On-line peak detection in medical time series with adaptive regression methods

Sequential analysis of medical time series has important implications when data concern vital functions of the human body. Traditional monitoring of vital signs is performed by comparing the level of time series with predetermined target bands. As in industrial control charts, alarm signals start when the series level goes beyond the bands. Many physiological time series, however, are non-stationary with stochastic cycles and are characterized by turning points, i.e. periods where the slope of the series changes sign and determines the beginning of rise and fall phases. Turning points are useful indicators of the effect of drugs and medications and must be timely detected in order to calibrate the drug dosage or perform interventions. Sequential and recursive filters are suitable methods for change and turning points; they are mainly given by double exponential smoothing, time-varying parameter regression and prediction error statistics. Their tuning coefficients can be selected in a data-driven way, by optimizing score functions based on the series’ oscillation. Extensive application to real and simulated data shows that adaptive techniques represent effective solutions to on-line monitoring and surveillance in medicine.