Unsupervised Machine Learning Elicits Patient Archetypes in a Primary Percutaneous Coronary Intervention Service

A primary percutaneous coronary intervention (PPCI) re-establishes blood flow in an obstructed coronary artery. PPCI referrals vary in admission criteria partly on the basis of ECG findings, hence, not all the referrals are accepted. The aim of the paper is to discover archetypes of accepted patients referred to the PPCI center. Cluster analysis was performed on a PPCI referral dataset to identify patient archetypes and identify any key patterns of patients who were accepted for PPCI. A k-means clustering algorithm was used with the elbow method for determining the optimum number of clusters (groups of patients). A silhouette plot was generated for within cluster validation. Among the accepted PPCI referrals, there were four different groups of patients. The patients within each group have similar characteristics. The largest cluster of patients include male patients being referred out of hours and with excessive door to balloon times (DTBTs) as compared to those referred in hours. Another cluster includes older female patients who are referred out of hour. Also, it was discovered that the false activation rate and DTBTs are higher in females as compared to male clusters. The smallest cluster include the most elderly patients in the whole referral dataset and mainly includes more males than female's who are referred out of hour and have the highest false activation rate, DTBT, and 30 days mortality rate. The cluster analysis of PPCI dataset revealed different patient archetypes. Each group of patients have a different mean age, out of hours referral rate, DTBT, false activation and 30 days mortality rate compared to other group. The identified clusters could be helpful for the clinicians to better understand their patients and utilize this information to aid the clinical decision making.