Learning temporal rules to forecast instability in intensive care patients

Abstract

INTRODUCTION
We study the utility of time series features and exceedences detected in a set of vital signs (VS: heart rate [HR], respiratory rate [RR], systolic blood pressure [SysBP], diastolic blood pressure [DiaBP] and peripheral oxygen saturation [SpO2]) of Step Down Unit (SDU) patients in forecasting their instability. For each vital sign, a range of acceptable values is defined. A patient is considered unstable if one of their VSs exceeds its acceptable range (Table 1) for at least four minutes. We aim at predicting onset of any such instability ahead of time.

In the experiments, we consider 305 stays at one SDU. VSs of each patient are sampled every 20 seconds, and decision rules from Table 1 are used to flag potential instability. Each event of exceedence is evaluated by expert clinicians and labeled either as true instability or a likely artifact due to caveats of the measurement procedures.

OBJECTIVES
Our primary objective is to understand how accurately, how confidently, and how far in advance we can forecast onset of episodes of true instability.

METHODS
We use temporal rule learning methodology to build forecast models for instability1. The training data consists of the clinician-confirmed actual episodes of instability (separately flagged due to specific vitals: HR, RR, SysBP, DiaBP, SpO2) to be predicted, and the preprocessed raw VS data as candidate predictors. Preprocessing involves computing basic statistics of VS waveforms such as moving averages, moving trends, and cross-correlations. Scalar values of thusly featurized vitals are used as state variables, and their abrupt changes as discrete events, to jointly form the input space for the rule learning algorithm. It then searches for human readable, statistically confident temporal association rules with sufficient support in data and high temporal predictive accuracy (see Fig.1 for examples). We evaluate the utility of inferred rules using cross validation: data is split into multiple disjoint folds so that no single SDU stay could be used both in training and testing phases of evaluation.

RESULTS
Several temporal association rules of significant utility have been derived from data. A subset of complementary rules with the highest confidence, support, and precision was then assembled and evaluated jointly as a forecasting expert system. Fixing the false discovery rate to one per day per patient we observe recall rates of instability episodes of 83%/49%/42% respectively for HR/RR/SpO2 alerts, when the forecast horizon is 4 to 30 minutes ahead.

CONCLUSIONS
Learning temporal rules from multivariate VS waveforms allow s forecasting instability in SDU patients. Presented approach can enable pre-emptive treatment of emerging instability.