BACKGROUND

There is still demand for algorithms that can be used at the point of care, especially when dealing with events that do not present with a single obvious clinical indicator. The Spiegelhalter-Knill-Jones (SKJ) method is an approach for the development of a clinical score that focuses on the effect size of predictors, which is more relevant in settings where events may be rare or data is scarce. However, it does require predictors to be binary or dichotomised.

METHODS

We developed an extension of the Spiegelhalter-Knill-Jones method that can include continuous variables and added additional features that make it more useful in a variety of settings. We illustrated our method on two historical datasets dealing with viral failure in HIV patients in Cambodia. We used area under the curve (AUC) and risk classification improvement (RCI) as metrics to evaluate the performance of resulting predictions scores and risk classifications.

RESULTS

All new features worked as intended. Scoring systems developed with the new method outperformed an earlier application of a classic version of SKJ method on the training dataset, while no significant difference was found on any of the performance measures in the test dataset.

CONCLUSIONS

This extension provides a useful tool for clinical decision-making that is much more flexible than the original version of SKJ, and can be applied in a variety of settings.