Cluster Analysis of Obesity Disease Based on Comorbidities Extracted from Clinical Notes

Clinical notes provide a comprehensive and overall impression of the patient's health. However, the automatic extraction of information within these notes is challenging due to their narrative style. In this context, our goal was to identify clusters of patients based on fourteen comorbidities related to obesity, automatically extracted with the cTAKES tool from the i2b2 Obesity Challenge data. Furthermore, results were compared with clusters obtained from experts' annotated data. The sparse K-means algorithms were used in both experiment at two levels: at the first level, three clusters were found, and at the second, new clusters were found by applying the same algorithm to each of the clusters from the former level. The results show that three types of clusters could be identified based on the number of comorbidities and the percentage of patients suffering from them. Diabetes, hypercholesterolemia, atherosclerotic cardiovascular diseases, congestive heart failure, obstructive sleep apnea, and depression were the diseases with the highest weights contributing to the cluster distribution.