7. Developing high-performance prediction models for medical outcomes.

Abstract

Abstract: A prediction model can be considered to have high performance when it gives correct prediction in at least 90% cases for qualitative outcome and at least 90% differences between the observed and predicted values are within clinical tolerance in the case of quantitative outcome. A model with an accuracy between 80% and 90% can be possibly tolerated in some situations, but any model with less accuracy implies an unacceptably large error in clinical applications. Most of the models developed so far do not meet these criteria. Moreover, prediction of the unknown is confused with classification of the known. Developing high-performance models requires a lot of extra efforts that are rarely seen in the present endeavors. For this, it is imperative that a large number of known and suspected predictors are considered and complexity in terms of interactions and nonlinearity is accepted. Models with many predictors is not a big limitation now because of wide availability of enormous computing power. More rigorous validation is needed than done now. Artificial intelligence-based models can be linked to the relevant literature so that they continuously update with the new development under the data science paradigm. This article highlights the problems with many of the existing models and describes steps to develop a high-performance model for medical outcomes. Many of the advices we give are rarely available in the literature.