Clustering diabetic patients with comorbidity patterns are necessary to learn relationships between diabetes patients’ clinical profiles and as an essential pre-processing stage for analysis tasks, like classification and categorization. Nevertheless, the heterogeneity of these data makes traditional clustering methods more difficult to apply, necessitating the development of novel clustering algorithms. In this paper, we recommend an effective and scalable clustering technique suitable for datasets made up of attributes which are atomic and set-valued. In these datasets, each record corresponds to a different diagnosis detail of a diabetic patient based on his or her hospital visit, where diagnosis details in each record are represented using the International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) codes. Our proposed technique involves three main stages. In the first stage, we selected the top-k diabetes-specific comorbidities patterns. In the second stage, we ensured that the co-occurring conditions in the selected top-k diabetes-specific comorbidities patterns really co-occur together or not using topic modeling and in the last stage, we constructed high quality clusters efficiently using average linkage agglomerative clustering with cosine similarity. Also, based on silhouette analysis, we assessed the efficiency and effectiveness of our proposed technique using a large, freely available MIMIC dataset (MIMIC-III and MIMIC-IV), comprised of over 14,222 and 68,118 distinct records, respectively. Our findings reveal that our technique finds clusters that: (i) preserve interrelations between demographics (age, gender) and diagnosis codes (ICD-9-CM codes), and (ii) are well-separated and compact. Finally, the founded clusters are beneficial for numerous investigative tasks like query answering, visualization, anonymization, classification etc.

Data driven social security fraud detection has been given limited attention in research. Recently, social schemes have seen significant expansion across many developing countries including India. The fundamental aims of social schemes are to alleviate poverty, enhance the quality of life of the most vulnerable and offer greater chances to those relegated to the fringe of society to engage more enthusiastically in the society. Although governments channel billions of dollars every year in support of these social schemes, quite significant number of the eligible people are excluded from the program mainly through fraud and dishonesty. Although fraud is considered an illegal offence and morally reprehensible, it is unfortunate that the prevalence of fraud in social benefit schemes is rampant and a significant challenge to address. In this paper, we studied the viability of machine learning techniques in identifying fraudulent transactions in the context of social schemes. We focus on the detection of the false income level claims made by the fake beneficiaries to get the privileges of government scheme. We used the standard classifiers like Logistic Regression, Decision Trees, Random Forests, Support Vector Machine (SVM), Multi-Layer Perceptron and Naïve Bayes to identify fake beneficiaries of the government scheme from those deserving people. The results show that the Random Forest Classifier perform best providing an accuracy of 99.3% with F1 score of 0.99. The outcome of this research can be used by the government agencies entrusted with the management of the schemes to wade out the abusers and provide the required benefits to the right and deserving recipients.