In most databases used to build general severity scores the median duration of intensive care unit (ICU) stay is less than 3 days. Consequently, these scores are not the most appropriate tools for measuring prognosis in studies dealing with ICU patients hospitalized for more than 72 h.
Purpose: To develop a new prognostic model based on a general severity score (SAPS II), an organ dysfunction score (LOD) and evolution of both scores during the first 3 days of ICU stay.
Design: Prospective multicenter study.
Setting: Twenty-eight intensive care units (ICUs) in France.
Patients: A training data-set was created with four ICUs during an 18-month period (893 patients). Seventy percent of the patients were medical (628) aged 66 years. The median SAPS II was 38. The ICU and hospital mortality rates were 22.7% and 30%, respectively. Forty-seven percent (420 patients) were transferred from hospital wards. In this population, the calibration (Hosmer-Lemeshow chi-square: 37.4, P = 0.001) and the discrimination [area under the ROC curves: 0.744 (95 % CI: 0.714-0.773)] of the original SAPS II were relatively poor. A validation data set was created with a random panel of 24 French ICUs during March 1999 (312 patients).
Measurements and main results: The LOD and SAPS II scores were calculated during the first (SAPS1, LOD1), second (SAPS2, LOD2), and third (SAPS3, LOD3) calendar days. The LOD and SAPS scores alterations were assigned the value "1" when scores increased with time and "0" otherwise. A multivariable logistic regression model was used to select variables measured during the first three calendar days, and independently associated with death. Selected variables were: SAPS II at admission [OR: 1.04 (95 % CI: 1.027-1.053) per point], LOD [OR: 1.16 (95 % CI: 1.085-1.253) per point], transfer from ward [OR: 1.74 (95 % CI: 1.25-2.42)], as well as SAPS3-SAPS2 alterations [OR: 1.516 (95 % CI: 1.04-2.22)], and LOD3-LOD2 alterations [OR: 2.00 (95 % CI: 1.29-3.11)]. The final model has good calibration and discrimination properties in the training data set [area under the ROC curve: 0.794 (95 % CI: 0.766-0.820), Hosmer-Lemeshow C statistic: 5.56, P = 0.7]. In the validation data set, the model maintained good accuracy [area under the ROC curve: 0.826 (95 % CI: 0.780-0.867), Hosmer-Lemeshow C statistic: 7.14, P = 0.5].
Conclusions: The new model using SAPS II and LOD and their evolution during the first calendar days has good discrimination and calibration properties. We propose its use for benchmarking and evaluating the over-risk of death associated with ICU-acquired nosocomial infections.