Data mining of common laboratory tests can be used to identify patients at risk

Abstract

In deteriorating hospital patients, delayed intervention is associated with increased risk. Common laboratory tests may help identify such patients earlier. We studied the relationship between commonly measured laboratory variables and the occurrence within the next day of a rapid response team (RRT) call or unplanned intensive care unit (ICU) admission or death in patients admitted to hospital for >24h or presenting to the Emergency Department (ED) over a 6 year period. We analysed 5.9 million individual measurements and 559,612 multi-test batches of laboratory tests. We specifically studied 405,826 batches in 41,417 ward patients (average age 64.8 ± 17.5 years) and 153,786 batches in 65,363 ED patients (average age 58.1 ±20.7 years). Among 106,780 patients admitted to hospital, 1024 had at least one RRT call, 142 at least one unplanned ICU admission and 1286 died. In ward patients, total CO 2 (bicarbonate equivalent) had an AUC-ROC of 0.713 for imminent unplanned ICU admission and plasma urea had an AUC-ROC of 0.774 for imminent death. In ED patients, pH had an AUC-ROC of 0.819 for imminent unplanned ICU admission and creatinine had an AUC-ROC of 0.756 for imminent death. In addition, combinations of variables resulted in AUC-ROC values in the 0.80-0.88 range. Laboratory tests could, on average, predict an RRT call 10.2 (ward) or 11.9 (ED) hours before it occurred. In conclusion, data mining of common laboratory tests can be used to quantify risk and assist in the identification of high risk ward or ED patients