INTRODUCTION. Unplanned readmission to intensive care is highly undesirable in that it contributes to increased variance in care, disruption, difficulty in resource allocation and may increase length of stay and mortality particularly if subject to delays. Unlike the ICU admission from the ward, readmission prediction has received relatively little attention, perhaps in part because at the point of ICU discharge, full physiological information is systematically available to the clinician and so it is expected that readmission should be largely due to unpredictable factors. However it may be that there are multidimensional trends that are difficult for the clinician to perceive that may nevertheless be predictive of readmission. OBJECTIVES. We investigated whether machine learning (ML) techniques could be used to improve on the simple published SWIFT score [1] for the prediction of unplanned readmission to ICU within 48 hours. METHODS. We extracted systolic BP, pulse pressure, heart and respiration rate, temperature, SpO2, bilirubin, creatinine, INR, lactate, white cell count, platelet count, pH, FiO2, and total Glasgow Coma Score from ICU stays of over 2000 adult patients from our hospital electronic patient record system. We trained our own custom multidimensional / time-sensitive algorithmic ML system to predict failed discharges defined as either readmission or unexpected death within 48 hours of discharge. We used 10-fold cross validation to assess performance. We also assessed the effect of augmenting our system by transfer learning (TL) with 44,000 additional cases from the MIMIC III database. RESULTS. The SWIFT score performed relatively poorly with an AUROC of around 0.6 which our ML system trained on local data was also able to match. However when augmented with an additional dataset by TL, the AUROC for the ML system improved statistically and clinically significantly to over 0.7. CONCLUSIONS. Machine learning is able to improve on predictors based on simple multiple logistic regression. Thus there is likely to be information in the trends and in combinations of variables. A disadvantage with this technique is that ML approaches require large amounts of data for training. However, ML approaches can be improved by TL. Basing prediction models on locally derived data augmented by TL is a potentially novel approach to generating tools that customised to the institution yet can exploit the potential power of ML algorithms. REFERENCES [1] Gajic O, Malinchoc M, Comfere TB, et al. The Stability and Workload Index for Transfer score predicts unplanned intensive care unit patient readmission: initial development and validation. Crit Care Med. 2008;36(3):676–82. Grant Acknowledgement This work was internally funded.