Combination of Dynamic Analysis and Data Mining for Stability Prediction after Cardiac Surgery.

Van Loon K.1, Guiza F.², Meyfroidt G.3, Aerts J.-M.1, Blockeel H.², Van den Berghe G.3, Berckmans D. 1

1Division Measure, Model &Manage Bioresponses, Katholieke Universiteit Leuven, Kasteelpark Arenberg 30, B-3001 Leuven, Belgium, e-mail: .

²Department of Computer Sciences, Celestijnenlaan 200a, B-3001 Leuven, Belgium

3Department of Intensive Care Medicine, University Hospital Gasthuisberg, Herestraat 49, B-3000 Leuven, Belgium

Objectives

Intensive care unit (ICU) capacity planning depends on estimates of future available beds, and future discharge of patients. The objective of this study was to develop an early warning monitor that predicts the probability that a patient can be disconnected from mechanical ventilation within the first hours after cardiac surgery, and to investigate whether dynamic information of the first four hours of ICU stay of individual patients can improve predictions based on static admission data.

Methods

The criteria to start the weaning from mechanical ventilation are: hemodynamic and respiratory stability, absence of bleeding and normothermia. Models will be built to predict the first window of 30 minutes where these stable clinical conditions are reached. The considered prediction task was stated as follows: Predict the probability that the patient will begin to satisfy the stability criteria within each of the following time frames (classes): class 1: earlier than nine hours after admission; class 2: later than nine hours after admission.

Data of a total of 203 patients admitted to the ICU after planned coronary bypass surgery were analyzed. We selected five physiological input variables that were routinely monitored in these patients and saved with a sample interval of 1 minute in the Patient Data Management System (Metavision ®, iMD-Soft ®): heart rate (bpm), systolic arterial blood pressure (mmHg), systolic pulmonary pressure (mmHg), blood temperature (°C) and oxygen saturation. On the basis of these five physiological variables, different dynamic features were extracted: the means and standard deviations of the signals (Avgstd), parameters of multivariate autoregressive (MAR) models and cepstral coefficients (CEP). These sets of features served subsequently as inputs for a Gaussian process (GP) and the prediction results were compared with the case where only admission data was used for the classification. GP models will be compared against a logistic regression predictive model.

Results

Table 1 gives an overview of the obtained aROCs as well as the Brier scores for each experiment with the GP. The dynamic features, especially the cepstral coefficients (aROC: 0.749, Brier score: 0.206), resulted in significantly higher performances when compared to static admission data (aROC: 0.547, Brier score: 0.247). In all cases, the Gaussian process classifier outperformed logistic regression (LOGREG), which was calculated for comparison purposes.

Table 1: The aROC values and Brier scores of the different experiments with the GP classifier.

aROC / Brier Score / LOGREG / GP
Admission data (7 parameters) / 0.543/ 0.249 / 0.547 / 0.247
Avgstd (20 parameters) / 0.628 / 0.241 / 0.713 / 0.214
MAR (25 parameters) / 0.591 / 0.250 / 0.708 / 0.219
CEP (25 parameters) / 0.542 / 0.247 / 0.749 / 0.206

Conclusion

The performance of GP models to predict respiratory weaning can be significantly improved by using advanced calculations of the dynamic properties of time series of heart rate, systolic arterial blood pressure, systolic pulmonary pressure, blood temperature and oxygen saturation, when using aROC and Brier scores as evaluation criteria. When compared to logistic regression, GP classifier results in better performances in all cases. Dynamic analysis of data measured by patient data management systems has the potential to offer a valuable additional tool for monitoring the status of individual patients at the bedside.