Learning Normal Representations for Blood Biomarkers.

Area of Science:

• Laboratory Medicine
• Biomarker Analysis
• Computational Biology

Background:

• Current blood biomarker interpretation relies on population reference intervals, ignoring individual variability.
• This can mask significant changes from a patient's baseline, potentially delaying disease detection.
• Existing personalized approaches may overfit sparse data, leading to false positives and unnecessary follow-up.

Purpose of the Study:

• To evaluate the performance of purely personalized biomarker interpretation.
• To develop and validate a novel framework (NORMA) for individualized laboratory interpretation.
• To improve the precision of predicting adverse clinical outcomes using blood biomarkers.

Main Methods:

• Analysis of nearly 2 billion longitudinal laboratory measurements from over 1.6 million individuals.
• Development of NORMA, a conditional transformer-based framework integrating patient history and population data.
• Validation of NORMA-derived intervals against clinical outcomes like mortality and acute kidney injury.

Main Results:

• Purely personalized intervals frequently overfit, labeling up to 68% of measurements as abnormal without clinical correlation.
• NORMA demonstrated higher precision in predicting adverse outcomes compared to purely personalized or population-based intervals.
• NORMA effectively balances individual patient data with population-level normal variation.

Conclusions:

• Over-personalization in laboratory medicine can lead to misinterpretation and unnecessary clinical actions.
• Anchoring individual biomarker trajectories to population priors, as achieved by NORMA, offers superior predictive accuracy.
• The NORMA framework and associated tools are publicly released to promote accessible, individualized laboratory interpretation.