The Inclusion and Intended Use of Prediction Models in Clinical Guidelines: A systematic review of five clinical domains in four countries

Objectives Prediction models can be used to support medical decision-making, but their integration into clinical guidelines remains unclear. Despite the abundance of published models, it is unknown how many are included in clinical guidelines and whether their recommended use in the guideline aligns with model's intended use at development. We investigated the inclusion of prediction models in clinical guidelines and examined any discrepancies between the intended use described in the model development or validation studies and their recommendation in the clinical guidelines in which they are mentioned. Study Design and Setting We systematically reviewed clinical guidelines across five clinical domains (pulmonary embolism, preeclampsia, sepsis, dementia, and lung cancer) in four countries (the United States, the United Kingdom, Canada, and the Netherlands). For each prediction model included in the guideline, we identified the corresponding model development and validation studies. Data on clinical setting and intended use of the prediction model were extracted from the guidelines, the development and external validation papers. Results A total of 20 clinical guidelines were included, identifying nine unique prediction models that were externally validated in 59 studies. Pulmonary embolism had the highest frequency of prediction models in guidelines (4/4), followed by sepsis (3/4), lung cancer and preeclampsia (both 1/4), and dementia (0/4). We identified large discrepancies between the setting and intended use of the models at development or in validation studies, and the guidance provided in guidelines, particularly on aspects such as the clinical context/setting, the timing of use, and the specific patient population. Conclusion Clinical guidelines occasionally recommended the use of prediction models. Many guidelines provided limited instructions on how to operationalize these models and often recommended using the model in a specific clinical scenario without addressing any further key differences from development or validation. When included, typically there were inconsistencies between the intended use of the model and guideline recommendation. There is need for greater standardization in how prediction models are incorporated into guidelines, and to ensure that their recommended use aligns with their intended purpose and validation evidence.