Robust regression methods for intensive care monitoring

Introduction

Alarm generation of modern patient monitoring systems still predominantly relies on simple threshold methods. This leads to an unacceptably high rate of false positive alarms. Many false positive alarms are generated by measurement artefacts and measurement noise. One approach to address this problem is to alarm on the underlying signal (that is, the noise-free time series of the physiological variable), instead of the raw measurement.

Methods

Monitoring time series were simulated. Against these data four robust regression methods were evaluated: least trimmed squares (LTS), least median of squares (LMS), repeated median (RM), and deepest regression (DR). Moreover, online monitoring series from critically ill patients during multiparameter monitoring were also compared.

Results

LTS and LMS showed comparable behaviour, as did RM and DR. LMS and LTS provided only 20% efficiency, DR 61% and RM 70% (least squares regression = 100%). RM and DR had smaller standard deviations and smaller mean-squared errors than LMS and LTS under different noise distributions (standard deviation of online estimates based on sliding windows of size n = 21 for simulated standard normal errors: LMS: 0.875, LTS: 0.887, RM: 0.500, DR: 0.533). Analyses with clinical monitoring data also showed that LMS and LTS preserve sudden level shifts but are unstable and perform poorly with trend changes; RM and DR blur shifts but yield more stable estimations.

Conclusion

All four methods allow one to extract the underlying signal from physiological time series in a way that is robust against measurement artefacts and noise. However, there are significant differences between the methods. Overall, repeated median regression seems the best choice for intensive care monitoring since it is not only the most stable but also the fastest method.