S110 Memorial Union
Early hospital readmissions can negatively impact patients’ quality of life and hospitals’ income. Intensive efforts have been made to develop 30-day hospital readmission risk prediction models. Most of the reported statistical and machine learning models were built with static datasets. The reality, however, is that data arrives sequentially and the trend may evolve over time. As a result, many static models failed in practice. To make the models up-to-date, the traditional approach is re-training them from scratch periodically, which can be expensive. Most importantly, it is difficult to determine and discard outdated knowledge from the models using this approach. In this work, we aim to develop a self-adaptive 30-day hospital readmission prediction model using a data stream mining approach. We used a diabetic inpatient encounter data stream to dynamically train and evaluate models based on incremental learning algorithms. This approach and preliminary results will be presented in the seminar.