To transition from the intensive care unit (ICU) to a general care floor (GCF), the medical team has determined that lower intensity monitoring poses no danger to the patient. A subset of GCF patients will decompensate and need readmission to the ICU. In this pilot study, we investigated whether there are early markers of decline in the electronic health record (EHR). ICU physicians, using clinical intuition, hypothesized that ICU readmission was preceded by tachycardia, desaturation, low hemoglobin, glycemic abnormalities, or hyperkalemia in the EHR. A remote surveillance system recorded a virtual alert (VA) when EHR values of these parameters crossed a threshold pre-determined by intensivists prior to the start of the study. Over 18-months, we followed 432 transfers out of the ICU, 66 (15.3%) of which were readmitted. The sensitivity of VA for readmission was 60.6% [95% CI 47.8 Ð 72.4] %, while the specificity was 11.2 [95% CI 8.2 Ð 14.9] %. The positive predictive value (PPV) of a VA for readmission within 48 hours of VA was only 5.2%. The low PPV of VAs chosen by expert clinicians in this study could lead to alarm fatigue, a sensory overload due to excessive alarms that can result in missing a true event. New data in this field suggest machine learning methods could more accurately predict ICU readmission by capturing nonlinear interactions between many various sources of real-time data in the EHR. In the future, we will use this approach to model daily ICU readmission risk.