View ORCID ProfileLaura J Bonnett tenure track fellow,
Kym I E Snell research fellow in biostatistics,
Gary S Collins professor of medical statistics,
Richard D Riley professor of biostatistics
Bonnett L J, Snell K I E, Collins G S, Riley R D.
Guide to presenting clinical prediction models for use in clinical settings
BMJ 2019; 365 :l737
doi:10.1136/bmj.l737
Re: Guide to presenting clinical prediction models for use in clinical settings
In their recent article, Bonnett et al. describe four ways to present clinical prediction models, namely a points score system, graphical score chart, nomogram and digital presentation using a website and/or mobile app. We would like to nuance their description of a points score system. We disagree with the statement that continuous predictors need to be categorised before regression.
Although it increases the ease of interpretation when regression coefficients are rounded (and scaled), unfortunately some predictive accuracy is sacrificed. Indeed, when predictors are kept on their original continuous scale and the regression coefficient is multiplied by the value of the respective predictor, an individual score can be obtained (although this might not be a rounded number). Next, after calculating the individual score for each patient, a range of scores with their corresponding range of predicted probabilities can be depicted, potentially rounded at this stage to ease interpretation. Granted, not all predicted probabilities corresponding to all point scores can be assessed, but when evaluating separate risk groups, for instance, this is often not necessary. This approach is similar to prediction model presentation using a nomogram.
However, it does not aspire to give exact outcome probabilities. Such a presentation could enhance interpretation (without loss of precision and without needing multiple categories in a regression analysis which could be problematic in smaller datasets). Furthermore, the linearity assumption can be addressed using transformations in the modelling steps upfront without creating unequal categories.
An example of such a model we have previously published involved prediction of failure after focal treatment for prostate cancer recurrences (see table 2 and 3 in the paper)1.
1. Development and internal validation of prediction models for biochemical failure and composite failure after focal salvage high intensity focused ultrasound for local radiorecurrent prostate cancer: Presentation of risk scores for individual patient prognoses. Urol Oncol. 2018 Jan;36(1):13.e1-13.
Competing interests: No competing interests