Background: Assessing severity of illness using available electronic medical record (EMR) data on admission and predicting inpatient mortality is very challenging. Lacking standardized practices around end of life issues, hospitalists use their clinical judgment in making these crucial decisions. Prolonged discussions may be needed in medically complex patients to direct an optimal plan of care resulting in Intensive Care Unit (ICU) escalations, increased length of stay and can negatively impact patient satisfaction. Thus, there is a critical need to assist hospitalists in accurately identifying patients at high risk for mortality and guide patient-physician discussions around end of life issues. Previously reported mortality risk stratification models are either disease specific, limited to ICU patients or predict post-discharge mortality. In this project we used administrative data available within the first 24 hours of transfer to predict 30-day inpatient mortality at Indiana University (IU) Adult Academic Health Center (AHC). This is the first known study to include all disease categories and all levels of inpatient care as our focus is to assist hospitalists who provide general medical care to a diverse inpatient population.

Methods: A total of 10,389 patients (18 years and older) transferred to IU-Adult AHC between 1/1/2016 and 12/31/2017 from other facilities were included. Administrative data was collected from the IU Health data warehouse and from the Indiana Health Information Exchange (IHIE), a resource that stores clinical and diagnostic data for patients across Indiana. 31 putative variables were selected based on a PubMed literature search using terms “early In-hospital mortality, Inter-hospital transfers” and by consideration of previously reported mortality prediction tools. We used a multiple logistic regression model in the full data set to identify 21 out of the 31 variables that were associated with 30-day inpatient mortality(p<0.05). The cohort was then randomly divided into derivation (n = 5194) and validation (n = 5195) samples. The rate of 30-day inpatient death was equivalent for the derivation sample and validation sample (9.09%). These 21 variables were tested for 30-day inpatient death in the derivation sample (Table 1). Area under curve -receiver operating characteristics (AUC-ROC) analysis was used to identify optimal risk threshold score with the highest sensitivity and specificity in the derivation sample. These thresholds were then applied to the validation sample to verify the accuracy of the mortality predictive ability of the model (Table 2).

Results: TThe final model was strongly discriminative (C statistic = .91) and had a good fit (Hosmer‐Lemeshow goodness‐of‐fit test, p = .14). The positive predictive value for 30-day inpatient death was 64.79% and the AUC-ROC was 0.91; (95% confidence interval 0.89 – 0.92, p <.0001). An outcome score of -2.32 maximized sensitivity (82.84%) and specificity (84.05%) in derivation sample. Applying the same threshold (-2.32) to the validation sample resulted in sensitivity of 75.42% and specificity of 84.01%. At values above this threshold in the validation sample, 32.04% (356/1111) of patients died within 30 days; below this threshold, only 2.84% (116/4084) of patients died.

Conclusions: This novel model can accurately predict 30-day inpatient mortality using readily available EMR data within the first 24 hours of a hospital transfer and facilitate complex goals of care discussions with patients and families.

IMAGE 1: Table1: Summary of Multivariate Analysis Predicting 30-day Inpatient Death in Derivation Sample (n=5194)

IMAGE 2: Table 2: Mortality Predictive Model Performance by Specificity Levels (95%, 90%, 85%, 80%, 75%):