Objective: To evaluate whether physiological trends and interaction terms could enhance predictive ability in a variant of the Richardson Score.
Method: We conducted a retrospective cohort study using data from 2,222 newborns of 34 weeks or more gestation with respiratory distress. The outcome variable was assisted ventilation longer 3 days or death. Using logistic regression and split validation we fit models for a variant of the Richardson score (gestational age, worst PaO(2)/FIO(2) ratio, lowest MAP, and a single interaction variable relating the lowest pH and the highest PaCO(2)) as well as models incorporating variables for trends and additional interaction terms. We assessed discrimination using the c-statistic.
Results: The 24-h Richardson score had a c-statistic of 0.83. In the validation dataset, adding pH trend significantly increased the c-statistic to 0.87. Adding PaO(2)/FIO(2) ratio trend increased the c-statistic to 0.86. Interactions with high significance were present in the data (e.g., adding all two-way interactions to our best trend model yielded a c-statistic of 0.92) but were unstably estimated with n = 2,222.
Conclusions: Incorporating trend and interaction terms in severity scores can enhance predictive ability. These modeling strategies have been underutilized in severity scoring, but in an era of increasing electronic availability of detailed clinical data the incorporation of trend and interaction effects in severity scoring will become both feasible and desirable.