Abstract ID#: 3147 Category: Health Sciences Undergraduate/Graduate: PhD CHF-AR-HMM and Analysis for Post-ICU Mortality Prediction Yilin Yin, Chun-An Chou Department of Mechanical & Industrial Engineering, Northeastern University

Overlook Modeling: Conduct Autoregressive hidden Markov model (AR-HMM) The (ICU) mortality prediction q The post-ICU dynamic risk (mortality) prediction is on demand by patients for timely decision making and consecutive care q Our goal is to predict the mortality within certain days (e.g., 5-10 days since admission) with the last 24-hour physiology data by improving the existing scoring system

State-of-the-art Scoring System q Risk scores converted from irregular time-series body vitals and laboratory test data are delivered for predicting mortality q In this case, Simplified Acute Physiology Score (SAPS) II [1] designed for mortality prediction is considered Fig 3. Methodology workflow and CHF-AR-HMM mixture model.

Challenges q Risk scoring systems are not capable of capturing the patient’s severity level in a dynamic manner q Current cumulative hazard Markov model method relies on the prior probability, which is highly biased

Methodology q We use the mixture model CHF-AR_HMM as combination of autoregressive Hidden Markov model (AR-HMM) [2] with cumulative hazard function (CHF) [3] to discover the latent states (survival or death) dynamic on the aggregated data from the risk scoring system for each patient

Data Processing q PhysioNet ICU Challenge 2012 MIMIC II [4] is used for model validation, including first 48-hour ICU (12 physiology variables and other patient attributes), laboratory tests data since ICU admission from 12000 patients, the demographic and mortality outcome information. q 1)Discretize the raw ICU data into SAPS I Score; 2) Split data into time intervals Fig 4. General approach and CHF-AR-HMM mixture model and 8-hour time interval example for joint probability classifier.

Result q Mortality prediction (20 rounds 5-fold cross validation, under-sampling)

Fig 1. A sample showing three variables (HR, SysABP, Temp) of the original data. The shadowed area is the first 8-hour time interval within the last 24-hour data. The boxplot and violin plot show the distribution of aggregated SAPS II scores for all samples in this time interval. In hospital death = 1 is survival; 0 is death.

Fig 5. The 8-hour time interval has the best performance. The probability of survival (Red is death; blue is survival). q From the output, the CHF AR-HMM model improves the mortality prediction performance of the original SAPS II scoring system as the original SAPS II output with sensitivity from 10% to 30% level and specificity from 80% to 90% level. The CHF AR-HMM model also responds to the dynamic regarding the number of days of discharge from the ICU. The survival probability for both death and survival cohorts decreases by the cut-off date (day 5 to day 10 since admission).

Reference Fig 2. The density plot of deceased samples. The red line is the exponential density line fitted by λtype1. Each time intervals training subset is assigned the λ parameter as the number of death in unit survival time. λ is the hazard function for given 1. Le Gall et al, “A simplified acute physiology score for ICU patients,”,1984 Critical Care 2. Stanculescu, Ioan et al., “Autoregressive Hidden Markov Models for the Early Detection of Neonatal ,” IEEE J. Biomed. Health Inf., vol.18(5), pp.1560-1570, 2014. patients in each subset. The transition probability from survival to death is the cumulative probability function. λtype2 is 3. DH Lee, E Horvitz , “Predicting mortality of intensive care patients via learning about hazard,”, 2017, Thirty-First AAAI Conference on Artificial Intelligence reversely calculated from the cumulative probability of survival, which is converging to 0 when the patient is dead. 4. http://physionet.org/challenge/2012/