Survey on clinical prediction models for diabetes prediction

As per literature, there are many forms of describing predictive analytics. It is inductive [8]. It doesn’t expect anything about data but it allows the data lead the way. It uses statics, machine learning, neural computing, robotics, computational mathematics and artificial intelligence to explore all data and find meaningful relationships and patterns. Predictive analytics is a set of business intelligence (BI) technologies that uncover relationships and patterns within large volumes of data that can be used to predict behaviour and events. To be more clear see the Fig. 1. Machine learning used by predictive analytics is a technique to train algorithm which can predict an output based on some input value. This leads to correlations and not to conclusions [9].

From previous work, there are two major types of predictive analytics such as supervised learning and unsupervised learning [8, 10]. Supervised learning is a process of creating predictive models using a set of historical data and produce predictive results. Examples are classification, regression and time-series analysis where as in Unsupervised learning does not use the previously known result to train its models. It uses descriptive statics. It identifies clusters or groups [8]. Further classification of predictive models are of nine types business rules, classification and decision trees [11‐13] naive Bayes, linear regression [10‐12], logistic regression [11, 12, 14], neural networks (NNs) [11‐13], machine learning, support vector machines (SVMs), natural language processing (NLP) [11]. In paper [13] the author has described seven types of regression models, each holding importance of its own. He listed seven types of regression models as linear regression, logistic regression, polynomial regression, stepwise regression, ridge regression, Lasso regression, elastic net regression. More versions of predictive models are described in two ways Smooth Forecast Model which describe smooth variable outcome for example profit and Scoring Model which describe binary outcome for example whether the blood report indicates disease or normal condition [9]. Another list of predictive models is linear models, decision trees, neural networks, clusters models, support vector machines, expert systems. The difference between simple linear model and generalized linear model is presented in Table 1 [10]. Furthermore, a list of models are Time-series analysis, Monte Carlo simulations, Statistics for spatial data [15].

It was listed that there are six steps for developing predictive models. Which are listed as follows project definition, exploration, data preparation, model building, deployment, model management [9].

Usage of predictive model by Organizations is summarized in paper [16] which is listed in Table 2.

A basic architecture about building predictive model consisting of three layers is presented in Fig. 2. What is important is how to implement model inputs and outputs, how model show growth when it is turned on or off, when to upgrade or replace according to ongoing time [11].

Depending on situation what model has to be selected is described as for segmentation use clustering algorithm, for developing recommender system use classification algorithm, Use decision tree when linear decision boundary is used, for predicting next outcome of time driven events use regression algorithms, to predict continuous values use regression, use naïve Bayes when features are conditionally independent, Machine learning is used for classifying text problems with ensemble model sometimes [11].

Deployment means using a model for the intended purpose. Most predictive models are built depending on the decisions made by the organization [17]. Various ways for deploying the predictive model are sharing the model, score the model, incorporate the model in a BI report, embed the model in application [8].

Predictive models are commonly assessed using c-statistics. This measure indicates whether the predictive model predicts positive or negative outcomes. If model c-statistics exceeds 0.7 it is considered as an acceptable prediction, if c-statistics is 0.5 then model prediction is not good [18]. c-statistics is equal to the area under ROC curve. Grading based on c-statistics is shown as in Table 3. Another way to assess the prediction model is Analysis of variance (ANOVA) when data is categorical [11]. The measures are listed in Table 4.

Some metrics that are considered for successful model are

Predictive analytics have huge applications in various fields like Homeland security, Crime prevention, Infrastructure management, Cyber security, Intelligent Transportation, Health care and bioinformatics, Text mining, Fraud detection, Social media and decision support [1], Credit scores, Credit card, fraud detection, Mail Sorting, Weather prediction, Hot dogs and hamburgers [16], not only about resource allocation but also about where or how should to allocate resource, what to expect from outcomes of model, how to manage the key drivers of the economic model for better outcome? [19]. There are four ways to monetize predictive models. The first way is for saving cost. The second way is using a predictive model for improving revenue. The third way is for returning investment and lastly is for risk management [20].

Predictive modeling tools [19] are summarized in Table 5.

From paper [16] author have listed existing predictive models and their usage

The clinical prediction model is very valuable as it can be applied to a various scenario like screening, prediction, medical decision making and education in health. In the medical field, the prognosis is considered as a fundamental component. In paper [22] a process of developing clinical prediction model is explained with five steps:

Predictive analytics uses regression models on available data for predicting outcomes mostly in the medical field. In past medical data was collected manually by handwritten, dictated or incomplete, this data was small for predictive modeling. But now a day because of EMRs, diverse and large digital data of health care is available as listed in Table 6. Natural processing languages are used to access unstructured data. This increases the quality of data and also quality prediction. The relative predictive power of a statistical model increases exponentially when using millions of patients instead of hundreds of patients. To discover relevant predictive variables clinical, claims, socioeconomic and care management data should be integrated to form one dataset [18].

A multi stage adjustment model with low misclassification rate which predicts which persons are most likely to develop diabetes is built by using KoGES dataset [23]. A physiological model which can predict the blood glucose level 30 min in advance was developed using five patients data by training SVR with physiological features. This helped in producing best results than doctors [24]. Another type of predictive model is sparse factor graph model. By using which diabetes complications are not the only forecast but also can discover the underlying associations between diabetes complications and lab test types. All algorithms were implemented in C++, and all experiments were performed on a Mac running Mac OS X with Intel Core i7 2.66 GHz and 4 GB of memory. The data set used for the experiment is collected from a geriatric hospital. The data set contain 1-year span data with 181,933 medical records, 35,525 patients data and 1945 types of lab tests. 60% of data was chosen for training the model and the rest for testing. The proposed model addresses two challenges feature sparseness and knowledge skewness [25].

A hybrid model has been developed to predict whether the diagnosed patient may develop diabetes within 5 years or not. A tool used for this purpose is WEKA and the data set was PIMA Indian diabetes data set. This hybrid model has achieved 92.38% accuracy [26]. The details of the hybrid model are shown in Fig. 3.

Another hybrid prediction model helps in producing optimal feature subset. This helps in detecting diabetes with high accuracy. To implement the model WEKA tool is used on PIMA Indian diabetic dataset. The proposed models have given an accuracy of 98.9247%. The procedure adopted by authors in developing predictive model is first preprocessed the dataset, then compute F-score values of features, select features with high F-score as discriminative features, then k-means algorithm is used to select feature subset that gives minimum clustering error and finally SVM is used to classification [27] as shown in Fig. 4.

In paper [28] the authors used two different types of neural networks to express which will output the accurate classifier in predicting diabetes. The two neural network models are multilayer neural network and probabilistic neural network. The dataset contains Pima Indian diabetes, having two classes and 768 samples. 576 samples were used for training and 192 were used for testing. The proposed methods were proved to better when compared with other previous methods.

In paper [29] the author developed a prediction model based on Hybrid-Twin Support Vector Machine (H-TSVM), which predicts whether a new patient is suffering from diabetes or not. They used Pima dataset for conducting an experiment. The factor that keeps this proposed method different from others is kernel function. The classifier produces an accuracy of 87.46%.

In paper [30] the author proposed a predicting model that classify type 2 diabetic treatment plans into three groups such as insulin, diet and medication. The dataset used for developing the model was JABER ABN ABU ALIZ clinic centre which contains 318 medical records. The model was developed using WEKA tool by applying J48 classifier and it has produced an accuracy of 70.8%.

In paper [31] the author developed a prediction model which predicts what are different types of disease a diabetic patient can develop. To develop the model a data set of 3 years span is collected from AR hospital with 739 patient details and 31 attributes. The pre processed data after deleting outliers by using distance based outlier detection (DBOD), is given as input to logistic regression model which was built by Bipolar Sigmoid Function that is calculated using Neuro based Weight Activation function. The model produced prediction accuracy of 90.4%.

In paper [32] a tool FNC was developed that can be used for diagnosing diabetes as shown in Fig. 5. The proposed model is developed by incorporating three techniques fuzzy logic, neural network and Case based reasoning with 200 patient details having 16 input attributes. Fuzzy logic and neural networks are implemented by using Matlab, Case based Reasoning is implemented by using MyCBR plug-in. After obtaining the result from three approaches rule based algorithm was applied to all three techniques to improve the accuracy. Finally, the best accuracy was obtained for case based reasoning.

In paper [33] author developed a hybrid model KSVM. The important criteria that make this model different from other methods are feature selection algorithm. PIMA data set was utilized to do experiments and results were produced. It was shown that diagnosis results using K-SVM are 99.74, 99.78, and 99.81 for learning experiments with amount 50, 60, and 70% data respectively, and 99.82, 99.85, and 99.90 for testing experiments with amount 50, 60, and 70% data respectively.

In paper [34] authors developed a prediction model that would predict whether a person would develop diabetes by considering daily lifestyle activities. To build prediction model PIMA diabetes data set was used and CART (Classification and Regression Trees) machine learning classifier was applied. The proposed model could provide an accuracy of 75%.

In paper [35] authors developed a prediction model that would predict whether a person develops diabetes or not. To achieve this PIMA diabetes dataset was used. In the proposed method first controlled binning technique is applied then multiple regression was used to improve the accuracy of the model. After incorporating all techniques an accuracy of 77.85% was achieved. The controlled binning technique which is innovative thought in this paper is calculated by using the Eqs. (1) and (2)

In paper [36] authors developed a decision tree model for the diagnosis of type 2 diabetes. They used Pima Indian diabetes dataset. Pre-processing techniques like attributes identification and selection, handling missing values, and numerical discretization was used to improve the quality of data. Weka tool was used, J48 decision tree classifier was applied to construct the decision tree model. The model produced an accuracy of 78.17%.

In paper [37] the authors developed a prediction model by using neural networks to classify and to diagnose onset and progression of diabetes. They have used 545 patients’ data from a diabetes clinic. First, they trained and tested neural networks with a different number of neurons and found a neural network with seven neurons has produced highest accuracy. The memetic algorithm is used to update weights which improved the accuracy of the model from 88.0 to 93.2%. this model was compared with other models too. But a neural network with seven neurons and application of memetic algorithm is observed as the best model (Figs. 6, 7).

In paper [38] the authors have developed an expert healthcare predictive decision support system that predicts diabetes. This model is trained on Pima diabetes dataset. Decision tree and K-nearest neighbor algorithms are used to develop the model and found that C4.5 algorithm has achieved 90.43% accuracy.

In paper [39] the authors have developed a prediction model using Chi squared test to find not only dependencies between factors but also independences. Then CART is applied to build a prediction model which has 75% accuracy. Data was collected through questioners from 200 people and model was built using R tool.

In paper [40] authors developed an elastic net model which improves the accuracy for estimating glucose. The authors have collected 45 experimental sessions data set from diabetic patients. The data was collected from a noninvasive glucose device i.e., a blood sample is not taken. Three models were constructed using regularized methods LASSO, Ridged and Elastic net model. The elastic net model has compared with LASSO, ridged and partial least square regression and found Elastic net model is best.

From all of the techniques and prediction models discussed above, we want a prediction model that predicts diabetes of a diagnosed person. Since this output can be obtained depending on the time we would lie to use regression model. Of all regressions, Elastic Net is most useful as categorical, numerical and image or signal form data can be given as input to the model. The elastic net regression model is a combination of LASSO (Least Absolute Shrinkage And Selection Operator) and Ridged Regressions. Thus elastic net regression support shrinkage of coefficients as well as grouping effect. One more interesting point is numerical, categorical and image form data can be given as input to the model.