Keywords

1 Introduction

Diabetes is emerged as a major healthcare problem in India and every year it is affecting large number of people. The data science based Knowledge Management System (KMS) in health care industry is getting attention to draw effective recommendations to cure the patient in its early stages [1, 2]. The knowledge augmented through KMS is an asset for society and incremental learning triggers knowledge augmentation [3, 4]. Online interactive data mining tools are available for incremental learning [5]. The threshold acts as a key in incremental learning to investigative formed closeness factors [6]. This approach in a way may change pattern of diabetes diagnosis [6–10]. In this study proposed TBCA is applied on the values of attributes that are collected from patient’s medical reports. TBCA implementation unleashes hidden relationships among attributes to extract impactful and non impactful attributes for diabetes mellitus.

In Sect. 2, TBCA is presented. In the following sections i.e., in Sect. 3 the methodology used for its implementation, in Sect. 4 analysis of obtained results, in Sect. 5 concluding remarks and at the last section, references used to carry out this study are listed.

2 TBCA

This section presents a high level pseudo code for TBCA in two parts to show TBCA is an extended version of Closeness Factor Based Algorithm (CFBA).

3 Methodology Used to Implement TBCA

TBCA data set considers medical reports of working adult diabetic patients having age group between 35–45 years for the year 2015–2016. TBCA works in three different phases as mentioned below:

  1. (1)

    In pre-processing input is taken as a CSV file and closeness factor value is calculated by taking into account different possibilities like sum wise, series wise, total weight and error factor for each data series set. The computed values are exported as a CSV file.

  2. (2)

    In clustering, clusters are formed based on closeness values that are generated through preprocessing for a particular data series and formed clusters are stored in a new CSV file in an incremental fashion.

  3. (3)

    Post clustering phase is used to extract values of attributes from the formed clusters for further analysis. The attributes related to diabetes mellitus are extracted on the basis of threshold where lower limit is mean of a cluster and upper limit is its higher value. These eight attributes are mentioned in Table 1 where first four are impactful and remaining are non impactful. The following figures represent processing done on 5 K data sets during phases of TBCA in a single and in multiple iterations.

    Table 1 Impactful and non impactful attributes for diabetes mellitus
    Full size table

4 TBCA’s Analysis

TBCA aims to find out impactful and non impactful attributes and for the same following types of analysis are carried out.

  1. (1)

    Related attributes analysis: The mean value of each attribute of every cluster is taken into account to analyze related attributes in a single and multiple iterations on data sets as shown in Figs. 1 and 2. The graphs for some of the related attribute analysis are shown below and they depict their behaviour pattern graphically (Fig. 3).

    Fig. 1
    figure 1

    Processing of 5 K data series in single iteration of TBCA

    Full size image
    Fig. 2
    figure 2

    Processing of 5 K data series in multiple iterations of TBCA

    Full size image
    Fig. 3
    figure 3

    HDL versus Non HDL Cholesterol, VLDL versus Non HDL Cholesterol analysis

    Full size image
  2. (2)

    Outlier analysis to extract impactful attributes: The outlier deviation analysis of datasets with extracted eight attributes is carried out which results in depiction of the deviation of the outlier values from the cluster deviation values. The generated pattern in shown in outlier analysis and it is observed that outlier detection in clustering plays a vital role. The patterns depicted via the statistical graph in Cluster 2 deviation versus outlier deviation for diabetes datasets in Fig. 4. In Fig. 4, after analysis of deviation of each cluster against the outlier deviation, it is observed that attributes BLOOD GLUCOSE FASTING, BLOOD GLUCOSE PP, CHOLESTEROL and TRIGLYCERIDES are the main factors that are responsible for the generation of the outliers as deviation of the other cluster attributes are overlapping with the outlier deviation. This pattern is cross verified through cluster 2 averages versus outlier average graph shown in another part of Fig. 4.

    Fig. 4
    figure 4

    Clusters, outlier average and clusters deviation, outlier deviation analysis

    Full size image

4.1 Accuracy/Purity of TBCA

The following formula is used for calculation of accuracy or purity of TBCA.

$$ = \left( {100 - \frac{{({\text{Clustering}}\,{\text{value}}\,{\text{of}}\,{\text{multiple}}\,{\text{iteration}} - {\text{Clustering}}\,{\text{value}}\,{\text{of}}\,{\text{single}}\,{\text{iteration}})}}{{{\text{Clustering}}\,{\text{value}}\,{\text{of}}\,{\text{multiple}}\,{\text{iteration}}}}{ \times }100} \right) $$

where clustering value = cluster count for cluster that contains maximum clustered data for a particular iteration.

The accuracy/purity of TBCA is based on clustering value for single iteration and in multiple iterations on same dataset. As shown in Figs. 1 and 2, the first cluster has the maximum weight age (42 and 46 % of the total data resides there) and hence it contains maximum clustered datasets. Therefore, the cluster count or clustering value of this cluster is used calculate the accuracy or purity of TBCA. This accuracy signifies processing of raw datasets and creation precise clusters in single as well as multiple iterations as shown in Figs. 1 and 2 over the same datasets. The multiple iterations on same dataset work in an incremental fashion and confirm cluster members independent of their order, CFBA parameters.

5 Concluding Remarks and Outlook

TBCA proved to be very useful in obtaining inter attribute relationship and outlier value knowledge over various iterations in an accurate manner which eventually triggered towards finding of key attributes related to diabetes mellitus. TBCA has showed 91.9 % of accuracy over single or in several iterations on data set under consideration. It can be effectively used in healthcare domain for prediction of a particular disease like diabetes mellitus. It involves novel mechanism of formation of clusters based on closeness factor and then by using threshold to extract required attributes leading to crisp prediction of impactful set of attributes among them for diabetes mellitus. If a person is suffering from diabetes mellitus properly keeps track of impactful attributes then he/she can manage to cure at early stages. These extracted impactful attributes can act as a catalyst for IT industries for those that are working on medical reports of patients in order to suggest life style management recommendations to cure them from certain diseases. These impactful attributes can also bring revolution in diabetic mellitus patient’s treatment in terms of test on a patient for its diagnosis. TBCA algorithm in turn plays a vital role in augmentation of generated knowledge for diabetes mellitus and may also change current way of pathology practices for diagnosis of diabetes mellitus. So, TBCA may prove best in all other disease prediction, being applied across domain, not restricted.