mSOFA: Optimize SOFA for ICU Mortality

--- tags: SOFA --- # mSOFA: Optimize SOFA for ICU Mortality ## Abstract ### Objective Although consensus-driven SOFA score was reasonably accruate in predicting mortality, the cutoff points for each organ system were somewhat arbitrary. This study proposed mSOFA, which aimed to optimize SOFA score on ICU mortality prediction for sepsis-3. ### Data Sources This research included sepsis-3 patient cohort derived from MIMIC-III (Medical Information Mart for Intensive Care III) database. ### Study Selection A novel mSOFA score for mortality prediction is developed for sepsis-3 by applying logistic regression to the raw variable values that SOFA considers. The performance on mortality prediction of mSOFA was compared with SOFA. ### Data Extraction The patient cohort was derived from MIMIC-III database following sepsis-3 definition. 13 raw variable values related to SOFA, SOFA subscores, SOFA socres and ICU outcome are extracted to support further analysis. ### Data Synthesis A total of 20,936 sepsis-3 adult patients were identified and enrolled, including 2,236 non-survivors and 18,700 survivors. The overall ICU mortality rate was 10.68%. The results showed SOFA was accurate in predicted mortality (AUC: 0.7, 95% CI: 0.7-0.8). Compared to SOFA, mSOFA trained using raw variable values could make more accurate mortality prediction and achieve better discrimination (AUC: 0.8, 95% CI: 0). ### Conclusions This study developed time-series mSOFA for mortality prediction in patients with sepsis-3 by applying logistic regression on raw variable values. Although mSOFA had good discrimination and calibration, further external validation are required. ### Key Words SOFA, sepsis, mortality prediction, logistic regression ---- ## Introduction Sequential (Sepsis-related) Organ Failure Assessment (SOFA) score is one of the most commonly used acuity score for assessing the degree of organ dysfunction/failure and tracking patient status during ICU stay. SOFA is calculated using the worst measurements over a given interval (typically 24 hours) in conjunction with predefined thresholds to assgin a score for each organ system. The sum of these component scores yields the overall SOFA score, which can be used to assess patient acuity and predict mortality. Although SOFA is well correlated with mortality, it's accuracy is hindered by fixed cutoff points for each component score. Besides, we can not assertain patients with same SOFA score carry the same mortality risk. In this study, we try to solve this problem by applying logistic regression directly on SOFA-related variables. ## Mterials and Methods ### Cohort selection In total of 61,532 ICU admissions of 46,520 unique patients at Beth Israel Deaconess Medical Center (BIDMC) between 2001 and 2012 were obtained from the Medical Information Mart for Intensive Care (MIMIC-III, Version 1.4). Adult patients (age ≥ 18) who met the Third International Consensus Definitions for Sepsis (sepsis-3) are enrolled for further study, resulting in 20,936 ICU admissions, including 18,700 survivors and 2,236 non-survivors. **Table 1 Comparison of survivor and non-survivor patients.** | | Survivors (n=18,700) | Non-survivors (n=2,236) | P | | ---- | ---- | ---- | ---- | --- | | Male (%) | 0.56 | 0.55 | 0.17 | | Age (year) | 65.31 (53.73-78.31) | 70.71 (60.07-82.51) | < 0.01 | | Maximum SOFA score | 8.58 (6-11) | 13.27 (10-16) | < 0.01 | | Length of ICU stay (days) | 5.14 (1.53-5.63) | 7.60 (1.90-9.68) | < 0.01 | > Results are given as mean (interquartile range). P < 0.05 is accepted as statistically significant. Age of Patients whose age > 89 were treated as 100. ### Data cleaning Sepsis-3 was defined as life-threatening organ dysfunction caused by a dysregulated host response to infection. To calculate SOFA and identify sepsis-3 patients, time-series measurements for 13 variables (partial pressure of oxygen, fraction of inspired oxygen, mechanical ventilation status, platelets count, serum bilirubin, serum creatinine, urine output, Glasgow Coma Scale score, mean arterial pressure, dosing for dopamine, dobutamine, epinephrine and norepinephrine) used in the original SOFA score were extracted. Imputation of missing data as normal is the recommended and accepted methodology for dealing with missing data in prognostic scoring system. Hence, missing variables are imputed as normal. The SOFA score is calculated at admission and every hour until discharge, using the worst measurements during the prior 24 hours. Patients with a SOFA score increase at least 2 points consequent to the infection were identified as sepsis-3 patients. The primary outcome of this study is all-cause ICU mortality. Patients who survived to ICU discharge were labeled with 0, otherwise 1. ### Development of mSOFA The accuracy of SOFA score in predicting mortality is hindered since fixed cutoff points are somewhat arbitrary. To address the challenge, logistic regression models were fitted on raw variable values. With a moving window of width $m$, the logistic regression model was fitted on raw variables for $k$-th window with time from $km$ to $km+24$. To train this logistic regression model, train:validation:test=8:1:1 By doing so, time-series coefficients and constants were obtained. ## Results ### mSOFA vs. SOFA in mortality prediction Across all time windows, AUC of mSOFA are higher than original SOFA. **Figure 1: Discrimination of mSOFA compared with origin SOFA.** Compare AUC. **Figure 2: Comparison of SOFA and mSOFA predictions against mortality rate.** Compare prediction probabilities against real mortality rate. ### Case studies - Find 2 patients with the same SOFA score, but different subscores and outcomes. In this case, SOFA fails while mSOFA works well ## Critical transition Coefficients and constants across time windows are changing. ## Discussion By applying logistic regression on raw variables, mSOFA returns well calibrated predictions as it directly optimizes log-loss. Although calibration may not be needed, further external validation of mSOFA are required. ## Conclusions This study developed a novel score called mSOFA which yields more accurate mortality prediction compared to original SOFA. Although SOFA score is well correlated to mortality, mSOFA can better discriminate patients with same SOFA score but different risk of mortality. ## Acknowledgements Thank God, thank you, and thank myself. ---- ## References **Pay attention to the styles** - SOFA definition paper Vicent - Sepsis3 definition paper - MIMIC-III paper - Use of SOFA for outcome - DeepSOFA - LASSO score - LODS score - Logistic EuroSCORE - SOFA evaluation ---- ### Figures ### Tables ### Supplementary 1. List of sepsis3 patients 2. Comparison of SOFA and mSOFA 3. Time-series coefficients and constants 4. Links for code ---- https://journals.lww.com/ccmjournal/Documents/CCM_Info_for_Auths.pdf