Logistic regression model to predict outcome after in-hospital cardiac arrest: validation, accuracy, sensitivity and specificity

• Published in: Resuscitation 1998-03, Vol.36 (3), p.201-208
• Main Authors: Dodek, Peter M, Wiggs, Barry R
• Format: Article
• Language: English
• Subjects: Anesthesia. Intensive care medicine. Transfusions. Cell therapy and gene therapy; Biological and medical sciences; Cardiac arrest; Confidence Intervals; Emergency and intensive cardiocirculatory care. Cardiogenic shock. Coronary intensive care; Female; Heart Arrest - mortality; Hospital Mortality; Humans; Inpatient; Intensive care medicine; Logistic Models; Logistic regression; Male; Medical sciences; Outcome; Outcome Assessment (Health Care); Predictors; Retrospective Studies; Sensitivity and Specificity
• ISSN: 0300-9572; 1873-1570
• DOI: 10.1016/S0300-9572(98)00012-4

Objective: To develop and validate a logistic regression model to identify predictors of death before hospital discharge after in-hospital cardiac arrest. Design: Retrospective derivation and validation cohorts over two 1 year periods. Data from all in-hospital cardiac arrests in 1986–87 were used to derive a logistic regression model in which the estimated probability of death before hospital discharge was a function of patient and arrest descriptors, major underlying diagnosis, initial cardiac rhythm, and time of year. This model was validated in a separate data set from 1989–90 in the same hospital. Calculated for each case was 95% confidence limits (C.L.) about the estimated probability of death. In addition, accuracy, sensitivity, and specificity of estimated probability of death and lower 95% C.L. of the estimated probability of death in the derivation and validation data sets were calculated. Setting: 560-bed university teaching hospital. Patients: The derivation data set described 270 cardiac arrests in 197 inpatients. The validation data set described 158 cardiac arrests in 120 inpatients. Interventions: none. Measurements and results: Death before hospital discharge was the main outcome measure. Age, female gender, number of previous cardiac arrests, and electrical–mechanical dissociation were significant variables associated with a higher probability of death. Underlying coronary artery disease or valvular heart disease, ventricular tachycardia, and cardiac arrest during the period July–September were significant variables associated with a lower probability of death. Optimal sensitivity and specificity in the validation set were achieved at a cut-off probability of 0.85. Conclusions: Performance of this logistic regression model depends on the cut-off probability chosen to discriminate between predicted survival and predicted death and on whether the estimated probability or the lower 95% C.L. of the estimated probability is used. This model may inform the development of clinical practice guidelines for patients who are at risk of or who experience in-hospital cardiac arrest.