Anomaly Detection in Clinical Data of Patients Undergoing Heart Surgery

We describe two approaches to detecting anomalies in time series of multi-parameter clinical data: (1) metric and model-based indicators and (2) information surprise. (1) Metric and model-based indicators are commonly used as early warning signals to detect transitions between alternate states based on individual time series. Here we explore the applicability of existing indicators to distinguish critical (anomalies) from non-critical conditions in patients undergoing cardiac surgery, based on a small anonymized clinical trial dataset. We find that a combination of time-varying autoregressive model, kurtosis, and skewness indicators correctly distinguished critical from non-critical patients in 5 out of 36 blood parameters at a window size of 0.3 (average of 37 hours) or higher. (2) Information surprise quantifies how the progression of one patient's condition differs from that of rest of the population based on the cross-section of time series. With the maximum surprise and slope features we detect all critical patients at the 0.05 significance level. Moreover we show that a naive outlier detection does not work, demonstrating the need for the more sophisticated approaches explored here. Our preliminary results suggest that future developments in early warning systems for patient condition monitoring may predict the onset of critical transition and allow medical intervention preventing patient death. Further method development is needed to avoid overfitting and spurious results, and verification on large clinical datasets.