All Times EDT
Keywords: Readmissions, Machine Learning, Predictive Modeling, Model Comparison, Classification, Survival Analysis, Random Forest
In this research, we compare different predictive models to examine whether including time information in the models would improve predictive performance compared to simply using a binary outcome. The outcomes for this study are three different disease-specific 30-day readmissions in patients over 65 years old including heart failure (HF) - 241,156 index admissions and 175 predictors (a factor with p levels is counted as p-1 predictors), myocardial infarction (MI) - 81,621 index admissions and 156 predictors, and pneumonia (PN) - 178,904 index readmissions and 167 predictors. The 30-day readmission rates are 20.57%, 19.41%, and 15.25% for HF, MI, and PN, respectively. The data are from the 2016 Nationwide Readmissions Database (NRD). Variables in NRD include ICD-10 diagnosis and procedure codes, anonymous demographic information, and data elements derived from HCUP software tools, including Clinical Classifications Software (CCS) code, chronic condition indicators, comorbidity indicators, procedure classes, and utilization flags. Models compared were logistic regression, logistic regression with LASSO shrinkage, Cox Proportional Hazard (CoxPH) regression, CoxPH regression with LASSO shrinkage, classification random forest, and survival random forest. Some models used a 50% subsample due to memory constraints. To compare survival and binary outcomes, the risk of 30-day readmission was computed from survival models as 1 – the probability of survival to day 30. Overall, the predictive performances from different models are very similar given the same outcome. Among the three outcomes, PN can be most accurately predicted, followed by MI, then HF. The AUCs for models predicting PN readmissions are around 0.62 - 0.63; for MI models, the AUCs are around 0.61 - 0.62; The AUCs for HF models are all around 0.59. Adding LASSO shrinkage slightly improves model performances. For the outcomes studied, the extra time information has no advantage in predicting 30-day readmissions.