Introduction: Despite advances, survival to discharge following in-hospital cardiac arrest (IHCA) remains less than 25%. Patients at risk of IHCA are often on continuous ECG monitoring, but no automated methods exist to leverage the extensive real-time physiologic information that it contains.Hypothesis: We hypothesize that identification of physiologic changes, particularly over a patient’s hospital course, as manifest on continuous ECG, by automated algorithms can help predict the risk of IHCA in real time.Methods: We conducted a retrospective study of 77 IHCA cases (PEA and asystole) and 1763 matched control patients. Continuous ECG data was processed automatically to derive signal-averaged metrics for PR, QRS duration (QRSd), ST, QTc, RR, QRS amplitude, and also presence of atrial fibrillation (AF) and pauses. We selected 2 consecutive 3-hour blocks (blocks 1 and 2) for each case and control: for cases these were selected from the 6-hour period immediately preceding IHCA (with block 2 immediately preceding IHCA); for controls these were chosen at random. Maximal positive and negative trends for each parameter was evaluated for each block, as well as the difference in these values between blocks 1 and 2. Multivariate logistic regression analysis was used to model each patient’s risk of IHCA during block 2, using only ECG-derived parameters. Bootstrap analysis with 500 iterations was used to estimate the performance of this model on a validation set.Results: Significant predictors of IHCA during block 2 included new AF and pauses (present in block 2 but not block 1), QRSD prolongation, ST elevation in lead II, decrease in QTc and increase in RR compared to the prior 3-hour period (block 1) (all p<0.001). Our model achieved an AUC of 0.81 on the training set and 0.80 (bootstrapped 95% CI 0.7958 – 0.8030) on the validation set. On the validation cohort, with a specificity of 90%, we were able to achieve a sensitivity of 63% for predicting IHCA.Conclusions: Significant ECG changes are observed in the 3 hours preceding IHCA, and our model was able to predict a majority of cases with high specificity using only changes in ECG parameters. With further refinement of these algorithms, ECG changes could be used to aid in the real-time prediction of patients at risk for IHCA.