JMIR Publications

ABSTRACT

Background:

Blood tests used to identify patients at increased risk of undiagnosed cancer are commonly used in isolation. Some prediction models incorporate changes over repeated blood tests to improve individualised cancer risk identification.

Objective:

Our aim was to critically appraise existing prediction models incorporating blood test trends for risk of cancer.

Methods:

MEDLINE and EMBASE were searched until 15th May 2023. Screening was performed by four reviewers. Data extraction for each article was performed by two reviewers independently. We narratively synthesised studies and performed a random-effects meta-analysis of the c-statistic for trends-based prediction models.

Results:

We included 13 articles: five developed seven models and eight externally validated models only. In the seven models, full blood count trends were most commonly utilised (86%, n=7 models). Common cancers were colorectal (43%, n=3), gastro-intestinal (29%, n=2), non-small cell lung (14%, n=1), and pancreatic (14%, n=1). Four models (57%) used statistical regression and three (43%) used machine-learning. The number of blood test combinations ranged from 1 to 26. The ColonFlag model utilising trends in the full blood count was commonly externally validated, with a pooled c-statistic=0.81 (95% CI=0.77-0.85, n=4 studies) for six-month colorectal cancer risk. Models were often inadequately tested, with only one external validation study assessing calibration.

Conclusions:

Our review highlights that trends in individual blood tests are predictive for some undiagnosed cancers. Prediction models incorporating blood test trend were not found for most cancer sites, were rarely externally validated, and were not assessed in combination with symptoms.