Advertisement

ML1 Comparing Mortality in Cardiac Patient Surgical Clusters with Machine Learning Clusters in the National Inpatient Sample

      Objectives

      This study investigates mortality in cardiac patient clusters based on surgery type versus patient clusters created through unsupervised machine learning (ML).

      Methods

      The 2017 National Inpatient Sample describes US patient discharges and is provided by the Healthcare Cost and Utilization Project (HCUP). Patients included in this study were ≥18 years old with a “Major Therapeutic” primary cardiac procedure per HCUP Procedure Classes and Clinical Classification Software, and with a complete discharge record. Clusters were created through two different methods: 1) based on the three most common cardiac procedures; 2) based on patient and hospital characteristics, independent of mortality, through the ML algorithm K-prototypes.

      Results

      A total of 170,326 discharges met inclusion criteria. The three prevalent cardiac procedures were percutaneous transluminal coronary angioplasty (PTCA) – 40.2%, coronary artery bypass graft (CABG) – 16.1%, and heart valve procedures (HV) – 15.0%. The prevalent procedures within each ML cluster were: Cluster 1: PTCA – 31.2% and CABG–22.6%; 2: HV – 30.1% and CABG – 20.5%; 3: PTCA – 73.7% and CABG – 8.6%. The surgery clusters contained 121,423 discharges, while the ML clusters contained all 170,326 discharges. While the average Elixhauser Comorbidity Indices (ECI) based on the surgery clusters were different (PTCA: 2.1; CABG: 3.6; HV: 4.6; p<0.0001), the ML clusters revealed a clear difference in the average ECI (Cluster 1: 9.8; 2: 2.9; 3: 0.8; p<0.0001). While the mortality rate within each surgical group was different (PTCA: 1.6%; CABG: 1.7%; HV: 2.3%; p<0.0001), the ML clustering exposed a stark distinction in mortality between clusters (Cluster 1: 7.6%; 2: 0.8%; 3: 0.7%; p<0.0001).

      Conclusions

      A novel application of unsupervised ML in cardiac surgical patients identified a high mortality cluster otherwise missed by traditional classification. This high mortality cluster warrants further research to understand the typical patient journey and support treatments that may reduce the mortality rate.