Predicting intervention onset in the ICU with switching state space models

Abstract

The impact of many intensive care unit interventions has not been fully quantified, especially in heterogeneous patient populations. We train unsupervised switching state autoregressive models on vital signs from the public MIMIC-III database to capture patient movement between physiological states. We compare our learned states to static demographics and raw vital signs in the prediction of five ICU treatments: ventilation, vasopressor administra tion, and three transfusions. We show that our learned states, when combined with demographics and raw vital signs, improve prediction for most interventions even 4 or 8 hours ahead of onset. Our results are competitive with existing work while using a substantially larger and more diverse cohort of 36,050 patients. While custom classifiers can only target a specific clinical event, our model learns physiological states which can help with many interventions. Our robust patient state representations provide a path towards evidence-driven administration of clinical interventions.

Figures

Figure 1:

Illustration of data processing pipeline. (1) We extract vital signs and lab results (xn) are extracted from the database for a filtered selection of patients. (2) A switching-state autoregressive model is used the model the time series, generating belief states bn (the probability of each state at each time). (3) Static features are extracted for all patients (sn) - these are based on admission data and do not change over the course of the subject’s stay. (4) Given three possible sets of features for each timestep t and patient n - sn, xnt, and bnt - we train a classifier to predict the per-timestep outcome of interest ynt (e.g. vasopressor administration). Our system predicts the outcome ynt using features from either the immediately previous timestep fn,(t–1), or some further delay fn,(t–d).

Figure 2:

AUC scores for different features predicting the onset of each intervention at a delay of d ∈ {1, 2, 4, 8} hours ahead of the current timestep. Features: Each bar color denotes one feature or feature concatenation: static observa tions s (10 dimensions using one-hot encoding), dynamic time-series observations x (18 dimensions), and belief state vectors b (K = 10 dimensions) from the switching state model in Eq. (2). Interventions: fresh-frozen-plasma transfu sion (ffp), platelet transfusion, red-blood-cell (rbc) transfusion, vasopressor administration, and ventilator intubation.

Figure 3:

Learned classifier weights for each belief state under each separate intervention task, using fixed delay of 1 hour. The learned set of K = 10 hidden states is indexed by an integer from 0, 1, … 9. Large weight values indicate a state’s presence will cause the logistic regression classifier to raise the probability of the intervention.

Figure 4:

Average value of dynamic features xnt assigned to timesteps strongly associated with state index 9. Values are z-score standardized per variable. State 9 had very low observed spo2 and bicarbonate levels as compared to other states. Lactate levels were the highest observed across all states - no other state had significant positive lactate z-scores.