Purpose: ICU delirium is a dynamic condition and occurs in an environment with daily changes in risk. Current prediction models do not account for daily changes in risk and outcomes. We sought to develop and internally validate a dynamic risk model to predict daily changes in acute brain dysfunction in the ICU.
Methods: Using data from a multicenter prospective study of critically ill patients we developed a daily transition prediction model, using multinomial logistic regression that estimated 15 transition probabilities for each ICU day conditional on daily measured risk factors. Transitions were from 1 of 3 cognitive states (normal, delirious, or comatose) to 1 of 5 possible outcome states (normal, delirious, comatose, discharged from ICU, and died). Candidate predictors were selected according to 1) strength of evidence, 2) expert input, 3) availability in research database, and 4) availability of coded variables in the electronic medical record (for future validation using electronic record-based predictions). We assessed predictive value (e.g., positive and negative predictive value) and calibration by plotting empirical versus model-estimated probabilities for each of 5 possible outcomes (from any starting state) and 15 possible transitions. Internal validation was performed via a bootstrap validation procedure.
Results: We analyzed data from 810 patients for the development and internal validation. Delirium affected 606 (75%) patients. The final model included 14 individual risk factors including current cognitive status, age, history of cognitive impairment, baseline and daily severity of illness, and daily administration of sedatives medications. The positive and negative predictive values for cognitive states were as follows: normal (PPV = 0.66, NPV = 0.88), delirious (PPV = 0.55, NPV = 0.83), and comatose (PPV = 0.64, NPV = 0.90). Outcome-level calibration for normal, delirious, comatose, and ICU discharge outcomes were outstanding as evidence by a linear calibration curve estimates with slope and intercept of 1 and 0, respectively (Figure). Transition-level calibration was excellent for frequent transitions (e.g., coma to delirium), and poor for rare transitions (e.g., coma to discharge).
Conclusions: A dynamic transition model successfully predicts cognitive states for each ICU day while simultaneously accounting for ICU discharge and death. Future applications of the daily acute brain dysfunction prediction model should confirm its validity and assess its value in cognitive outcome surveillance.
