Comment on: Permanent stoma rates after anterior resection for rectal cancer: risk prediction scoring using preoperative variables

Alison Bradley, Stephen Knight, Mei M Chin, Susan Moug

Research output: Contribution to journalLetterpeer-review

Abstract / Description of output

Dear Editor

We read with interest the article entitled ‘Permanent stoma rates after anterior resection for rectal cancer: risk prediction scoring using preoperative variables’ by Back et al.1 This is an informative example of how preoperative data can be utilized to provide a personalised predictive model to support shared decision-making. We congratulate the authors for their important and timely work. However, if such a model is to provide clinical utility, we believe there are outstanding methodological questions that should be addressed.

The absence of reporting according to the TRIPOD statement limits transparency. The use of complete case analysis can result in loss of statistical power and introduce selection bias into the study population as missing data rarely occurs randomly but often pertains to disease or participant characteristics. Multiple imputation of missing data is instead advocated as a means of reducing bias whilst maintaining statistical power, which appears possible with the SuperLearner methodology. Furthermore, candidate predictors were not blinded to outcome and were limited to predictors included in the Swedish Colorectal Registry.

Finally, the reported area under the receiving operating characteristic curve (AUROC) was only 0.67. Impact analysis and classification measures (sensitivity, specificity, predictive value) are lacking. Reliance on AUROC alone to assess model performance is insufficient to determine real-world performance. The absence of calibration specifically limits conclusions on the ability of this novel risk model to support shared decision-making, as poorly calibrated models can lead to false expectations.

Overall this paper marks an important step towards personalised predictive medicine and, with some fine tuning, could become an invaluable tool in supporting shared decision-making. The issues highlighted are not unique to this study. If personalised predictive modeling supporting shared decision-making is to become a clinical reality, a wider exploration of their methodological quality and clinical impact is warranted.
Original languageEnglish
JournalBJS Open
Volume109
Issue number2
DOIs
Publication statusPublished - 9 Nov 2021

Fingerprint

Dive into the research topics of 'Comment on: Permanent stoma rates after anterior resection for rectal cancer: risk prediction scoring using preoperative variables'. Together they form a unique fingerprint.

Cite this