Abstract
Simple blood tests are a cornerstone of primary care, yet their utility in cancer detection is often limited by a reliance on single-point "snapshot" measurements. Current clinical guidelines typically trigger investigations only when a result crosses a population-wide threshold. However, recent evidence suggests that longitudinal trends—how a patient's results change over time relative to their own baseline—could be far more predictive. The BLOod Test Trend for cancEr Detection (BLOTTED) study is an observational and prediction model development study designed to explore this hypothesis. By leveraging English primary care electronic health records, the study aims to develop clinical prediction models that incorporate simultaneous trends in multiple blood tests to identify cancer risk earlier and more accurately.
The Problem: The Limitation of Snapshot Testing
In contemporary primary care, a "normal" blood test result can be misleading. GPs currently look for results that fall outside a specific reference range. However, a patient may experience a significant, steady decline in a marker like hemoglobin that remains within the "normal" range but represents a major departure from their personal historical baseline.
Existing evidence for cancer detection is almost entirely based on these isolated snapshots. While certain abnormalities, such as raised platelets or low albumin, are known red flags, they often lack the specificity to pinpoint a particular cancer type or they appear too late in the disease progression. The BLOTTED study addresses whether the trajectory of a patient's blood markers over several years is a more sensitive predictor of malignancy than any single measurement, potentially catching cancer during the critical window when it is most treatable.
Methodology and Data Integration
As a protocol for a large-scale observational study, BLOTTED outlines a sophisticated three-aim statistical approach using the Clinical Practice Research Datalink (CPRD) Aurum database. This dataset includes records from approximately 40 million patients in English primary care, linked to several national registries to ensure robust outcomes:
- NCRAS: The National Cancer Registration and Analysis Service for definitive cancer diagnoses.
- HES: Hospital Episode Statistics for secondary care data.
- ONS: Office for National Statistics for mortality data.
Analytical Framework
- Descriptive Epidemiology (Aim 1): Researchers will use LOWESS (Locally Weighted Scatterplot Smoothing) to map how blood markers naturally fluctuate across different demographics and comorbidities, establishing a foundation for what constitutes "normal" variation over time.
- Association Modeling (Aim 2): The team will employ Joint Models, which simultaneously analyze longitudinal data (the blood test trends) and time-to-event data (the cancer diagnosis). This allows for a direct correlation between a marker's trajectory and the risk of developing cancer. Mixed-effects models will be used to account for variation between individual patients.
- Predictive Modeling (Aim 3): The study will develop multivariate models, benchmarking traditional statistical methods against Machine Learning approaches. Neural networks will be utilized for supervised clustering of trajectories to determine if specific "shapes" of data curves can predict specific cancer types.
To manage the complexities of real-world medical records, the protocol even proposes using Generative Adversarial Networks (GANs) for missing value imputation—a cutting-edge method for handling the sporadic intervals at which patients are typically tested.
Key Insights: The Personal Baseline
The core innovation of the BLOTTED study is the shift from population-wide thresholds to individual trajectories. By identifying a "personal baseline," the model can detect deviations that would otherwise be ignored. Furthermore, the study investigates Multivariate Synergy—how simultaneous changes in different markers (e.g., a decline in hemoglobin paired with a rise in platelets) might provide a unique "fingerprint" for specific malignancies, such as colorectal cancer.
Preliminary work justifying this protocol found that such simultaneous shifts were significant predictors of a colorectal cancer diagnosis within two years, even when all individual results remained within the standard reference ranges.
Challenges and Future Implementation
The researchers acknowledge several significant hurdles, including the massive computational burden of analyzing longitudinal data for millions of individuals. There is also the challenge of "clinical indication bias," where tests are more likely to be ordered for patients already suspected of being ill.
Ultimately, the success of the BLOTTED project will be measured by its clinical utility. The goal is to integrate these predictive models into GP software to provide real-time decision support. By identifying high-risk trends before symptoms become overt, the study aims to increase cancer yield, reduce unnecessary referrals, and provide more patients with the opportunity for early intervention and improved long-term outcomes.
Summary updated