OBJECTIVES:

Inspired by the well-known four-phase hierarchy for the clinical evaluation of new pharmaceuticals, several frameworks were proposed for clinical prediction model research. However, these frameworks focused on the steps of model development and evaluation process rather than how to define phases to distinguish the maturity of the models. We aim to construct a conceptual framework with phases of clinical prediction model research, to define the criteria of each phase, and to describe the paths from one phase to the next phase.

METHODS:

Existing frameworks in prediction model research were reviewed, compared and criticized. A new phases definition is adapted from biomarker development and diagnostic test research, allowing their similarity to prediction model research.

RESULTS:

An architecture with 5 consecutive phases is proposed for clinical prediction model research, which contains: Phase 1 Predictors exploratory, Phase 2 Model derivation, Phase 3 Model validation, Phase 4 Evaluation of model impact on patient outcome, Phase 5 Implementation and effects on population. Phase 2 is further divided into two sub-phases based on the study design, and Phase 3 is further divided into two sub-phases based on data source (narrow vs broad). Each (sub-)phase is elaborated with example studies, and the proposed phases and the paths from a lower phase to an advanced phase are presented as a board game, to illustrate actions needed in each phase.

CONCLUSIONS:

The proposed framework can help researchers to design a road map for their prediction model development projects, and help model developers and model users better understand a model belongs to which specific phase. It will ensure the success of novel prediction model development, from concept to implementation.