Using Machine Learning Models to Predict the Likelihood of Patient Readmission to ICU

Intensive care units (ICUs) cater to individuals with severe injuries and illnesses. Therefore, the patients within suffer from acute anatomic and physiologic derangements requiring constant monitoring and more intense support from hospital staff. Their conditions necessitate rapid diagnosis and intervention of abnormalities to facilitate a period of recovery. A challenge arises, however, in the recognition of sufficient resolution of the pathophysiologic state such that the patient can be safely discharged to a lower intensity environment. The decision to discharge is currently based on the expertise of ICU clinicians, but there is currently no formal method to assist clinicians in predicting a patient’s chance of success or readmission, so this process is imperfect. As such, ICU readmission rates range from two to 20 percent. Readmission rates depend on a variety of factors, including demographic characteristics, comorbidities, severity of illness score, duration of index ICU stay, type of ICU, discharge destination, etc. Regardless of what factors led to readmission, a major problem manifests in the rates of in-hospital death for those who are readmitted to the ICU. Compared to patients who are successfully discharged, those who are discharged but return to the ICU are two to 10 times more likely to die in hospital. Unfortunately, current predictive models are insufficiently accurate. In general, these models are built on static parameters and don’t take advantage of the large amount of complex data available in the EHR, nor do they take time varying covariates into account. So, generating an algorithm that is able to use all of the available data in order to accurately predict readmission to the ICU would have several important impacts. By improving physicians’ ability to determine resolution of the pathophysiologic state and reduce premature discharge, we could expect morbidity and mortality rates to decrease as well as reduced healthcare costs for patients. Similarly, the information a predictive model provides would allow for better allocation of resources, by letting hospital staff know which ICUs have higher readmission rates and require more attention. Finally, the features examined during this project could potentially be generalized for use in future predictive models.