Biostatistical Models for Predicting Mortality in Intensive Care Units

Physician Perspectives on ICU Mortality Prediction Models

Abstract

Intensive care Units clinicians often rely on risk scores. However, the use of a prediction model depends on how it is built, updated, explained and introduced into clinical work. The researchers surveyed 223 physicians working in intensive care settings in the West Bank and asked about eight features of biostatistical models used for mortality prediction: input variables, model type/methodology, calibration and discrimination, temporal resolution, handling of missing data, interpretability, external validation and clinical integration. Accordingly, the responses were analyzed in SPSS using descriptive statistics, Cronbach's alpha, collinearity diagnostics and multiple linear regression. All the eight dimensions showed positive and statistically significant associations with the physician-reported mortality-prediction outcome. The model type/methodology had the largest standardized coefficient (β = 0.288; B = 0.256; p < 0.001) and followed by the temporal resolution (β = 0.229; B = 0.202; p < 0.001). The full model yielded R = 0.905 and R² = 0.819. The findings showed how the physicians viewed the usefulness of the model characteristics; they did not represent the validation of an ICU mortality model using patient-level outcomes.