A new tool for screening severe sleep apnea syndrome

To design and validate a new “score” (the Clinique Pasteur score) to detect severe OSAS in patients with unexplained nocturnal polyuria.

To validate this “score” (internal validation)

Development of the score
Step 1: logistic regression model
The first step in constructing the score was to perform a multivariate logistic regression analysis. All variables associated with severe OSAS according to a p value <0.2 in the logistic regression were selected. For continuous variables, the log-linearity assumption had to be fulfilled to ensure the validity of the model. If this assumption was not fulfilled, the continuous variable was categorized. The continuous variables included in the final model were age and BMI, and the assumption of log-linearity was not fulfilled. The thresholds chosen for the categorization of BMI were 25 kg/m2 and 30 kg/m2 according to anthropometric definition (normal/overweight/obese). The threshold for age (70 years) was much more arbitrary and was choose for sample size and powerful of the prediction reasons. The results of the model are expressed by means of odd-ratio together with their 95% confidence intervals computed by Wald’s methods.
The performance was assessed by the rate of prediction error, the receiver operating characteristic (ROC) curve and a graphical illustration of the specificity/sensitivity of the model. The area under this curve (AUC) and its 95% confidence interval computed by bootstrap procedure (2000 replicates) indicated the predictive performances of the model.
The internal validity of the model was investigated by splitting the database into two cohorts: a learning cohort (65% of the sample size) to create the model and a validation cohort (35% of the sample size) to assess. Individuals were randomly assigned to one of the cohorts. The predictive performance was assessed by the rate of prediction error and ROC curves.
Step 2: construction of the score
As the variables involved in the model were discrete, it was possible to construct a simplified score by choosing intergern proportional values to rounded values of the logistic regression coefficients. This simplified score was constructed by using the connection between the values of the score and the predicted probabilities for a patient presenting with unexplained NP to have a severe OSAS. The choice of the cut-off score was made according to the plot of PPV and NPV as a function of the threshold.