Design of an intelligent IoT enabled healthcare responsive framework for emergency scenarios

Abstract

Timely assessment and response to critical health scenarios are important for survival of patients. Smart wearables help in non-interrupted patient tracking whereas advanced intelligent models enhance early risk detection. But nowadays, heavy road traffic is causing delays in arrival of ambulance service thereby decreasing emergency service efficiency. Existing frameworks either address patient health monitoring or traffic control in a separate manner. Thus, a model which can integrate risk analysis and adaptive traffic management for ambulance service is lacking. The aim of this research is to design an intelligence based responsive health model for patients needing emergency help by tracking vital metrics with an advanced risk predictive model. Real time traffic support is also desirable to reduce ambulance service delay. The framework consists of a smart wristband ‘BioTrace-G’ to collect patient’s vital signs. This data is sent to the patient’s smartphone where an application ‘E-response’ is configured. The application hosts GA-DNN (Genetic algorithm-Deep neural network) model used for feature optimization and critical risk level prediction. When the detected risk is high or mid type, emergency ambulance service is automatically triggered which is supported by a traffic unit to facilitate faster emergency service. The model upon evaluation recorded a promising outcome. The mean risk prediction accuracy with GA-DNN was 95.2% in context to sensor readings while it is 94.2% when number of patients are considered. The computed mean inference latency was only 57.8 s. Also, the GA-DNN generated the least mean false negatives and false positives of 6.9% and 13.4% respectively. The framework optimized the patients prioritization and ambulance dispatch delay as compared to conventional approach. The model with integrated traffic support showed better results when validated against metrics like response delay, number of signal stops and ambulance speed. Hence, the integrated responsive framework serves as a prototype for early risk identification and categorization with reduced response delay and enhanced patient care.

Introduction

Nowadays, issues related to smart connected health is a universal challenge¹. The burden on medical centers can be significantly reduced if vital health physiological metrics like pulse rate, body temperature and body humidity can be appropriately dealt with at the right time. An advanced medical system should make provision for adequate and emergency healthcare facilities for patients anywhere and anytime in a user-friendly manner. With relevance to modern healthcare service, the capability to facilitate quick and efficient emergency response is important to reduce mortality rate and improve patient care. Any delay in detecting critical health related issues may lead to further complications. Traditional emergency systems are primarily sensitive to fluctuations and depend upon manual interference for making decisions. It may cause deviations in patients priority and inconsistent resource usage. In recent times, rise in wearables and intelligent transport mechanisms have generated curiosity in emergency medical service. The rapid use of sensors based applications has made it feasible to incorporate automation and communicating interfaces². With this, physical entities can get connected any time, and any place thereby can detect themselves to other commodities³. Global Positioning System (GPS) is also an enhancement which makes use of it⁴. Patients can be tracked online and doctors can directly contact patients continuously, and patients’ health parameters can be easily maintained⁵. In spite of enhancements in responsive healthcare, delays still happen because of restricted and unreliable data with inadequate management between healthcare system and traffic control staff. Pulse rate, body temperature and body humidity are the most responsive vital metrics. Over the years, smart health devices like fitness bands and smartwatches have gained a lot of popularity as they aim to monitor these health parameters. Though the user can keep track and maintain a log of it, the full-duplex nature of the health care device is not met as it is of little help to users in case of emergency health discrepancies. There is now a substantial need for an intelligent and smart system that can track different health metrics of a user, detect health disorders and provide medical services in real-time scenarios and emergencies. So, a responsive healthcare model can be of great help. The need of a responsive model to deal with medical emergencies related to vital health metrics arises due to following factors:

• Extensive delays in ambulance service is one reason for avoidable mortality cases such as risks related to heart attacks or strokes. So, if the severity level of patients can be predicted and support from traffic units can be provided then ambulances can be routed effectively with reduction in response time.

• In majority cases, patients fail to recognize their risk severity level at the time of need of medical service. Wearables can help with continuous monitoring of vital metrics and detecting fluctuating trends thus facilitating preemptive emergency notification so that users can be made aware of his health status.

• Using medical emergency service without proper priority may overload the healthcare resources at the centers creating more delay for health critical patients. Thus, an optimized predictive model can process vital metrics and estimate risk level of patients prompting early action.

• Human based judgement may cost the life of patients in case of emergencies. Here, a predictive intelligence driven model optimizes the accuracy in predicting risks making it more reliable.

Thus, a converged integration of machine intelligence with Internet of Things (IoT) can transform the emergency medical system by providing real time, information driven and intelligence based health response service. As compared to traditional models, this convergence speeds up the health crisis detection, permits robust health monitoring, automated decision making, facilitates alert notification to emergency units more reliable and supports adaptive response mechanisms with least delays. The research study makes use of this convergence approach for emergency health risk level assessment of patients thereby applying wearable based health analysis, predictive modelling with cloud computing. It coordinates with medical ambulance service and road traffic units to enable real time patient care.

The aim of the research is to collaborate the medical facilities with the traffic control system in case someone faces serious health issues due to fluctuations in vital signs and is unable to reach out for help. Thus, the user does not need to contact the support instead, medical support will reach out to him. The study designs a robust and lightweight wristband ‘BioTrace-G’ that integrates sensors to monitor vital signs of patients and also tracks their location through the inbuilt GPS of their smartphone. Vital signs include pulse rate, body temperature and body humidity of the patient. The acquired sensor data are processed using a microcontroller implanted in the band. The processed data is transferred to the user’s smartphone in case of any fluctuations observed through wireless communication. A connected mobile application ‘E-response’ captures the data and processes it using an optimized trained deep learning model GA-DNN (genetic algorithm-deep neural network). The model predicts the risk level of the patient as either ‘high’, ‘mid’ or ‘low’ category based on the aggregated vital signs from sensor data. For high risk patients, it generates alert notification to emergency contacts, nearest traffic control service and nearby medical centers for emergency ambulance service. The medical centers as well as the traffic units are well coordinated with the application upon login. The traffic staff and nearby hospitals are able to trace the location of the patient through the GPS and Google map. Upon receiving notification, traffic units set the signals green along the route of the ambulance and if heavy congestion is found in the path then the ambulance can be redirected to alternate routes by passing the traffic information to medical centers.

The main contributions of the research study are summarized as:

• The research presents a novel system framework to integrate advanced wearables, smartphone based predictive analytics and intelligent transport model for responsive, effective and emergency health service to patients in real time.

• A smart lightweight wrist band ‘BioTrace-G’ is designed which is able to track a patient’s vital signs - mainly pulse rate, body temperature and body humidity periodically. The accumulated data is transmitted to the user’s connected smartphone to detect health risk level at prior.

• An optimized GA-DNN model connected by the user’s smartphone application ‘E-response’ and trained in cloud is used to optimize and precisely categorize the risk level of a patient as high, mid or low using the vital signs data.

• Further, based on detected risk level, an automated emergency is activated which enables dispatching of ambulance service at nearby medical centers and interfacing with local traffic units for dynamic control over traffic signals which leads to reduction of response delay.

• Upon implementation, the model shows promising results when evaluated with different performance metrics like accuracy, inference delay, precision, recall, f-score, false positives and false negatives among others.

The paper is divided into several units. Section 1 introduces the need and motivation for emergency healthcare in today’s scenario thereby highlighting the main contributions of the work. Section 2 discusses the relevant existing works undertaken in the domain along with identified research gaps. Section 3 elaborates the proposed model and its constituents in detail. Section 4 presents the implementation outcomes in tabular and graphical representations. Section 5 summarizes the findings and concludes the work.

Related works

Several types of work concerned in the healthcare sector by the application of IoT have been done in the past using different technologies. This section consists of a brief description of the works undertaken in the concerned area. Muhammad et al.⁶ devised a system that can be used to measure ECG signals for home health monitoring using convolutional neural networks. Almotiri et al.⁷ presented an application that can collect real-time data from several wearable devices and store the data in a network server. This information is useful in the effective treatment of patients. Barger et al.⁸ developed a smart house model to detect and verify if the system is capable of outsmarting the behavioral pattern of a patient by using a sensor network that would monitor and track the usual behavioral pattern of the patient. Chiuchisan et al.⁹ developed an architectural model to report inconsistencies in inpatient health and room atmosphere to doctors and relatives so that appropriate steps may be taken to avoid any mishap. Dwivedi et al.¹⁰ built a healthcare prototype to secure patient details that need to be communicated on the internet by using a combination of Public Key Infrastructure, Smartcard and Biometrics technologies. Gupta et al.¹¹ developed a framework that determines and keeps track of details of ECG and various health parameters using Raspberry Pi. The data can be of immense help in medical diagnosis and treatment. Gupta et al.¹² presented a model that utilized the Intel Galileo development unit, which can keep a record of various health parameters and store it in the server. This data can be used to smoothen the patient diagnosis process. The patient does not have to visit the hospital time and again. Hossain and Muhammad¹³ proposed a framework that can be used to benefit patients by using the latest IoT device to capture speech and image signals of a patient. Nagavelli and Rao¹⁴ developed a method which can find the degree of severity of disease risk aversion of a patient based on his medical records by using a statistical approach. Sahoo et al.¹⁵ studied the health care management system. They proposed a method that forecasted future medical parameters of a patient with the usage of massive amounts of health data saved on the big data analytics platform. Tyagi et al.¹⁶ proposed to extensively use IoT to build a network of patients, doctors, family and friends with the consent of the concerned members. It would help reduce the dependence on traditional medical system by interconnecting the patients and doctors, thereby relieving patients of expensive clinical trials and shortage of doctors, thus providing enhanced medical services.Xu et al.¹⁷ proposed to build an emergency IoT based medical data platform using Ubiquitous Data accessing method so that it can be accessed globally anytime, anywhere with ease. Deepika Agrawal et al.¹⁸ proposed to develop an IoT based efficient health care management system which can monitor and store various health parameters like patients heart rate, blood pressure and ECG and send alerts to the doctors for faster and constant medical supervision. Sapna Tyagi et al.¹⁹ proposed a nation-wide IoT based cloud framework where all healthcare delivering entities like doctors, patients, Labs, Pharmacists, Nurses will be connected under one umbrella. Alexandru Archip et al.²⁰ proposed to build a low-cost health care monitoring system to facilitate smooth patient monitoring in hospital wards (post ICU discharge). It is made available on mobile phones for faster and better tackling of medical emergencies. It uses low-power dedicated sensor arrays for EKG, SpO2, temperature and movement. S. Sivagami et al.²¹ proposed a model to monitor the environmental conditions of a hospital for which hospital staff are responsible for using RFID for monitoring the condition. The system would use a set of complementary technologies such as RFID, WSN and smart devices such as mobile. All the technologies would be inter-operating with each other through a Constrained Application Protocol (CoAP)/IPv6 over low-power wireless personal area network (6LoWPAN)/representational state transfer (REST) network infrastructure. Nitha K. P. et al.²² proposed a model to integrate various technologies in health care by using IoT to enable connection among various smart devices ranging from smart wrist wearables to health care systems. Amin and Hossain²³ proposed an intelligent system to ensure the security of health data and prevent its tampering by using innovative technologies such as big data, ambient intelligence, IoT etc. Danilo De Donno et al.²⁴ proposed a Smart Health System (SHS) for automatic monitoring and tracking of patients, personnel and biomedical devices within the hospitals and nursing organizations by using RFID, WSNand smart mobile technologies. Cecilia Occhiuzzi et al.²⁵ proposed an Ambient Intelligence platform Night Care monitor any anomalies in patient behaviour and hospital ambience during night time and report to hospital personnel and families through an alarm. The system is based on Passive RFID technology. The system deploys miniature wearables, inpatient clothes and small devices in the ambience. The RFID reader detects and reports the presence or the absence of the user in the bed, his/her jerky movements and the motion patterns, accidental falls, prolonged absence from the bed and prolonged periods of inactivities such as fainting, unconsciousness or even death by processing the electromagnetic signals generated through the interaction between the patient and surrounding environment. Mohamed Adel Serhani et al.²⁶ proposed a Service Oriented Architecture framework to collect data from patient wearables generated through bio-sensors, store it in the cloud and make it available at the remote monitoring station so that physicians can access it. Yang et al.²⁷ designed a cloud-based framework that used wearable computing to monitor ECG signals of patients. It was quite an efficient healthcare embedded unit. In²⁸, A.Banerjee developed a smart clinical computing module which recorded Used to track and notify all clinical related signal and context data through various sensors. O.Ogunduyile²⁹ used biomedical sensors embedded in smartphones which recorded heart rate and oxygen saturation level of elderly patients admitted in medical units. Zhang³⁰ proposed a real-time information streaming model to monitor and determine biomedical signals with the use of biomedical sensing devices. Varshney in³¹ developed and implemented a pervasive mobile healthcare prototype which was quite helpful in tracking blood pressure, heart rate and alcohol level of newly admitted patients on an emergency basis. S. U. Amin et al.³² proposed a deep learning method to classify EEG data for cognitive health. Alhussein et al.³³ uses deep learning and IoT-cloud platform for smart health monitoring³⁴. Some of the renowned works studied in the literature survey are summarized in Table 1.

Table 1.

Relevant research works undertaken in providing emergency healthcare solutions.

Authors	Technology used	Functionalities
Muhammad et al.6	Deep convolutional network	It is used to quantify ECG signals in home health monitoring
Almotiri et al.7	Wearable devices and body sensor network.	It is used to gauge the overall health status of patients at remote centres.
Barger et al.8	Smart house	It is used to determine ' ‘patient’s general behavioural patterns.
Chiuchisan et al.9	Smart ICUs	Used to monitor any inconsistency in their health status or their body movements
Dwivedi et al.10	Multi-layered clinical data framework combining Public Key Infrastructure and Biometrics model	They are used to secure medical data to be transmitted for examination.
Gupta et al.11	Raspberry Pi prototyping model	It is used to record ECG and other vital health indicators.
Nagavelli and Rao14	Statistical mining approach based on the degree of the disease probability threshold	It is used to forecast the seriousness level of clinical attention needed by tracking medical record of patients.
Sahoo et al.15	Cloud oriented big data Analytics interface	Used to monitor medical indices and predict future health status of patients.
Xu et al.17	Resource oriented Ubiquitous Data monitoring model	It is used to access and record patient health conditions.
Deepika et al.18	IoT based effective medical care management system	It is used to control and access several clinical factors such as blood pressure, heart rate, ECG among others.
Sapna Tyagi et al.19	IoT based cloud prototype	It is used to coordinate all medical entities to be interconnected under a single cohesive unit.
Alexandru et al.20	Cost-effective medical coordination framework	It is used to facilitate trouble-free patient monitoring in medical wards after discharge from ICU.
Sivagami et al.21	RFID and WSN in mobile	It is used to track environmental conditions of medical centers.
Amin and Hossain23	IoT integrated with edge for smart health	It describes the techniques and the widespread deployment of Internet of Things (IoT) solutions in edge systems for EEG classification.
Danilo De Donno et al.24	RFID based Smart Health System (SHS)	It is used to monitor and track personnel, patients and biomedical equipment in hospitals.
Cecilia Occhiuzzi et al.25	RFID technology-based Ambient Intelligence interface	It is used to access any ambiguity in patient behaviour and medical ambience during night hours.
Serhani et al.26	Service-Oriented Architecture model with wearables	Used to aggregate patient-related data and make it available for physicians.
Yang et al.27	Wearable ECG sensors and Cloud for the processing	It is used to monitor ECG signals.
Banerjee et al.28	Mobile healthcare computing approach	Used to track and notify all clinical related signal and context data through various sensors.
Ogunduyile et al.29	Smartphone with biomedical sensors	It is used to record heart rate and oxygen saturation level of elderly patients.
Zhang et al.30	Real-time healthcare data streaming framework	It is used to monitor generic biomedical signals using biomedical sensors.
Varshney et al.31	The pervasive Mobile healthcare model	It is used to track and record pulse rate, blood pressure and alcohol level in newly admitted patients.
Amin et al.32	Deep learning methods for EEG classification	CNN is used to classify the EEG motor imagery data
Alhussein et al.33	(IoT) and Cloud	Deep learning is used in smart healthcare scenario

As observed in the literature review, though there exists models to track health discrepancies in patients using advanced technologies still it has some major flaws. Existing IoT based smart devices mostly focus on one or two health metrics while ignoring the combined assessment of all vital signs. Also, though GPS is inbuilt in smartphones, it is hardly coupled with any medical support system using wearables. Majority systems apply simple alert based models which are not so reliable. Apart from these, maximum wearables are fitness oriented and are not validated clinically to suit their usage in medical centers for patients diagnosis. Traditional prediction models using simple classifiers are static and not so accurate as they lack optimization capability for patient data personalization. Advanced deep learning based models help to detect early abnormalities and also filters out excess false notifications. Moreover, the current systems lack emergency ambulance service support. Recently, the mortality rate based on vital signs has increased due to delay in recognizing critical health issues of patients which result in subsequent delay in ambulance dispatch service. Also, most of the critical responsive models fail to integrate traffic units with the patient location which leads to poor path for ambulance.

Hence, to address these limitations, a personalized and clinically driven real time responsive healthcare system is needed to deal with emergency scenarios. The system should capture real time patient’s vital metrics data which can be transferred to an user interface to monitor, alert and facilitate emergency service seamlessly.Simultaneously, a more precise and optimized risk estimation model can be more reliable for risky cases to alert immediately to patient’s close contacts. Also, it should be able to automatically trigger the emergency ambulance service with integrated traffic coordination for smooth ambulance service to patient location.

Problem statement

Instant response to health emergencies is crucial to reduce medical risks and improve patient care. But at present, the medical emergency responsive systems are struggling with various challenges which affect their reliability. Some major limitations include late detection of high health risks, inadequate risk analysis processes, disorganized resource support and time wastage due to heavy urban traffic. As a consequence, the existing models primarily rely upon manual decision making which results in fluctuations and delays in emergency patient care.

The overall problem is structured as a multi-purpose challenge which involves three constituents:

Real time monitoring of patient and detecting emergencies

Suppose M = {m1, m2, m3,….mn} denote the vital health metrics like pulse rate, blood pressure, body temperature, level of pain, oxygen saturation, body humidity etc. The metrics are captured in real time using wearables to identify fluctuations in vital signs readings and recognize potential health risk emergencies. To achieve this, a predictive intelligence model is needed such that: f(M) → R. Here R represents the level of risk (high, mid or low).

Health risk detection with optimized prediction approach

The defined function f(M) can be fulfilled by applying a suitable feature optimized method with an advanced predictive model to classify the health risk levels. The aim is to enhance prediction accuracy (acc) while minimizing computational latency (l) and false negatives (fn). The problem may be denoted as:

Medical emergency provision with traffic support

If a ‘mid’ or ‘high’ health risk is identified then the system is supposed to alert the medical emergency unit as well as predict the ambulance path (P) with consideration to the traffic information T_i. The system also coordinates with traffic units to set signals along the path P. The objective is to reduce the ambulance arrival time A_a from the patient’s site S_p to the medical center S_m such that:

Thus, the main objective of the research study is to propose and apply an integrated hybrid framework to accurately identify health emergencies utilizing real time vital signs data of patients and further categories the risk severity level through an optimized and computationally enhanced predictive model. This will be facilitated by effective coordination with road traffic units for ensuring quick and priority based ambulance service delivery to patients in need.

Proposed methodology

The proposed model is an attempt to converge predictive intelligence with sensory wearables to take advantage of the healthcare sector. The sole purpose of the device is to provide users with a knowledge of their basic health parameters and provide for faster medical aid for the patients by reducing the delay in ambulance services. This system can be used by members who may belong to any age group, to keep a check on their primary health conditions, with their details to be updated to the databases.

System architecture of model

The model prototype design architecture is illustrated in Fig. 1 and is discussed here. The proposed medical emergency model is an integration of wearable intelligence, cloud computing and urbanized road traffic design. The framework collaborates a lightweight smart wrist band BioTrace-G with the user smartphone application. The band driven by in-built microcontrollers and biosensors, regularly monitors patient’s vital metrics. The pulse rate, body temperature and humidity readings are sensed through the wrist band and is sent to the smartphone through Bluetooth. The mobile application preprocesses the received data and makes it normalized. The genetic algorithm of the GA-DNN module helps in optimization of neural network parameters and the sensor readings thereby selecting only relevant data readings. The DNN algorithm uses this optimized data to estimate the level of health risk. This hybrid predictive model GA-DNN is executed on the user’s mobile to predict risk level. Further, an integration of emergency ambulance service is coordinated with the traffic control unit to reach the patient in quick time. The smartphone activates the decision process and issues emergency alerts if the risk level threshold is crossed. The cloud back-end also works in parallel and authenticates user requests, store events and GPS location to coordinate with emergencies to dispatch the nearest ambulance service. A well dedicated road traffic unit is also associated with the model to facilitate efficient route optimization and priority based traffic signal release to reduce ambulance travel delay. The model used a periodic data sampling frequency of 12 s with an additional event sampling driven by abnormal vitals metrics. The transmitted data among smart band, smartphone app and cloud is secured through Bluetooth Low Energy encryption media. Data uploaded to the cloud using AES (advanced encryption standard) method. Data anonymization is retained by assigning each volunteer a random unique number. No personal information of participants is collected. This anonymized data is stored for training the model. This embedded framework acts as a responsive, intelligent and context sensitive health emergency support.

Fig. 1.

System architecture of the proposed responsive model for emergency healthcare.

The threshold values for the sensors to assess the patient vital metrics in BioTrace-G band is defined in Table 2. As shown, every 30 s period the sensor’s reading is sent. These thresholds act as the decision rules in the GA-DNN model such that any input beyond the normal range can trigger emergency alerts. The model ensures that estimations are aligned with real-world health standards through mapping these raw sensory readings into explainable domains.

Table 2.

Threshold range details of patient’s vital signs captured by biosensors in BioTrace-G band.

Sensor	Data range	Data size	time interval
Heart Rate	60–100 BPM	1 byte	30 s
Body Temperature	97.5 °F and 98.6 °F	1 byte	30 s
GPS Position	0-180 degree	4 bytes	30 s
Relative Humidity	35–60	1 byte	30 s

The risk level for the vital metrics was classified into low, mid and high labels on the basis of the threshold range of the parameters. The patient is at ‘low’ risk if all three metrics are in normal range while it is ‘mid’ if at least one metric does not fit in the normal range. The risk is ‘high’ if more than one metric falls outside the normal range. Based on pseudocode 1, it is observed that for each vital sign, their domain range is different. Moreover, if the recorded risk is in ‘high’ category, then the emergency ambulance service is activated. It signifies that multiple vital signs of the patient are not in normal threshold range and this triggers the alert notification.

Pseudo code 1.

Health risk categorization based on vital signs.

Thus, based on the pseudocode, if the predicted risk is ‘low’ then the patient’s health condition can be referred to as ‘normal’. but if the detected risk is either ‘mid’ or ‘high’ then it falls in the ‘abnormal’ category. Abnormal conditions need attention such that if the risk is ‘high’ then immediate emergency service is triggered accordingly.

The GPS in the patient’s smartphone is used to track the location of the patient. Cloud server acts as the communication layer to store the patients records and their risk level information. Also, the GA-DNN model is trained in the cloud itself. The emergency service libraries in the cloud are responsible to send alert notifications to close contacts of patients and to both nearby medical centers and local traffic control units. If the detected risk is high or mid then the ambulance service is enabled and receives the patient’s location. It dispatches the nearest ambulance to reach out to the patient. The traffic control unit works in collaboration with emergency service as it resets the traffic signals in the ambulance routes for rapid vehicle transport.

Structural modules of model

End-user device module, Staging server module, Predictive intelligence module and End server module forms the structural units of the model as shown in Fig. 2.

Fig. 2.

Graphical view of functional modules of the proposed healthcare model.

• End-User Device Module It serves as the primary interactive module that acts as an interface with the end-user patients. Resource Code and Controller Client form its two components. Resource Code is the set of instructions that collect the sensor data from pulse, temperature, humidity and GPS functionaries. This data is diverted to the Controller Client which processes the readings and pushes them to the controller server unit on the staging server module along with the patient information like name, contact number, customer id among others from the in-built memory local instance of the database.

• Staging Server Module It is the intermediary server that retrieves all sensor readings of patients along with information related to patients and allots with a specific timestamp. Its two sub-components include Controller Server and End Server Client. Controller Server fetches the customer sensor readings from the controller client at a particular timestamp and stores them in a table of the in-memory local instance of the database. Then it pushes the customer vitals at a specific timestamp to the end server client. End Server Client stores the customer vitals at a particular timestamp in a separate table from the controller server in the same in-memory local instance of the database. Then it takes vital information from that table and inputs it to the predictive intelligence module. If any vital metrics deviations is obtained from the IoT intelligence component, then it will send the customer vitals along with their personal information to the End Server-side Server process.

• Predictive Intelligence Module It is responsible for providing the predictive capability to the healthcare model. The intelligence unit is composed of a hybrid combination of genetic algorithm with deep neural network (GA-DNN) to improve the performance of health risk estimation through automated hyperparameters fine-tuning. The GA-DNN model is trained using a vital signs dataset captured through sensor readings from the BioTrace-G band used by 116 volunteers chosen randomly from local housing colonies. The volunteers include elderly people, cardio risks individuals, high stress youths and few patients. By using the aggregated data of vital metrics, GA is helpful to DNN in enhancing predicting accuracy and making the model dynamic.

The GA-DNN procedure for health risk prediction is shown in pseudocode 2. It begins by capturing the vital metrics of patients like pulse rate, body humidity and temperature from the BioTrace-G wearable band. These metrics are the input features to the DNN model for training which is later used for classification of risk into high, mid or low type. But to design an efficient DNN, its hyperparameters need to be precisely chosen. Accurate number of layers, neurons, learning rate, activation function and optimizer are some crucial parameters of DNN that need to be set correctly. Here, GA helps in automating the method as it encodes every feasible layout as a chromosome. By imitating processes of natural evolution like selection, crossover and mutation, it scans the optimum blending of hyperparameters which leads to maximal prediction accuracy. For every generation, chromosome layouts are used to train DNN by applying the accuracy which acts as the fitness score. Over subsequent rounds, the new offspring set develops into a superior optimized model. This hybrid GA-DNN approach makes the tasks of health risk prediction more dynamic, reliable and precise, specifically suited to real time and personalized medical emergency scenarios where quick risk analysis is needed for rapid response support.

Pseudo code 2.

GA-DNN module.

Table 3 highlights the hyperparameters of DNN used to encode a chromosome in GA. These hyperparameters define the design and training of DNN. Here, the values are narrowed intentionally down so as to find a balance among efficiency, accuracy and feasibility of implementing the model in a smartphone. Each chromosome is used to represent a unique DNN layout. For example, a chromosome [3, 64, 1, 0.0004, 32, 0, 0.2] defines the inclusion of 3 hidden layers, 64 neurons per layer with ReLU activation function using 0.0004 learning rate, Adam optimizer with 0.2 as dropout rate.

Table 3.

Values of DNN hyperparameters used for a chromosome in GA.

Hyperparameter	Datatype	Values encoded
Hidden layers count	Numeric	2–3
Number of neurons in each layer	Numeric	32/64/128
Activation function	Categorical	0 = Sigmoid, 1 = ReLU
Learning rate	Floating	0.0001–0.0005
Batch size	Numeric	32/64
Optimizer	Categorical	0 = Adam, 1 = SGD
Dropout rate	Floating	0.2

Hidden layers are kept in the range of 2–3 to make the model adequate in capturing non-linear associations and limit the computational cost. Less neurons in each layer minimizes memory requirement during run time and inference latency. Low learning rate avoids slow convergence and also manages step size during optimization. ReLU is used to handle non-linearity in data patterns while Sigmoid is used for more smooth output probabilities. A small batch size of 32 is suited for lesser memory based smartphones while batch size of 64 helps in better model convergence. A low dropout rate of 0.2 prevents overfitting by deactivating random neurons during the model training process. Small values are appropriate for less complex data like patients vital metrics. This narrow hyperparameters space is helpful for effective tuning of GA based DNN models for achieving optimum accuracy and lightweight deployment. Every chromosome encodes these hyperparameters with an initial population of 30 randomly sampled candidates over 50 generations such that each candidate is determined by training the DNN for 15 epochs. The fitness score is computed on basis of accuracy rate using the equation

Here, ‘FS’ denotes fitness score while ‘Acc’ is the validation accuracy rate and ‘N’ represents the total number of neurons used in the model. 0.01 is the fitness coefficient chosen to balance between model accuracy and model complexity. For example, if the neurons are increased by 100 then the fitness score gets reduced by 1.0 thereby maintaining a good trade-off between accuracy and model size.

Further, the hyperparameters set is refined by iteratively applying tournament selection, single point crossover and mutation. Tournament selection (size 3) picks 3 chromosomes form the population set and the one with the highest fitness score is the parent. Single-point crossover selects a random crossover point in the chromosome and exchanges parts of the two parent ones to form two new offspring. It is followed by enabling every hyperparameter with a 4% probability mutation in each generation. The procedure terminates when no enhancement happens in 5 successive rounds and thus the best configured set with highest fitness score is used for training the model. It signifies faster inference and higher prediction accuracy in smartphone-like resource limited set up.

End Server Module This module deals with the retrieval of customer data and health metrics, thereby storing in the database so that it is accessible to police and medical units. End server process fetches the customer information and health vitals from the end server client and then stores them in an in-memory local database instance to be consumed by the map view system along with the customer and police control units. Further, map view shows the live nearby vital metrics deviation cases on a real-time basis using map zippers through the customer and police control systems.

Smartphone application layouts

The sensor readings of the patient’s vital signs are received by a smartphone application ‘E-response’. The functional screen layouts of the application is illustrated in Fig. 3. It constitutes two modules: Patient module (PM) and Control module (CM). PM handles all user-related issues, interactions and queries. But monitoring along with administration of patients by medical staff and traffic units are done through CM. PM includes six different screens. When a user visits the system, he/she is welcomed by a welcome screen. It displays the home page receiving a user to the healthcare system. After accepting a user, a registration screen appears where a user needs to register himself/herself to take the privilege of the facilities of the system. The user information is required to be verified, and for this purpose, a validation screen is available. Once the user’s details are validated, he/she is provided a unique password to be able to log in the system through a login screen. A meta-data screen contains details of all users successfully registered with the system. At any instant, if any user is inserting or deleting or changing his/her data then for that purpose, an updated screen is also provided to all registered users. The meta-data gets automatically updated whenever any changes occur in the profile of a user which is subsequently reflected in the database. The other component is the CM which deals with coordinating and monitoring aspect of the system model. It has a control login screen which facilitates the medical staffs and police personnel to register themselves and monitor the health status of patients. With the help of location map screen, the administrative team can monitor and trace the exact location of patients nearby and according can take appropriate step to help a needy patient. The server screen constitutes a running remote server which controls the device interaction with the hospital portals. A medical site can retrieve the real-time readings for a specific patient from this server interface.

Fig. 3.

E-response Smartphone application modules.

Patient module (PM)

The PM deals with all interactions concerning patients. It introduces the patients to the application with a welcome message. Registration of first-time users is performed in this unit. Validation of users, login of users, storing customer details and any update concerned with customers are all handled by this unit. Below are the views of individual modules discussed in brief.

Welcome screen Figure 4 illustrates the welcome interface of a customer portal where the general information concerning the healthcare system is summarized, and the user can view it. It helps the customer in getting familiar with the application, especially those who visit the system for the first time.

Fig. 4.

Welcome screen in E-response smartphone application.

Registration screen Figure 5 presents a view of the registration interface where the customer needs to fill in the details like Customer ID, Customer Name, Email ID, address, among other information to register with the healthcare model. It is essential for the customers who visit for the first time and are willing to use the model. Without a valid registration, a customer cannot take advantage of the utilities available with this prototype.

Fig. 5.

Patient registration screen in E-response smartphone application.

Validation screen Validation interface is a vital part of PM. It is concerned with the customers who register themselves with correct information in this healthcare model. Once a user expresses his/her data, it is verified by the system. If it is found to be accurate, then a green message dialogue box will pop up notifying the same. Along with it, an email will be sent to the customer for validation of correct email as seen in Fig. 6.

Fig. 6.

Patient validation screen in E-response smartphone application.

Login screen After successful validation, the patient can log in with the system at a convenient time using the allotted ID. The customer needs to login to the portal using customer ID and password. The log in the module is highlighted in Fig. 7.

Fig. 7.

Patient login screen in E-response smartphone application.

Meta-data screen Meta-data Interface as shown in Fig. 8 is the data repository of the proposed healthcare system. It stores the details and information of all customers who have successfully registered with the system. Any registered customer can view his details at his convenience. Figure 8 shows the Meta-data interface.

Fig. 8.

User meta-data screen in E-response smartphone application.

Update interface An updating link is also provided to customers for any kind of insertion, deletion or modifications required at any instant as depicted in Fig. 9. It facilitates a customer to change any of his details except the unique Customer ID.

Fig. 9.

Patient update screen in E-response smartphone application.

Control module (CM)

Besides the patient view in PM, there is also an administrative unit available which is called control module (CM). It is responsible for effective monitoring and control of traffic personnel units as well as nearby medical centers for assisting emergency patients. Here the traffic staff and the medical emergency ambulance personnel are required to log in, and the coordinates of nearby hospitals and police control rooms will be traced. Location of all patients will also be available to administer at any instant. The modules of CM are discussed below.

Admin-Login screen The log in page is available for the clinical staff and policemen to register themselves as shown in Fig. 10.The hospital and police control room need to login to their portal using a unique response ID that will be generated by the system.

Fig. 10.

User control login screen in E-response smartphone application.

Location screen A map view is presented in Fig. 11. The google map was generated using JavaScript 3.58 https://developers.google.com/maps/documentation/javascript/versions. The map view automatically shows the nearby locality of the respective hospital and traffic control room upon their login to quickly lookout for emergencies. The green zipper offers the non-responsive users, and the blue zipper shows the already attended/responded users. After the hospitals answer the users, the blue zipper stays for a while and then disappears, clearing the map view area. The hospital can see the map view after logging into the web portal through its credentials. It shows the location of all the patients (addressed or unaddressed) near the locality. On clicking on the zipper of a patient which shows the patient details such as name, ID and phone number, it will redirect to another screen from where the hospital can see the live real-time readings of the concerned patient.

Fig. 11.

A view of Location Map screen showing the patient coordinates location.

Server screen A server running in the cloud acts as a controller service for the interaction of devices endpoints to the hospital portal. The hospital fetches the live stats for a customer from this server. A real-time sample recording of a patient is highlighted in Table 4. It includes live GPS Coordinates based on Time interval and the corresponding pulse readings with body temperature, and humidity level are also available. Figure 12 highlights the live readings of a specific patient displayed on the server console derived from the device endpoints. It constitutes not only the pulse reading, body humidity and temperature of patient but also the GPS coordinates indicating the current location of patient.

Table 4.

A sample readings of a patient at 40 s intervals received at the server.

Pulse rate (in BPM)	Body temp (in °F)	Body humidity (in %)	Live GPS coordinates	Risk level (Low/Mid/High)
87	97.4	48.29	20.324504,85.813646	Low
96	97.5	48.31	20.324504,85.813647	Low
94	97.4	48.32	20.324504,85.813648	Low
107	97.3	48.29	20.324504,85.813649	Mid
97	97.4	48.29	20.324504,85.813650	Low
98	99.2	62.42	20.324504,85.813651	High
95	97.4	48.29	20.324504,85.813652	Low
104	99.6	61.23	20.324504,85.813653	High
95	97.7	48.29	20.324504,85.813654	Low
96	97.6	45.44	20.324504,85.813655	Low
108	98.8	63.75	20.324504,85.813656	High
97	97.2	61.92	20.324504,85.813657	Mid
109	98.9	48.44	20.324504,85.813658	High
96	97.7	48.44	20.324504,85.813659	Low
98	97.4	54.31	20.324504,85.813660	Low
104	97.7	50.34	20.324504,85.813661	Mid
95	97.1	55.64	20.324504,85.813662	Low

Fig. 12.

A view of Server screen receiving ‘patient’s live data from the smartphone.

Result analysis

Several relevant works as discussed in the literature survey are concerned with collecting, aggregating and storing healthcare-related information and data using suitable mechanisms. Few research works also dealt with the processing of the aggregated data but in a non-real time scenario. None of the existing smart healthcare models incorporated a real-time response mechanism. Unlike previous models, our system can monitor the health parameters of its user and thereby help in providing timely medical services to him in case of any vital metrics deviations. Our proposed smart, responsive and intelligent healthcare model captures health-specific input data of a patient from sensors and determines the critical health parameters based on the input received. Health risk level is detected using predictive intelligence service of the model. Accordingly, in case of any critical risk detected, a signal is transmitted from the patient location to both the traffic police control site as well as to the medical care on the occurrence of health issues. Thus it is fruitful in integrating police control and medical centers, thereby coordinating the entire process. Hence preventive actions can be taken on an emergency basis on a smartphone application interface for faster attainment of patients and thereby reducing the time of medical services to reach the patient. Through this system, the patient is no longer required to get to the medical services. Instead, the medical services will automatically advance to the patient in case of emergency by the use of the system’s smart and robust mechanism and architecture. A comparative analysis of the technology used in the proposed healthcare model is done with some popular works discussed in the literature survey and is summarized in Table 5.

Table 5.

Comparison of existing models with proposed framework using multiple metrics.

Authors	Developed Framework	Technology implemented
		WSN	BAN	Client- Server prototype	Exclusive-Pair IoT Model	Scalable Back-end Architecture	GPS Module
Almotiri et al.7	Wearable computing	NO	YES	NO	YES	NO	NO
Chiuchisan et al.9	Smart ICUs	NO	YES	NO	NO	NO	NO
Nagavelli and Rao14	Statistical mining approach	NO	NO	NO	NO	NO	NO
Sahoo et al.15	Big data Analytics	YES	YES	NO	NO	NO	NO
Xu et al.17	Ubiquitous Data monitoring model	YES	YES	NO	YES	NO	NO
Sapna Tyagi et al.19	IoT based cloud prototype	YES	NO	YES	NO	NO	NO
Alexandru et al.20	Medical coordination framework	YES	YES	NO	NO	NO	NO
Sivagami et al.21	RFID- WSN	YES	NO	NO	NO	NO	NO
Amin and Hossain23	Big data Edge Intelligence and IoT	YES	NO	YES	YES	YES	NO
Danilo et al.24	RFID Smart Health System	YES	YES	NO	NO	NO	NO
Cecilia Occhiuzzi et al.25	RFID based Ambient Intelligence	YES	YES	NO	NO	NO	NO
Mohamed Adel Serhani et al.26	Service-Oriented architecture	NO	YES	YES	NO	NO	NO
Yang et al.27	Wearable ECG sensors	NO	YES	YES	YES	YES	NO
Banerjee et al.28	Mobile healthcare computing	YES	YES	YES	NO	NO	NO
Ogunduyile et al.29	Mobile biomedical sensors	YES	YES	YES	NO	NO	NO
Zhang et al.30	Real-time healthcare data streaming framework	YES	YES	YES	YES	NO	NO
Varshney et al.31	Pervasive Mobile healthcare	YES	YES	YES	NO	NO	NO
Proposed Responsive model	Health emeregncy	YES	YES	YES	YES	YES	YES

Privacy of information and consent of patients are other important factors that are taken care of in the model. A written consent form is provided to all participants which explains the objective of data collection (training the GA-DNN model with emergency detection), type of information gathered (vital signs only) and data privacy issues (anonymize data with encryption). These signed consent data were aggregated before collecting data. The participants took part in the process voluntarily and they were notified that the model’s purpose was to monitor and alert the volunteers. The framework uses SSL protocol for data encryption and sharing among BioTrace-G band, mobile device, cloud and other interfaces like traffic units with medical centers. Data stored on smartphone and cloud was encrypted with AES (advanced encryption standard) protocol. Also, attributes of patients like name and contact data are masked before sending it to the GA-DNN model training. Patient consent is also taken utmost care during usage of the mobile application E-response. A patient needs to allow the model to collect and share data along with prediction. Consent regarding capturing data like patient vital signs and location are acquired from the patient before using it. Access entities and timing of alert notification are also scheduled with the patient’s knowledge. Every participant was provided unique consent ID and any personal information of participants like name, phone number, address etc. were not retained. This ensured anonymization of participants.

The proposed IoT based healthcare responsive model was further tested on a random patient admitted to a nearby local clinical centre. Some health parameters like resting heart rate, walking heart rate, anticipation heart rate, temperature fluctuation and relative humidity level were measured and analyzed using the discussed model. The result was very positive and impressive. It confirmed the normal functioning of the body over time. In Fig. 13, resting heartbeat (not while sleeping) (in bpm) vs. time interval graph has been illustrated which shows a range from 67 bpm to 116 bpm which is typical for adults because of the normal functioning of the heart nodes. The graph shows the changes of the patient’s resting heart rate over a fixed 40 s time. Natural fluctuations in heart rate is seen around baseline with slight deviations because of anxiety and change in posture. The representation helps the modle to differentiate between normal resting changes and abnormal patterns. By subjecting this data into GA-DNN model, it captures short-term anomalies as well as ling-term anomalies which ensure accurate predictions in emergency situations.

Fig. 13.

Resting Heart Rate analysis of patient with the proposed model.

In Fig. 14, the anticipation heartbeat (in bpm) vs. time interval graph shows the range of 76 bpm to 130 bpm, which appears to be more when compared to the resting heartbeat. This happens because the electrical impulses that flow through the atria, which causes its contraction, are more due to the phenomenon of anticipatory rise. The graph denotes variation in heart rate in anticipatory state like in stress and sudden awareness. This creates a temporary rise in heart rate above a person’s baseline. By processing this variation, the model can distinguish between short term, stress driven response and sustainable abnormality which makes sure that alerts are notified only when medically relevant deviation happens.

Fig. 14.

Anticipation heart rate analysis of patient with the proposed model.

The walking heartbeat determines the cardio response of patient during mild physical activities thus giving vital results for general heart state in context to fitness level and walking speed. Here, the smart band regularly track patient’s vital metrics when the patient is in motion. This information is sent to smartphone for processing in real time. In Fig. 15, the walking heartbeat (in bpm) vs. time interval graph ranges from 87 bpm to 147 bpm due to the massive flow of electrical impulses causing repetitive contractions of the atria. The sinus node signals the heart rate to speed up in case of exercising, walking or in a situation where adrenaline flow is involved. This further adds to the increased bpm in walking heart rate.

Fig. 15.

Walking heart rate analysis of patient with the proposed model.

In another analysis, the temperature-time study is done to assess the variations in body temperature of patient over time to monitor any anxiety level or fever. This is constant during resting time while slightly increases during any activity or exposure to environment. In case of sustainable ris ein temperature over time, the model indicates the possibility of any infection, stress or dehydration. Figure 16. shows the rise in the body temperature of an individual in motion concerning the time interval. The graph shows the minor rise in temperature over a particular interval because of the time taken by the body to warm up due to the continuous sweating due to physical activities or the presence of mild fever in the body.

Fig. 16.

Temperature-Time analysis of patient with the proposed model.

Humidity time analysis is also important a sit keeps track of the body humidity level that surrounds the patient. The captured data from smart band is sent to E-response application in smartphone for further analysis. Minor variations correspond to emtional stress or light activity form patient. Abnormal rise in body humidity may signal presence of fever or anxiety while a sudden fall in level indicates dehydration. Through the humidity analysis, the model reduces false positive rate which increases pesonalized health assessment of patient. Figure 17 highlights the rise in the relative humidity of the body with respect to the time interval. The graph shows the rise in the humidity over a specific interval of time due to the excessive sweating or increased rate of the circulation of the blood and increase in the rate of the respiration.

Fig. 17.

Humidity-Time analysis of patient with the proposed model.

The intelligence capability of the model driven by the optimized GA-DNN is responsible for superior performance to a wide extent. To validate the model, it is compared with other popular variants like DNN, PSO-DNN and GS-DNN techniques. An analysis was carried out for risk level prediction accuracy using the GA-DNN model with respect to sensor readings obtained from BioTrace-G band as shown in Fig. 18. It was observed that GA-DNN generated a consistent performance throughout the procedure. 93.5% is the least accuracy rate with 100 readings, while 96.8% is the optimum accuracy generated with 1700 sensor data. The mean risk prediction accuracy noted was 95.2%. As seen, the accuracy drops and is inconsistent when applied to other models. The higher accuracy of the GA-DNN model with increase in data is because it provides superior feature selection and noise minimization. Also, it allows the GA to capture a rich search space for better model layout and choose optimum DNN hyperparameters. It makes the model more fine-tuned.

Fig. 18.

Health risk prediction accuracy of model with respect to sensor reading data.

A scalability analysis was performed using the proposed GA-DNN prediction model. A demonstration of prediction accuracy rate with respect to the number of patients was undertaken, and its result is illustrated in Fig. 19. As the number of patients increased, risk prediction accuracy decreased in case of the variants. Though DNN and PSO-DNN models showed promising results still GA-DNN outperformed them marginally. Maximum accuracy of 96.4% was recorded with 100 patients while the mean accuracy was 94.2%. Thus, GA-DNN was found to offer very stable and scalable performance. With the rise in the number of patients, the prediction rate of accuracy remains relatively high and almost consistent throughout. It is because of the global optimization ability of GA-DNN and its generalization capability that prevent it from overfitting. Use of lightweight protocols and ability to buffer is helpful in load management. Pruning enables the model to execute effectively on mobile phones thereby minimizing load. Also, cloud driven push protocols are very robust. All these features make the model scalable.

Fig. 19.

Health risk prediction accuracy of model with respect to patient count.

Inference latency plays an important role in medical emergency scenarios. It refers to the prediction delay of a model on unseen test input data. When this analysis was done on the different DNN models, it was observed that the latency was least with GA-DNN approach in risk prediction. As observed in Fig. 20, GA-DNN approach outperformed other models in context to the rise of the sensor’s data readings. When the sensory readings reached 500, then prediction with GA-DNN took only 112.5 s as compared to other models. The mean delay recorded was only 57.8 s and hence risk prediction with GA-DNN approach proved to be faster than the brute-force approach. GA can evolve with shallow connected DNN with fewer neurons, effective size of layers and optimum activation functions. This makes the network simple, structured and extremely optimized thereby reducing the inference latency time.

Fig. 20.

Inference latency analysis of model with respect to sensory readings.

Further, to validate the model performance, false positives (FP) and false negatives (FN) are taken into consideration. Both these metrics are very crucial for patient safety in health risk detection of patients mainly in emergency scenarios. If the false negatives are high then the predictive model fails to detect a critical health risk and it could lead to potential serious implications to the patient’s life. In false positives, the model wrongly estimates a risk when in reality there is no risk. This may lead to unnecessary medical resource wastage. The GA-DNN model balances both these metrics very well such that critical cases are not missed out while reducing false alerts simultaneously. The GA component finds the most significant sensor readings, fine-tunes the hyperparameters of the DNN model and minimizes the noise by selecting relevant features which may lead to more correct predictions. The global search space of DNN is parsed and as a result the model achieves more generalization. While other models like PSO-DNN and GS-DNN can get trapped in local minima but GA applies mutation with crossover to avoid local minima thereby discovering coupling of features to minimize error rate. An analysis for both these metrics are done for a 4 weeks duration using different models. As seen in Fig. 21, the GA-DNN generates the least FN in all weeks while recording a mean value of 6.9%. Other models like PSO-DNN and GS-DNN records mean FP of 10.7% and 9.3% respectively.

Fig. 21.

False negatives rate analysis of model over 4 consecutive weeks period.

Similarly as shown in Fig. 22, the FP values with GA-DNN is also less as compared to that of other models considered in the study. A mean of 13.4% FP with GA-DNN is observed which means it successfully detects the false alarms thus prevents resource wastage like ambulance dispatching service. Mean FP of 15.4% and 16.6% were recorded for PSO-DNN and GS-DNN respectively which shows the effectiveness of GA-DNN approach.

Fig. 22.

False positives rate analysis of model over 4 consecutive weeks period.

Evaluating the model through real-world pilot deployments is essential to ensure its generalization, reliability, and acceptance in practical health scenarios. The BioTrace-G device with embedded smartphone intelligence model was tested on a sample of 100 emergency cases related to vital signs fluctuations that are communicated to the local emergency ambulance service over a distance of 10 km radius from the patient’s location. The 100 patients include elderly citizens with heart risks and stress intensive middle aged youths. The participants were provided with BioTrace-G band and smartphone equipped with E-response application. The consent form from participants were also collected prior regarding sharing of data. The duration of the set up was four weeks in continuous usage. The back end services include hosting of cloud driven GA-DNN model with emergency local ambulance dispatch and integrated traffic module support for route optimization. Pilot data is used to fine-tune the GA-DNN model. Vital metrics were captured in these emergency cases through BioTrace-G devices and sent to smartphones. The GA-DNN model analyzed these readings to estimate the risk level of the patient (low, mid or high). The risk level was instantly communicated to nearby medical centers. High risk patients were detected and ambulance service was given priority to these cases. It arrived in quick time avoiding delays. Thus, the response time in dispatching ambulance service was significantly reduced as compared to the traditional approach.

As observed, the proposed model was compared with the traditional approach to test its effectiveness. The false negative rate with the proposed model was significantly reduced by 4% while the emergency patients prioritization was also enhanced by 8%. Also, the average ambulance dispatch time got reduced by 4.1% which justifies the relevance of the model. Thus, the deployment of the model for the risk prediction system resulted in rapid and intelligent ambulance service which led to overall reduction in response time, specifically for high-risk patients. The outcome is shown in Fig. 23.

Fig. 23.

Analysis of proposed responsive model with conventional process in context to emergency ambulance service.

Based on patient sensor data, the high risk cases are detected through the GA-DNN model. The notification was instantly sent to the local traffic control units. The model used RESTful APIs to transfer routing with GPS information of patients and ambulance to the traffic control through the smartphone application. A cohesive integration of traffic context routing with the proposed model makes sure that when a critical medical scenario is identified by the GA-DNN module, the ambulance dispatch service is provided along with an optimum route sleection to minimize delay. It is feasible through a smart traffic signal cooridnation and a GPS navigated system. Here, the dispatch service instantly sends the GPS position of ambulance along with route details to the traffic control server. Sensors driven traffic lights with network connectivity accepts this information and adapt their signal patterns to generate a ‘green passage’ for the ambulance. It decreases the number of red signal stopping points and thus reduces congestion based delays. Also, the navigation model utilizes the GPS coordinates with the live traffic contexts like road population, reported accidents, construction sites, climatic conditions etc. to determine the optimum route. Accodingly, it reroutes the passage if traffic hurdles arises. The simulation set up is developed and tested with the same 100 emergency cases. A comparative study is done on the impact of the API based integrated traffic system. A moderate population density with 7 intermediate junctions and mild congestion was picked for the testing. Thus, by combining traffic context control and dynamic GPS mechanism, the model enhances important functional parameters like increased mean ambulance speed, lesser signal stops and less response latency as compared with the model without any traffic integration.

As seen from Table 6, the mean response delay which is the duration between ambulance dispatch from medical center to the time of arrival for patients is reduced by 3.9 min. The average ambulance speed recorded from point of dispatch to the patient location is also minimized by 7.8 km/hr. Number of signals where an ambulance had to wait due to a red signal is also limited to only 2. Besides, the mean signal delay is also optimized to only 1.9 min. Hence, this coupling ensures that emergency medical service can reach to patients in least possible time thereby enhancing survival rates in critical scenarios like cardiac risks or respiratory issues.

Table 6.

Impact analysis of integrated traffic support in proposed responsive model.

Parameter	Without traffic integration	With traffic integration
Mean response delay(min)	13.3	7.4
Mean ambulance speed(km/hr)	26.8	34.6
Signal stops count	4	2
Mean signal latency(min)	3.4	1.9

Also, the emergency responsive GA-DNN model is assessed with standard benchmark datasets using common evaluation metrics to validate the robustness of model in assessing health risk prediction performance. WESAD (Wearable Stress and Affect Detection), MIMIC (Medical Information Mart for Intensive Care), PAMAP2 (Physical Activity Monitoring) and DEAP ( Dataset for Emotion Analysis using Physiological Signals) are the four datasets used for the purpose. These datasets were applied along with the real data collected from BioTrace-G band and are preprocessed so as to adjust with the sensory abilities of the proposed model. The steps include resampling, normalization and labeling. The physiological metrics from the datasets corresponding to the vital signs used in the model are extracted. Missing values are handled with forward-fill imputation. The retrieved data is resampled to adjust the sampling frequency of the model. Bad-pass filtering is further used for noise removal and is followed by a min-max method for data normalization. Finally, data is labeled on the basis of high/low threshold values set for emergency scenarios. After preprocessing, the datasets are input in the GA-DNN module. Accuracy, precision, recall and f-score metrics are further used to evaluate and compare the performance of these benchmark datasets with the proposed model.

As highlighted in Table 7, the GA-DNN model performed better with the real time data captured from patients as compared to other standard datasets in terms of accuracy, precision, recall and f-score metrics. But overall, the model remains consistent throughout. It is to be noted that models trained using WESAD or DEAP datasets recorded good accuracy for detecting emotions like stress and anxiety but they fail to generalize in adaptive and critical health scenarios.

Table 7.

Comparison of proposed model with standard benchmark datasets.

Datasets	Accuracy (%)	Precision (%)	Recall (%)	F-score (%)
WESAD	92.5	91.8	91.1	91.5
MIMIC	90.7	89.9	88.6	89.3
PAMAP2	90.3	89.7	88.7	89.3
DEAP	91.9	91.4	90.5	90.8
Proposed	94.6	93.8	93.1	93.4

Thus, as observed, the proposed model offers superior performance in emergency risk detection of patients. Still, it suffers from a few drawbacks. The potential limitations of the model are as follows:

• Restricted data diversity: The training dataset is generated from limited pilot implementations. The benchmark data samples fail to represent complete population demography, medical risks and environmental elements corresponding to diverse patient populations.

• Simple vital metrics as model input: The model depends upon a simple vital metrics set (pulse rate, body temperature and body humidity). It is unable to include other crucial parameters like respiratory rate, clinical history, ECG data etc. which can impact the accuracy rate.

• Hardware dependency: The performance of predictions is dependent on smart band quality and hardware configuration of smartphones which vary among individuals.

The ethical considerations of the model can be outlined as follows

• Participants provided an explicitly written consent form prior to collecting data.

• Anonymity of participants collected data was restored through unique features thereby retaining their privacy while maintaining model’s training usability.

• Encryption of data using AES (advanced encryption standard) and TLS (transport layer security) are enforced to maintain communication privacy over the network.

• The model is primarily designed only for monitoring and alerting tasks but it can not replace professional healthcare staff and their diagnosis approach.

Conclusion

The study designed a framework which integrated a wearable smart band ‘BioTrace-G’ to collect vital signs data of patients. The data is transferred to a smartphone application ‘E-response’ to manage patient profiles and it is enabled by integrated traffic support and an intelligent predictive capability of GA-DNN model that classifies and predicts the critical patients who need immediate ambulance care. The model makes sure of timely notification to ambulance service. GA helps in hyperparameters tuning thereby improving the generalization across patients. Unlike all previous other health care devices, this framework can monitor its user’s health vital parameters and thereby help in providing timely medical services to him in case of any critical deviations in vital metrics. It can measure specific health parameters with inputs from sensors and detect health discrepancies with the help of predictive intelligence service and sending output GPS signals of the exact location of the device bearer to medical care on the occurrence of health issues. It also integrates and coordinates traffic police and preventive actions on a smartphone enabled application interface for faster attainment of the emergency and reduces the time of medical services to reach the patient. The model upon implementation showed promising results in context to several evaluation metrics. By using this device, we were able to reduce the ambulance response times to a much smaller extent by testing it on a sample of a few emergency cases, as discussed in the results section. By reducing the traffic jam by directly integrating the healthcare services with the police control room and timely detecting a critical health risk, we reduced the time to almost 50% of the actual time taken otherwise. Upon implementation, the model gave promising results. The mean risk prediction accuracy noted was 95.2% in context to sensor readings while the mean accuracy was 94.2% with respect to the number of patients. The mean inference delay recorded with the GA-DNN model was only 57.8 s. 13.4% FP with GA-DNN is observed and it generated the least FN of 6.9%. Also, the deployment of the model resulted in rapid and intelligent ambulance service which led to overall reduction in response time for high-risk patients. It also validated the importance of integrated traffic support in superior emergency service to critical patients. Further, when tested with other benchmark datasets, the model generated consistent performance. By the use of this device, the patient is no longer required to reach the medical services. Instead, the medical services will automatically get to the patient in case of emergency by the use of the system’s smart and robust mechanism and architecture.

In future, the proposed model can be integrated for real-time deployment in hospitals for direct connectivity with medical records and its informative systems thereby facilitating automated patient assessment and decision support. The model will further be upgraded for deploying in rural infrastructure by utilizing power constrained communication technologies and edge intelligence to operate in regions with restricted internet-connectivity and facilities. Also, the study intends to improve model transferability thus permitting for generalization across varying patient population and medical state to manage personalization of clinical risk estimations.

Author contributions

Sushruta Mishra and Hrudaya Kumar Tripathy: Concept, methodology, and simulation setup. Himansu Das: Performed model implementation, and prepared the original article. Mohammad Shahbaz: Review and Drafting and Supervision . Surbhi Bhatia Khan: Review and Drafting and Validation . Ahlam Almusharraf: Formal analysis and Review and drafting.

Data availability

The data that support the findings of this study are available on request from the corresponding author.

Declarations

Competing interests

The authors declare no competing interests.