A Smart Healthcare Recommendation System for Multidisciplinary Diabetes Patients with Data Fusion Based on Deep Ensemble Learning

The prediction of human diseases precisely is still an uphill battle task for better and timely treatment. A multidisciplinary diabetic disease is a life-threatening disease all over the world. It attacks different vital parts of the human body, like Neuropathy, Retinopathy, Nephropathy, and ultimately Heart. A smart healthcare recommendation system predicts and recommends the diabetic disease accurately using optimal machine learning models with the data fusion technique on healthcare datasets. Various machine learning models and methods have been proposed in the recent past to predict diabetes disease. Still, these systems cannot handle the massive number of multifeatures datasets on diabetes disease properly. A smart healthcare recommendation system is proposed for diabetes disease based on deep machine learning and data fusion perspectives. Using data fusion, we can eliminate the irrelevant burden of system computational capabilities and increase the proposed system's performance to predict and recommend this life-threatening disease more accurately. Finally, the ensemble machine learning model is trained for diabetes prediction. This intelligent recommendation system is evaluated based on a well-known diabetes dataset, and its performance is compared with the most recent developments from the literature. The proposed system achieved 99.6% accuracy, which is higher compared to the existing deep machine learning methods. Therefore, our proposed system is better for multidisciplinary diabetes disease prediction and recommendation. Our proposed system's improved disease diagnosis performance advocates for its employment in the automated diagnostic and recommendation systems for diabetic patients.


Introduction
A recent development in biotechnologies and high throughout computing progressively contribute to quick and affordable e-healthcare data collections and disease diagnosis. e efficiency and reliability are dependent on accurate model building from e-healthcare big data. One of many lifethreatening diseases is diabetes disease (DD) [1][2][3].
Diabetes disease arrests 422 million adults all over the world [4]. e death rate due to diabetes disease is 1.5 million, and 3.7 million deaths are due to diabetes and high blood pressure [4]. Diabetes disease is a multidisciplinary disease that arrests the human body's significant parts like the kidney, eyes, lungs, and heart. e diagnosis of this disease was done either manually by a medical practitioner or by any automatic device.
All of these types of measurements for diabetes disease have some benefits and some drawbacks. Any experienced medical expert cannot manually find the diabetes disease early due to some hidden side effects on the human body. With the intelligent recommendation system's help and application of deep machine learning (DML) and artificial intelligent methods, this disease can be predicted at the earlier stage [5][6][7][8][9] with a minimal error rate.
ere are some healthcare automated systems for detection and recommendation of human diseases in recent researches. Myocardial infarction [10] is an acute disease for blood circulation in the heart. In this paper, deep CNN is applied for detection and to prevent humans from a heart attack. A computer-aided diagnosis (CAD) system [11] is used by applying the transfer learning technique for accurate and timely response to reduce the extensive calculation. In this paper, a CAD system works efficiently and accurately to detect and prevent heart attacks. Internet of Healthcare ings (IoHT) and Decentralization Interoperable Trust (DIT) [12] framework are a better healthcare system. In this paper, blockchain is used for data privacy and security. In this research, data is collected via IoHT at each point and transformed through blockchain for smooth and accurate healthcare data for better system accuracy.
Many ensemble learning models have been used in recent healthcare researches for better accuracy. For example, hepatocellular carcinoma [13] is a hazardous cancer disease in the human body. In this paper, an automated prediction system is developed using a stack learning approach for deep learning and examining healthcare data about this deadly disease. Stack learning is an ensemble learning technique. In this paper, evolutionary computational techniques are also used to examine the healthcare data about hepatocellular carcinoma disease. In cervical cancer [14] diagnosis, an ensemble machine learning approach is used. In this paper, two approaches are used for predicting disease on an images dataset.
In mobile edge computing [15], an automated recommendation system is proposed for the joint computation of multiuser offloading and task caching. In this paper, Q-learning and Deep-Q-Network-based algorithms are proposed for this system. Multilevel vehicular edge cloud computing [16], secure federated learning for 5G [17], and augmented Coronavirus disease detection [18] used an advanced ensemble deep learning approach for better results.
For detection of COVID-19 disease, a deep learning approach is adopted with an augmented approach [18] and it achieved 100% accuracy. In industrial mobile edge computing [19], the deep ensemble learning approach is used for resource allocation and data security. e importance of a smart healthcare recommendation system is increasing day by day for better and timely prediction. To minimize the risk of life-threatening human diseases, we need an efficient system for diagnosing and effectively recommending life-threatening diseases such as diabetes. Electronic health records (EHRs) play an essential role in smart healthcare recommendation systems for predicting life-threatening diseases, especially for multidisciplinary and life-threatening diabetes diseases. However, the data collected from sensors and EHRs are unstructured. To manage adequately such kinds of multisourced data for further examining is a challenging task.
Further, extracting the critical features and fusing them in a structured form is also a hectic and skills-demanding task. erefore, this section is further divided into two parts: a wearable sensors-based diabetes prediction system and extracting information from EHRs textual data. en, data fusion is essential for better results and accurate prediction of diabetes disease with DML.
Many recommendation systems for healthcare are already proposed in recent researches. e significant contribution of this research is to enrich the healthcare dataset for the best prediction of multidisciplinary diabetes disease. We have collected the patients' data through wearable sensors and EHRs in the textual record form of each patient. After collecting the records of each patient, essential data from both ends are fused to enrich the healthcare dataset. e Ensemble deep learning approach works accurately and produces better results in larger healthcare datasets. Finally, we have developed a better recommendation system by collecting patients' records and applying an ensemble machine learning approach for accurate and timely prediction and recommendation of multidisciplinary diabetes disease patients. e organization of the paper is as follows: Section 2 describes the most recent developments of diabetes disease detection and recommendation from the literature; Section 3 provides research methodology, data fusion, and proposed DML model; Section 4 presents the dataset selection, preprocessing, data fusion, and results and discussion; Section 5 describes the conclusion and future work; and the last section is devoted to references.

Related Work
Many researchers contributed to diagnosing diabetes disease. ey have used machine learning (ML) classifier and artificial intelligence (AI) assistance for the prediction of diabetes disease. With the help of artificial intelligence, we can easily collect healthcare data. After collecting the big data from the healthcare center, we can easily predict human diseases, including multidisciplinary diabetes diseases.
In early detection of diabetes disease [20], the k-nearest neighbor (KNN) classifier model was used and the result was compared with that of the support vector machine (SVM) model achieving 85.6% accuracy. In this paper, the KNN machine learning classifier was used for predicting diabetes disease. e results comparison was made with another machine learning classifier called SVM to authenticate the work. For the detection of diabetes type II disease [21], the authors used a convolution neural network (CNN) and compared their work with the linear regression (LR) model and multilayer perceptron (MLP). In this paper, the neural network was used to diagnose diabetes type II disease. For results comparison, two machine learning classifiers were used for the authenticity of the work. An accuracy of 77.5% was achieved in the area under the curve (AUC). Analysis of early detection of diabetes disease with feature selection technique [22] was carried out using SVM classifier and their results were compared with random forest (RF), naïve Bayes (NB), decision tree (DT), and KNN classification models. e highest accuracy achieved with SVM was 77.73%. In this paper, the feature selection technique was adopted. With the help of the feature selection technique, we can reduce the system's computational capability and improve accuracy. Multiple machine learning models were applied for comparison and authenticity of results. Bloodless techniques for the prediction of diabetes disease are used with computational tools [23]. e accuracy achieved through this technique was 91.67%. In diabetic retinopathy detection [24], the deep (DNN) technique was adopted. e accuracy achieved via CNN was 74.4%. Detection of multiclass retinal disease [25] was done with with AI. e CNN classifier was used and it achieved 92% accuracy.
A data-driven approach is used for predicting diabetes and cardiovascular diseases [26] with ML.
is paper adopted an extreme gradient boost and compared it with the LR, SVM, RF, and weighted ensemble model. An accuracy of 95.7% was achieved in the area under the ROC curve. In type II diabetes disease prediction [27], an ensemble classification model was adopted. e accuracy achieved via this model was 82.2% in the AUC. A new methodology, smartphonebased diabetes detection [28], was presented. In this paper, image data was considered for diagnoses and further directions. A microcontroller-based agent [29] was used to measure the blood glucose level of patients. A sensor integrated therapy [30] for diabetes disease was used to monitor glucose levels in a diabetic patient. A self-recommendation smart app [31,32] was used and trained on recorded health data like patients' daily physical activities and other important parameters related to diabetes. e valuable information extraction from wireless sensor data and the patient's electronic medical record is also challenging for predicting multidisciplinary disease. To handle this challenge, different models have been presented for extracting the most valuable information from the healthcare textual data [33][34][35] for making a dataset for the prediction of diabetic disease. e textual dataset collected through EHR's was preprocessed and converted into a meaningful format as per smart healthcare recommendation system demands. e wireless sensor data of healthcare was also collected through wireless devices. After collecting data through a wireless communication device, data was preprocessed for removing noisy wireless data to make a rich and accurate healthcare dataset. In this way, we can easily apply machine learning algorithms for predicting multidisciplinary diabetes disease.
After collecting and converting meaningful data from textual data and fusing preprocessed wireless sensor data [36][37][38] for making a rich healthcare dataset of diabetes, Table 1 shows the comprehensive limitation of previously published approaches.

Smart Healthcare Recommendation System for Multidisciplinary Diabetes Patients
is section explains the overall structure of a smart healthcare recommendation system for multidisciplinary diabetes disease patients (SHRS-M3DP) in detail. e proposed approach is divided into distinct layers for an accurate description of each layer working. In the end, the ensemble DML structure is presented, which is further deployed in the whole SHRS-M3DP to predict and recommend diabetes disease in the patients. It should deliver a concise and accurate representation of the experimental results, explanation, and the conclusion of experiments that can be drawn.

Smart Healthcare Recommendation System for Multidisciplinary Diabetes Patients.
is part explains the overall structure of a smart healthcare recommendation system for multidisciplinary diabetes disease patients (SHRS-M3DP). Initially, the general structure of the proposed SHRS-M3DP is described. en, a proposed system's structure is divided into distinct layers for an accurate description of each layer working. Finally, the ensemble deep learning model structure is presented, and it is further deployed in the whole SHRS-M3DP to predict and recommend diabetes disease in the patients. e proposed structure of the SHRS-M3DP system is exhibited in Figure 1. It is divided into two main segments: (1) training phase and (2) validation phase. ese phases are essential for the accurate prediction of multidisciplinary diabetes disease. ese phases communicate via a cloud. e training phase comprises seven levels: (i) sensory layer, (ii) EHRs layer, (iii) raw feature layer-1, (iv) raw feature layer-2, (v) fused raw feature layer, (vi) preprocessing layer, and (vii) application layer. e sensory layer comprises input parameters, including age, family history, glucose, skin thickness, blood pressure (BP), pregnancies, insulin, and body mass index (BMI). e sensory layer's input values are collected and transferred to the database, raw feature layer-1, through the Internet of medical things (IoMT). Because wireless communication is applied, data collected from multiple feature nodes and stored in the database may be inaccurate. For that reason, we considered such kind of data as feature raw data. EHRs layer consists of lab reports, questions, observation, and the patient's medical history. All the data collected from the EHRs layer are reports and need some methodology to convert them into a structured format for further processing.
e Framingham risk factors (FRFs) methodology used in the smart healthcare monitoring system for heart disease prediction [39] is adopted to extract data from the EHRs, as shown in Figure 2, and stored in raw feature layer-2. e data fusion approach is then applied for fusing the common features of both sensory data and EHRs to generate enhanced healthcare data on multidisciplinary diabetes disease.
ese fused feature data are then stored in the fused feature layer for further processing to predict diabetes disease. e following preprocessing layer plays a crucial role in the model. All deficiencies received through the sensory layer previously via wireless communication and EHRs layer are preprocessed in this phase.
ese missing values are managed by applying moving averaging and normalization methodology to mitigate noisy results. Subsequently, after  is layer is further divided into two sublayers: (i) prediction layer and (ii) performance layer. In the prediction layer, an ensemble DML model is applied to predict multidisciplinary diabetes disease. e ensemble deep learning combines several individual models to obtain better generalization performance of any predictive classification problem. e convergence process of the ensemble ML model is implemented in three ways: (1) max voting, (2) averaging, and (3) weighted average for classification. e ensemble ML model used an advanced boosting technique for regularizing, limiting the overfitting issue, and producing better accuracy compared to other ML models. As a result, the response rate of the ensemble ML model is ten times faster compared to other ML models. In the ensemble ML model's boosting technique, the trained dataset is divided into multiple weak learners. e average error rate of one weak learner is updated in the next weak learner. Resultantly, the final strong learner was found to have a minimal error rate for prediction.
Ensemble deep machine learning classifier can be expressed as follows: where F(x) is a strong learner of ensemble classifiers, α m is weight calculated by considering the last iteration's error, and h m (x) is a weak learner In this way, we can achieve maximum prediction accuracy. e operational flow of the advanced boosting ensemble DML model is shown in Figure 3 (Algorithm 1). e results are then sent to the performance layer. In this layer, data received from the previous layer is calculated. e performance layer results are evaluated based on accuracy, precision, recall, F1, root mean square error (RMSE), and mean average error (MAE) achieved by the SHRS-M3DP model. After comparing results, "YES" indicates that our proposed SHRS-M3DP model successfully predicted diabetes disease, and "NO" means the prediction layer of the proposed SHRS-M3DP model will be modified till the learning criteria objectives are attained. After successfully training the proposed SHRS-M3DP model, the trained fused model moves to a cloud to further import and predict diabetes disease. e validation phase comes in the last where trained fused SHRS-M3DP model is imported for prediction to authenticate whether the patient is affected with multidisciplinary diabetes disease based on the results. e results are then sent to the next layer, called the performance layer. In this layer, the data received from the prediction layer is evaluated. e performance layer results are then evaluated based on accuracy, precision, recall, F1, root mean square error (RMSE), and mean average error (MAE) achieved by the SHRS-M3DP model. After comparing the results, "YES" indicates that our proposed SHRS-M3DP model successfully predicted diabetes disease, and "NO" means the prediction layer of the proposed SHRS-M3DP model will be updated until the learning criteria are achieved. After successfully training the proposed SHRS-M3DP model, the trained fused model moves to a cloud to further import and predict diabetes disease. e last phase is the validation phase. In this phase, the trained fused SHRS-M3DP model is imported for recommendation to validate whether the patient is affected with multidisciplinary diabetes disease based on the results.

Experiments
e data collected from EHRs and sensors were discussed previously in the proposed SHRS-M3DP model. In addition, the fused feature database was also discussed in the last

Performance Evaluation.
e experiment was carried out to indicate the proposed SHRS-M3DP model's performance for diagnosing diabetic disease. Initially, the data was collected from sensors, which were transferred through IoMT to the feature database. Similarly, the patients' data collected through lab reports, questions, observations, and medical history were converted from unstructured format to structure format for further preprocessing. After collecting the features from sensors and EHRs, both datasets' features were combined to make a rich health dataset for better prediction and recommendation of diabetes disease. Finally, (1) Initialize model with a constant value: (1) Compute so-called pseudo-residuals: (3) Compute multiplier r m by solving the following one dimensional optimization problem : ALGORITHM 1: Ensemble ML-based diabetes prediction. Computational Intelligence and Neuroscience the processing module analyzed the final combined, fused feature dataset for further processing. Furthermore, the Hospital Frankfurt Germany diabetes dataset and Pima Indians diabetes dataset were then utilized for training the diabetes disease prediction model. For evaluation purposes, the proposed ensemble deep learning model was compared with some other classifiers: SVM, LR, KNN, NB, RF, and DT. e proposed SHRS-M3DP model was used before and after the feature selection and performance was compared. e datasets were divided randomly into 80% and 20%, respectively, to train and test the models mentioned above in the proposed model.

Evaluation
Metrics. Dissimilar evaluation metrics were used to conclude the model's overall efficiency, as shown in Table 3

Results.
is section presents the results of the abovementioned proposed model and a comparison with other classifiers, respectively. e complete details of all classifiers for diabetes prediction are divided into three parts: prediction of diabetes disease, Pima Indians diabetes dataset consisting of 768 cases with eight features, Hospital Frankfurt Germany diabetes dataset consisting of 2000 patients with eight features, and finally with fused features dataset having 2768 cases with eight features as shown in Table 2, respectively. e proposed SHRS-M3DP ensemble deep learning model prediction accuracy with other baseline classifiers is shown in Figure 4. e comprehensive explanation of each classifier before data fusion and after data fusion is as follows: e proposed ensemble deep machine learning model performed outstandingly as compared to all the baseline classifiers. e proposed ensemble DML classifier's accuracy in dataset 1 is 72.7% for predicting diabetes disease with 786 cases. However, it is low due to the small dataset. In dataset 2, ensemble ML performed better and achieved 91% accuracy with 2000 cases, higher than all other classifiers. In the final data fusion dataset, the accuracy of prediction of diabetes disease with the proposed ensemble ML model is 99.6%, having a minimal error rate. e summary of performance metrics, accuracy, precision, recall, F1, root mean square error, and mean absolute error of selected datasets individually and data fusion datasets, is presented in Tables 4-6.
Different DML classifiers are compared with the proposed model by various evaluation metrics, as shown in Table 4. In this experiment, only Pima Indians diabetes dataset is considered, without feature selection technique. e performance of each metric on a given dataset is precisely shown in Table 4. e proposed model's overall performance is less compared to the other classifiers due to the small dataset and the absence of feature selection technique.
In Table 5, only the Hospital Frankfurt Germany diabetes dataset is considered, without feature selection technique. Different DML classifiers are compared with the proposed model by various evaluation metrics, as shown in Table 5.  Figure 4: Accuracy of models before and after data fusion.    Figure 4. erefore, our proposed deep machine learning model can achieve more accurate results concerning a higher dataset ratio. In this experiment, the proposed model achieved 91% accuracy, much higher compared to other DML classifiers. On the other hand, in this experiment, RMSE and MAE are also very low compared with different DML classifiers's RMSE and MAE.
In Table 6, both diabetes datasets are being considered with the data fusion technique. Different DML classifiers are compared with the proposed model by various evaluation metrics, as shown in Table 3. e performance of each metric on a given dataset is precisely shown in Table 6. e proposed model's overall performance is much higher compared to the other classifiers due to the fused technique by making a rich healthcare dataset. Our proposed model performed outstandingly due to the data fusion technique and produced extraordinary results. In this experiment, the proposed model achieved 99.64% accuracy, much higher compared to other DML classifiers. In this experiment, RMSE is 0%, and MAE is only 0.06%. As shown in Table 6, all other classifiers produce results less than 85%.
Our proposed model accurately performed well with the help of data fusion technique. Our proposed ensemble DML model has achieved higher accuracy compared to other studies done in the recent past. e details of recent studies on diabetes with their authors are also summarized and shown in Table 7. e significant contribution for higher accuracy in our model is due to data fusion. In this way, we have made a rich healthcare dataset for the prediction of multidisciplinary diabetes disease. In this way, we have achieved higher accuracy compared to other studies, which is 99.6%.

Conclusion
e prediction of human diseases, particularly multidisciplinary diabetes, is challenging for better and timely treatment. A multidisciplinary diabetes illness is a lifethreatening disease worldwide which attacks major essential human body parts. A proposed SHRS-M3DP model is presented to predict and recommend multidisciplinary diabetes disease in the patients quickly and efficiently. e ensemble deep ML model and data fusion technique are used for fast response and better accuracy rate. e proposed model efficiently predicted and recommended whether the patient is a victim of multidisciplinary diabetes disease or not. e proposed SHRS-M3DP model can also identify the effect of human body parts: Neuropathy, Retinopathy, Nephropathy, or Heart. e proposed SHRS-M3DP model simulation is made by using Python language. Finally, the study of this research concluded that the proposed SHRS-M3DP model's overall performance is 99.6%, which is outstanding compared to previously published approaches.

Contribution.
Many recommendation systems for healthcare have already been proposed in recent researches. e significant contribution of this research is to enrich the healthcare dataset for the best prediction of multidisciplinary diabetes disease. We have collected the patients' data through wearable sensors and EHRs in the textual record form of each patient. After collecting the records of each patient, essential data from both ends are fused to enrich the healthcare dataset. e ensemble deep learning approach works accurately and produces better results in larger healthcare datasets. Finally, we have developed a better recommendation system by collecting patients' records and applying an ensemble machine learning approach for  Computational Intelligence and Neuroscience accurate and timely prediction and recommendation for multidisciplinary diabetes disease patients. e overall performance of our recommendation system is 99.6%. In this way, future academic research and practices will be helpful for new researchers in this medical field, especially for automated prediction and recommendation systems for human diseases.

Future Work.
e proposed SHRS-M3DP recommendation system achieved overall good performance. However, there is still a need to work for a better generalized efficient prediction and recommendation system for all human diseases. e complexity of the deep ensemble algorithm will also be considered in the near future for accurate and quick results of this algorithm.
Data Availability e data used in this paper can be requested from the corresponding author.

Conflicts of Interest
e authors declare that they have no conflicts of interest regarding the publication of this work.