of Things) can also be used to further automate the process of hospital management and patient monitoring³⁷. Using wearable body sensors, e.g., pulse oximeters and temperature sensors, patients’ vital signs can be monitored continuously in real time³⁸. The sensors send data wirelessly to a central gateway, which processes and aggregates the data before sending it to a decision-support system. The cloud system then runs the decision logic described, e.g., checking if the patient’s temperature is below 99°F and oxygen levels above 96%. If these are the criteria, the system automatically checks if the 14-day isolation period has elapsed. After confirmation, it sends a discharge notification, updates the hospital bed management system to indicate the newly released bed, and informs the concerned staff.

Fig. 4.

Number of rising cases of the COVID-19.

Fig. 5.

Epoch-wise reduction in MAE for the LSTM model.

If the vitals of the patient exceed the safe threshold, the IoT system automatically raises an alarm, alerting hospital management through SMS, email, or dashboard alerts. This is an automated process that saves time wasted in manual checks, reduces human error, and allows timely intervention. Further, the merging of IoT data with electronic health records (EHR) and hospital management systems allows enhanced resource planning and trend analysis. Overall, IoT deployment automates the entire process, allowing better patient care, operational efficiency, and bed turn-over rates without compromising on high accuracy for monitoring critical health parameters. Table 3 below indicates the various aspects of the IoT based patients monitoring system.

Table 3.

IoT-based patient monitoring and bed allocation.

Aspect	Details	Benefits/outcome
IoT devices used	Wearable sensors (pulse oximeters, temperature sensors)39	Real-time monitoring of vitals
Data transmission & gateway	Wireless data sent to central gateway, which processes and forwards to decision-support system	Seamless data aggregation and decision-making40
Decision logic	Check if temperature < 99°F, oxygen > 96%, and 14-day isolation completed	Enables automatic discharge decisions
Post-discharge actions	Sends notification, updates bed availability, alerts staff	Efficient bed management and communication
Emergency alerts	Alarms raised if vitals are critical; alerts via SMS, email, dashboard	Timely intervention and reduced manual monitoring
System integration	Integrated with EHR and hospital management systems	Enhanced resource planning and trend analysis
Automation impact	End-to-end automation of monitoring and decision-making processes41,42	Reduces human error, saves time, and improves patient care
Overall outcome	Use of IoT for monitoring, alerting, and process optimization43	Better operational efficiency and accurate critical parameter monitoring

The results are based on the historical data taken through the KAGGLE. Figure 6 shows the plot of patients who were admitted to the hospitals. The dataset utilized in this research was sourced from Kaggle and provides a comprehensive view of key COVID-19 metrics, offering crucial insights for understanding the progression and impact of the pandemic. The dataset contains several time-series plots that capture important trends over time, each serving a specific purpose in analyzing the dynamics of the pandemic. One of the key visualizations is the plot of patients admitted to hospitals, which illustrates the fluctuations in hospital admissions throughout the observed period. This plot is essential in understanding how healthcare facilities were impacted as the number of cases surged.

Fig. 6.

Epoch-wise reduction in RMSE for the LSTM model.

Additionally, the dataset includes a plot of recovered cases, which provides a clear picture of the recovery trends and helps to evaluate the effectiveness of treatment and healthcare interventions during different phases of the pandemic. A plot of the death rate within the hospital is also included, offering critical insights into the severity of the disease and the challenges faced by healthcare systems. Moreover, the dataset tracks the total test results and the increasing number of cases, showing the relationship between testing efforts and the rise in reported cases. This is particularly important for understanding how testing infrastructure scaled and its role in detecting the spread of the virus. Finally, the dataset presents a plot of COVID-19 positive cases, reflecting the overall trajectory of the virus in the population and providing an essential context for understanding how the pandemic evolved over time. Together, these time-series plots offer a rich dataset for analyzing the various facets of the COVID-19 pandemic, and they form the foundation for our research, enabling a detailed exploration of trends, correlations, and patterns within the data.

The preprocessing of the dataset was carried out using several essential steps to ensure its suitability for machine learning (ML) and deep learning (DL) models. First, we began by handling missing data. For any rows with missing values, we applied imputation techniques, such as mean or median imputation, to replace missing numerical data. Rows with excessive missing information were removed to maintain the dataset’s integrity. Categorical features were transformed using one-hot encoding to convert non-numeric values into a format compatible with machine learning algorithms. Next, the focus has been given on normalizing numerical features. We used Min–Max scaling to ensure that all numerical values fell within a uniform range (0 to 1), preventing any one feature from disproportionately influencing model training due to differences in scale. Given that the dataset includes time-series data, it was critical to preserve the chronological order of the observations. We formatted the data to ensure it accurately reflected the temporal dependencies in the dataset. For time-series forecasting, we used a sliding window approach to create sequences of past data points to predict future outcomes. This allowed the models to capture the dynamic trends over time, such as hospital admissions, recoveries, and death rates. Feature engineering played a key role in enriching the dataset. We created lag features, capturing past values of vital variables like patient admissions and recoveries, which help the models account for historical trends and dependencies. Additional temporal features, such as day-of-week or month, were also generated to capture any seasonal patterns. Finally, the dataset was split into training and testing subsets, typically using an 80/20 split. K-fold cross-validation was implemented to evaluate the models robustly, ensuring they were not overfitting to any particular subset of the data. By preprocessing the dataset in these ways, we ensured that it was ready for modelling.

Results and discussions

Test bed setup for experimental evaluation The experimental evaluations were conducted on the environment equipped with an Intel(R) Xeon(R) CPU E5-2620 v2 operating at 2.10 GHz, featuring 6 physical cores and 12 logical processors. The system was supported by 128 GB of DDR4 memory. Storage was configured with five 300 GB drives. For GPU-based computations, an NVIDIA Tesla K80 graphics card with 128 GB memory capacity was utilized. The dataset used here was downloaded from the publicly hosted COVID-19 time series data by the Center for Systems Science and Engineering (CSSE) at Johns Hopkins University, URL: https://github.com/CSSEGISandData/COVID19/tree/master/csse_covid_19_data/csse_covid_19_time_series. Cumulative confirmed United States cases from the dataset were extracted. After preprocessing, there was data aggregation by date, and the cumulative case counts were converted to daily new cases by differencing two consecutive days’ values. Through this, analysis of the spread of COVID-19 over time was made in a more defined manner. Preprocessing was then followed by a split of the dataset into the training and the testing set by an 80/20 ratio. A sliding window approach was used for the generation of sequence, where for each input, there were 10 consecutive days of daily new case counts in an attempt to predict the next day’s value. These were reshaped accordingly to fit into the input parameters of deep learning models. The model development was centered on three deep learning architectures: LSTM, GRU, and Transformer. Hyperparameter tuning was used for each of the models in order to provide optimal performance, and evaluation included metrics such as Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Loss.

The LSTM model was constructed with 64 units of memory, dropout of 0.2 to prevent overfitting, and trained for over 50 epochs with the Adam optimizer at learning rate 0.0001. The Mean Squared Error (MSE) as the loss function was employed. The GRU model also employed a similar structure, swapping the LSTM layer for a GRU layer but utilizing the same hyperparameters such as the number of units, dropout rate, optimizer, and loss function. Mechanisms with 4 attention heads that permit the simultaneous attention of different components of the input sequence. A light Transformer was created with the hyperparameters tuned including the number of layers, embedding size, and dropout rate. Training was conducted for 50 epochs using the Adam optimizer and MSE loss function. Throughout the training phase, there was a consistent drop in loss over epochs, and the models were exhibiting increasingly better performance. After training, the models were tested on the test set specified. The Transformer model demonstrated better results on all metrics of evaluation, closely followed by the LSTM and GRU models. Visual plots were created to represent the intercorrelations of the evaluation metrics. Graphical plots of MAE vs. Loss, RMSE vs. Loss, and RMS vs. Loss revealed similar trends, justifying the seen hierarchy of performance between the models. Moreover, values for the x-axis and y-axis were easily identified and shown for each graph to enhance interpretability. The thorough evaluation pointed out that deep learning models, particularly the Transformer model, hold enormous potential in forecasting trends in COVID-19 cases, demonstrating considerable enhancements over conventional sequential models.

Discussion on Hyperparameter Selection and Tuning In the LSTM model, a number of hyperparameters were selected with care to maximize performance and avoid overfitting. In particular, the usage included the utilization of a total of 64 LSTM units, which allowed for a balance between learning capacity and computational cost. This number of units was empirically selected to provide sufficient memory capacity without making the model excessively complicated. In another effort to enhance generalization, a dropout rate of 0.2 was utilized, where 20% of the neurons were randomly dropped out during training, thereby minimizing the chances of overfitting. The model was trained for 50 epochs, a time period which was selected based on observation of the trend in loss, such that this number of iterations allowed for sufficient accommodation of temporal patterns without inducing overfitting. A batch size of 32 was utilized, as it was discovered to offer the best balance between computational cost and training stability. The Adam optimizer, with a learning rate of 0.0001, was utilized due to its adaptive learning characteristics and ability to deal with sparse gradients, features highly sought after in time series forecasting activities. Model optimization was conducted using the Mean Squared Error (MSE) loss function, which is generally known for suitability in regression tasks, where large prediction errors were given a proportionately larger penalty. The same hyperparameters were used in the GRU model for consistency and to allow fair comparisons. The setup of the GRU network was 64 units, a reflection of the structure of the LSTM network because the same number has been found to offer an optimal balance between training time and model performance. Another 0.2 dropout rate was applied to prevent overfitting. Training was carried out for 50 epochs, with a batch size of 32, to allow for uniform updating of the gradients and efficient memory management. The Adam optimizer, using a learning rate of 0.0001, offered fast convergence and responsiveness to the changing learning needs of the GRU model. Like the LSTM network, the MSE loss function was used to reduce the gap between the real and predicted values, with accuracy in regression tasks in mind.

For the Transformer model, some modifications were made to accommodate its architecture, the same has been depicted in Fig. 4. The model used 4 attention heads, which allowed it to attend to various sections of the sequence data simultaneously, thereby capturing complex interdependencies well. Two multi-head attention layers were utilized, balancing between depth and computational tractability. A hidden size of 128 was chosen to offer sufficient representational capacity, enabling the model to learn complex patterns without overfitting. A dropout rate of 0.2 was applied uniformly to further generalize the model. Training was carried out for 50 epochs, and a batch size of 32 was utilized, just like in the recurrent models. The Adam optimizer with a learning rate of 0.0001 facilitated efficient and adaptive optimization during training. MSE was once again used as the loss function to maintain uniformity across models as well as to appropriately penalize large prediction errors using PyTorch or TensorFlow can be provided for time-series forecasting with transformers.

Discussions of the result achieved:

The performance of the deep learning models—LSTM, GRU, and Transformer—was evaluated based on their Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Loss values after training and testing over the COVID-19 daily new cases dataset and the same has been depicted in the Fig. 5, 6, 7, 8, 9, and 10. The figigure indicates the values as, the LSTM model achieved a MAE of 4213.76, a RMSE of 6050.09, and a corresponding Loss value of 78,426,017.38. These values suggest that while the LSTM was reasonably effective in capturing the temporal dependencies of the COVID-19 daily new cases, there were still notable prediction errors. Mathematically, the RMSE being higher than the MAE indicates that larger errors occurred occasionally because RMSE penalizes larger errors more than MAE. Similarly, the GRU model achieved a MAE of 4580.59, a RMSE of 7061.71, with a Loss value of 91,656,112.95. The GRU’s MAE and RMSE were slightly higher than the LSTM’s, which suggests that while the GRU was able to model the temporal structure to some extent, it underperformed relative to the LSTM model. The difference between MAE and RMSE for GRU also points toward the occurrence of significant errors on some days, negatively impacting the model’s reliability. In contrast, the Transformer model achieved a MAE of 4018.71, a RMSE of 5838.42, with a Loss value of 67,516,675.40. These numerical results demonstrate a superior performance in comparison to both the LSTM and GRU models. The MAE was the lowest among the three models, indicating that the Transformer made the smallest errors on average. Furthermore, the RMSE for the Transformer was also the lowest, highlighting that the Transformer was not only minimizing average errors but also ensuring that extreme prediction deviations were fewer compared to LSTM and GRU.

Fig. 7.

Epoch-wise reduction in MAE for the GRU model.

Fig. 8.

Epoch-wise reduction in RMSE for the GRU model.

Fig. 9.

Epoch-wise reduction in MAE for the transformer model.

Fig. 10.

Epoch-wise reduction in RMSE for the transformer model.

Mathematically, the following observations has been observed:

MAE comparison:.

Substituting the actual values: 4018.71 < 4213.76 < 4580.59.

RMSE comparison:.

Substituting the actual values: 5838.42 < 6050.09 < 7061.71.

This mathematically validates that the Transformer model yielded consistently lower errors both in terms of MAE and RMSE. In addition, considering the Loss values, which indicates the Mean Squared Error (MSE), the Transformer again achieved a minimum value (67,516,675.40), followed by LSTM (78,426,017.38) and GRU (91,656,112.95). A lower loss confirms that the model’s predictions were closer to the actual values on average.

Visual Interpretation of Results The plots of MAE vs. Loss and RMSE vs. Loss further corroborated the numerical observations. In these graphs, the Transformer’s data points were consistently located lower than those of LSTM and GRU models, visually indicating better performance. The MAE vs. Loss graph showed a near-linear trend for the Transformer, implying that its error reduction was consistently aligned with minimizing loss. Similarly, the RMSE vs. Loss plot depicted that the Transformer maintained lower root mean squared deviations compared to other models. The fine mini-grid style used in the graphs enabled a clearer visualization of minor variations in performance, and the x- and y-axis values extracted demonstrated a visible gap where the Transformer’s metrics outperformed others.Thus, based on the mathematical analysis of MAE, RMSE, and Loss, and the visual evidence from the plotted graphs, it can be conclusively proven that the Transformer model provided the most accurate and reliable forecasts for daily COVID-19 new cases. This superiority can be attributed to the Transformer’s ability to capture complex long-range dependencies in the data without the sequential bottlenecks that hinder recurrent architectures like LSTM and GRU. The same has been depicted in the Table 4

Table 4.

Analysis of the results obtained using the various metrics.

Metric	LSTM	GRU	Transformer	Trade-off Observations
MAE	4213.76	4580.59	4018.71	Transformer achieves the lowest MAE, indicating smaller average errors, followed by LSTM and then GRU
RMSE	6050.09	7061.71	5838.42	Transformer outperforms both LSTM and GRU in RMSE, with GRU showing the highest error variance
Loss	7,84,26,017.38	9,16,56,112.95	6,75,16,675.40	Transformer has the lowest Loss, indicating the most accurate and consistent model. LSTM performs better than GRU
Error variance	Higher due to higher RMSE	Higher due to higher RMSE	Lower due to lower RMSE	Transformer shows fewer extreme errors, making it the most reliable model. LSTM and GRU both show significant error spikes
Model complexity	Moderate (Sequential)	Moderate (Sequential)	High (Self-attention mechanism)	Transformer’s complexity might be a trade-off for better performance, while LSTM and GRU are simpler but less effective

To statistically compare the performance of the deep learning models (LSTM, GRU, and Transformer) with regard to the Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Loss, a pair-wise comparison through a statistical test such as the paired t-test or Wilcoxon signed-rank test can be done. The test would identify if the differences in the measures of performance between the models that are being observed are statistically significant or by chance. Because these models are tested on the same dataset, a paired t-test would be suitable to compare their performance on each of the metrics.First, the null hypothesis (H₀) can be stated as “there is no significant difference between the MAE, RMSE, and Loss values of the models,” and the alternative hypothesis (H₁) would be that “there is a significant difference between the performance of at least one model.” The paired t-test would then compare the MAE, RMSE, and Loss between the models pairwise (LSTM vs. GRU, LSTM vs. Transformer, and GRU vs. Transformer) for each measure. Since the Transformer model performed better than both LSTM and GRU in all three measures (lower MAE, RMSE, and Loss), the outcome of these tests will most likely have large differences, particularly between the Transformer and the other two models. The Statistical significance can be tested using a p-value cutoff (usually p < 0.05). If the p-value is less than this value for a particular metric comparison, the null hypothesis can be rejected, meaning that the performance difference is statistically significant. For instance, if the Transformer-LSTM comparison has a p-value less than 0.05 for the MAE or RMSE, we can safely conclude that the performance of the Transformer model is considerably superior to LSTM in terms of prediction accuracy and error reduction. Additionally, effect size measures like Cohen’s d may also be computed to quantify the size of the difference in performance. The larger the value of Cohen’s d, the larger the difference in model performance would be, thereby substantiating the argument that the Transformer model performs better than LSTM and GRU. According to the numeral discrepancies shown in the MAE, RMSE, and Loss values, the statistical testing and effect size calculation provided strong evidence to affirm the performance superiority of the Transformer model in estimating COVID-19 daily new cases. Visual evidence from the graphs shown, based on which the Transformer consistently outperforms, is that which confirms these statistical findings and supports the ability of the model to tackle intricate dependency in the data.

Challenges The findings obtained through this study bring forth a number of challenges from the perspective of model performance and interpretation. As much as the Transformer model outperformed both the LSTM model and the GRU model by way of MAE, RMSE, and Loss, an observation was made that there were huge errors committed in the prediction, especially where the models were dealing with complex patterns in the COVID-19 daily new cases dataset. The LSTM model, though efficient in handling temporal dependencies, had greater MAE and RMSE values than the Transformer, which implies that greater prediction errors were experienced during certain periods. This indicates that the ability of the model to generalize over various trends was compromised, resulting in deviations from true values. The GRU model also suffered from the same problems, with greater MAE and RMSE than the LSTM, indicating that GRU was unable to capture the temporal dynamics with the same precision. The vast disparity between the MAE and RMSE for LSTM and GRU also indicated that sporadic sheer errors were severely punished by the RMSE, which affected their overall usability. Such models, limited by their sequential natures, could have struggled with learning long-distance dependencies without experiencing large errors when sudden changes in the data, like spikes or sudden drops, were present. The Transformer model, however, although uniformly demonstrating better performance compared to all other models, also struggled to reduce prediction errors to a degree that would be deemed negligible in real-world use. Regardless of its capacity for capturing long-distance dependencies and the lower Loss figures, even predictions by the Transformer showed fluctuations which could potentially undermine real-world forecast accuracy. Such challenges are multiplied by the erratic and complicated nature of the data on the COVID-19 pandemic, making precise prediction especially troublesome. In addition, while the graphical results do indeed show convincing outperformance of the Transformer model, it is a challenge to interpret and graph results across various measures (MAE, RMSE, and Loss) and thus ensure the most important performance measures are emphasized. These challenges necessitate further tuning of model designs and further development of different approaches, like ensemble learning or composite models, for even better, more precise prediction of dynamic datasets like COVID-19 cases.

Conclusion

In this research paper, the prediction of new COVID-19 cases on a daily basis was tackled with the implementation of deep learning models, i.e., LSTM, GRU, and Transformer models. Following appropriate preprocessing of the dataset and input sequence preparation, the models were created, trained, and tested using standard performance measures such as MAE, RMSE, and Loss. The findings show that the Transformer model outperformed the LSTM and GRU models in all evaluation metrics. The capacity to learn complex temporal relationships without the constraint of processing sequential data was a significant strength. Visual examination of the models’ performance also corroborated the findings, with the Transformer recording lower error values and showing improved generalization on new data. It can be concluded from the experimental findings that Transformer-based models offer a very efficient solution to time series forecasting problems such as COVID-19 case prediction. The findings show that future research in epidemic forecasting can benefit a great deal from the use and extension of Transformer models.

Future work

This section highlights the future work on COVID-19 datasets and also highlights the privacy and data security challenges in healthcare IoT frameworks.

• Integration of Heterogeneous Data Sources ⁴⁴

Future efforts must focus on merging data from diverse origins—such as hospital electronic health records, public health databases, social determinants, and real-time IoT sensor feeds. This integration enables comprehensive analyses, facilitates cross-validation of insights, and supports more accurate predictive models. By combining structured and unstructured data, researchers can capture a fuller picture of COVID-19 dynamics, ultimately leading to improved public health responses.

• Standardization of Data Collection Protocols ⁴⁵

Uniform data collection methods are critical to ensure consistency and interoperability. Establishing standard operating procedures across healthcare institutions minimizes variability in data quality. Consistent protocols allow datasets from various regions and sources to be compared effectively, reducing errors during data aggregation and enabling robust multi-center research collaborations.

• Advanced Anonymization Techniques ⁴⁶

Given the sensitive nature of health data, it is imperative to develop sophisticated anonymization methods that safeguard individual identities while preserving data utility. Techniques such as differential privacy, k-anonymity, and synthetic data generation can help maintain confidentiality. Future work should focus on balancing data utility with privacy, ensuring that analytical insights remain valid without exposing personal information.

• Real-Time Analytics Frameworks ²⁸

Implementing frameworks capable of processing and analyzing data in real time is crucial for timely decision-making during health crises. These systems would monitor outbreak trends as they develop, allowing public health officials to deploy resources more effectively. Future frameworks should support streaming data inputs from various sources, thus enabling rapid response strategies during emerging outbreaks.

• Interoperability of IoT Devices ⁴⁷

Healthcare IoT ecosystems consist of numerous devices—from wearable monitors to smart sensors—each potentially using different communication protocols. Ensuring that these devices can seamlessly share data is essential. Future research should promote the development of universal standards and middleware solutions that bridge protocol differences, thus fostering an integrated and responsive healthcare network.

• Scalable Data Infrastructure ⁴⁸

As COVID-19 datasets grow in volume and complexity, the underlying infrastructure must scale accordingly. Cloud computing and distributed systems can support high-volume data processing and storage. Future infrastructure development should focus on flexibility and scalability, ensuring that both current and future data needs are met without compromising performance or reliability.

• Enhanced Predictive Modeling ⁴⁹

Machine learning and statistical models are vital for forecasting outbreak trends and resource requirements. Future work should refine these models to incorporate evolving data streams and emerging viral variants. Continuous model validation and updates are necessary to ensure that predictions remain accurate over time, thus enhancing preparedness for future waves or pandemics.

• Secure Data Sharing Protocols ⁵⁰

Facilitating the secure exchange of data between institutions is paramount for collaborative research. Developing robust protocols that encrypt data during transmission and authenticate users can help prevent unauthorized access. Future initiatives should work towards establishing standardized, secure data sharing agreements that protect sensitive information while fostering innovation and collaboration.

• Privacy-Preserving Data Mining ⁵¹

Privacy-preserving techniques—such as federated learning and secure multiparty computation—allow multiple entities to analyze data collectively without the need for centralization. These methods minimize the risk of large-scale breaches by keeping raw data at its source. Future work should focus on refining these approaches to ensure they are both secure and effective in yielding actionable insights.

• Robust Encryption for Data Protection ⁵²

Protecting healthcare data at rest and in transit requires state-of-the-art encryption standards. Employing strong encryption algorithms and regularly updating them in response to emerging vulnerabilities is essential. Future initiatives should also explore end-to-end encryption protocols specifically tailored for IoT networks to ensure that sensitive information remains confidential.

• Dynamic Consent Management Systems ⁵³

Patient consent for data usage should be adaptable to changes over time. Dynamic consent models allow individuals to modify their data sharing preferences as circumstances or regulatory frameworks evolve. Future systems must integrate user-friendly interfaces and secure mechanisms for consent management, ensuring that patients retain control over their personal health information throughout its lifecycle.

• Real-Time Intrusion Detection ^37,54

With the growing number of connected devices, healthcare IoT networks are increasingly vulnerable to cyberattacks. Real-time intrusion detection systems (IDS) can monitor network traffic, identify suspicious behavior, and alert administrators instantly. Future efforts should prioritize the development and deployment of IDS that are optimized for the unique characteristics of healthcare environments.

• Blockchain-Based Data Integrity ^41,55

Blockchain technology offers a decentralized approach to recording data transactions in an immutable ledger. By incorporating blockchain, healthcare systems can verify the authenticity and integrity of data, reducing the risk of tampering. Future applications may include tracking data provenance and ensuring that each transaction is securely recorded, thereby enhancing overall trust in the system.

• Compliance with Global Privacy Regulations ⁵⁶

Data handling practices must adhere to international privacy regulations, such as GDPR, HIPAA, and other regional standards. Achieving and maintaining compliance can be challenging due to differing requirements across jurisdictions. Future work should focus on developing adaptive compliance frameworks that automatically adjust to regional regulatory changes, ensuring that privacy standards are consistently met.

• Continuous Risk and Vulnerability Assessments ⁵⁷

Cyber threats evolve constantly, necessitating regular vulnerability assessments and risk evaluations. Ongoing security audits can identify weak points within the system and prompt timely countermeasures. Future initiatives should implement automated risk assessment tools that continuously monitor system integrity and flag potential security breaches before they occur.

• Ethical Integration of Artificial Intelligence ⁵⁸

The application of AI in healthcare requires a strong ethical foundation to prevent biases and ensure equitable treatment of all patient groups. Integrating ethical guidelines into AI development can help mitigate risks associated with discriminatory practices. Future work should emphasize transparency, fairness, and accountability in AI systems, ensuring that these technologies serve all communities without prejudice.

• Adaptive Security Measures Against Evolving Threats ⁵⁹

The cybersecurity landscape is dynamic, with new threats emerging regularly. Security systems must be designed to adapt rapidly to these changes. Future security strategies should include machine-learning-based threat detection and automated patch management systems, ensuring that defenses are continuously updated to counteract novel attack vectors.

• Ensuring Data Accuracy and Validation ⁶⁰

Reliable analysis hinges on the accuracy of the underlying data. Ensuring data quality involves rigorous validation processes, regular audits, and the use of error-detection algorithms. Future work should focus on developing standardized validation protocols and employing cross-verification techniques to maintain the integrity of COVID-19 datasets.

• Cost-Effective Security Implementations ^37,61

While robust security measures are essential, they must also be economically viable—especially for smaller healthcare providers with limited budgets. Future research should explore cost-effective solutions that do not compromise on security. This may include open-source security tools, cloud-based security services, and scalable security architectures that can be tailored to the financial constraints of various institutions.

• User Education and Training ⁶²

The human factor remains one of the most significant vulnerabilities in data security. Regular training programs can equip healthcare personnel with the necessary skills to recognize and mitigate security threats. Future initiatives should develop comprehensive education campaigns that emphasize best practices in data handling, cyber hygiene, and incident response, thereby reducing the risk of human-induced breaches.

• Edge Computing for Reduced Latency ⁶³

Processing data at the edge of the network—closer to where it is generated—can significantly reduce latency and enhance the speed of critical decision-making. Edge computing enables real-time data analysis, which is particularly valuable in time-sensitive healthcare scenarios. Future work should investigate integrating edge devices with centralized systems to create a hybrid architecture that balances speed and data centralization.

• Resilient Backup and Recovery Systems ⁶⁴

Ensuring data continuity in the event of system failures or cyberattacks requires robust backup and recovery protocols. These systems must be designed to restore data quickly and accurately, minimizing downtime and preventing data loss. Future efforts should focus on creating resilient backup architectures that incorporate redundancy, regular testing, and automated recovery processes.

• Adoption of Standardized Security Frameworks ^40,65

The implementation of industry-recognized security frameworks—such as ISO/IEC 27001 or NIST—can provide a systematic approach to managing data security risks. Standardized frameworks ensure that best practices are followed consistently across all levels of an organization. Future work should encourage the adoption of these frameworks to streamline security processes and foster a culture of continuous improvement in risk management.

• Cross-Sector Collaboration on Data Security ⁶⁶

The complexity of healthcare IoT systems often necessitates collaboration between public institutions, private enterprises, and technology vendors. Such cross-sector partnerships can drive the development of innovative security solutions and promote shared learning. Future initiatives should focus on creating collaborative networks that facilitate knowledge exchange, joint research projects, and the pooling of resources to tackle common security challenges.

• Continuous Monitoring and Auditing Mechanisms ^42,67

Ongoing surveillance of data access and usage is essential for early detection of anomalies and ensuring regulatory compliance. Automated monitoring systems, coupled with periodic audits, can help identify suspicious activity and trigger prompt remedial actions. Future work should invest in advanced auditing tools that leverage artificial intelligence to continuously scan for deviations from established security norms, thereby maintaining the integrity of healthcare IoT ecosystems.

Author contributions

Vivek Kumar Prasad: Conceptualization, Methodology, Software, Writing- Original draft preparation Debabrata Dansana: Data curation, Visualization, Investigation. S Gopal Krishna Patro: Investigation, Visualization, Investigation. Ayodeji Olalekan Salau: Data curation, Methodology, Writing- Reviewing, Editing and Validation. Divyang Yadav: Data curation, Visualization, Investigation. Brojo Kishore Mishra: Validation.

Funding

Authors declare no funding for this research.

Data availability

The datasets generated during and/or analysed during the current study are not publicly available but are available from the corresponding author on reasonable request.

Declarations

Competing interests

The authors declare no competing interests.