Estonian Biobank Study Maps Diagnoses by Shared Laboratory Fingerprints
Routine laboratory results from electronic health records can reveal how diseases cluster in real life, beyond the boundaries of billing codes and specialty labels. In a large Estonian Biobank analysis, investigators used linked EHR laboratory data from 211,994 participants spanning about 20 years to reconstruct disease groupings from the patterns embedded in common blood and chemistry tests.
The work centered on a familiar but stubborn problem in clinical informatics: laboratory data are plentiful, yet fragmented. Different units, naming conventions, and missingness patterns make it difficult to compare measurements across large cohorts. Here, the team harmonized and pre-processed the available laboratory values and then used them to train LightGBM models for each main ICD-10 code, with age, sex, and BMI included as covariates. From those models, SHAP feature-importance profiles were generated for each diagnosis code, turning a conventional EHR archive into a map of clinical signatures.
The result was not a generic machine-learning clustering exercise. It was a biologically grounded attempt to see whether diagnoses with similar laboratory behavior would group together in ways that make physiologic sense at the bedside.
Shared Lab Patterns Reassembled Disease Structure
Across the 1,995 ICD-10 codes examined, hierarchical clustering produced 33 clusters that remained stable across all three resolutions tested, while self-organizing maps produced 27 clusters. Eight clusters were robust across both clustering methods and all resolutions. In a field where many algorithmic groupings collapse when the input changes slightly, that kind of cross-method stability is the first sign that the signal is not purely statistical noise.
The clustering recovered established disease biomarkers as positive controls and then went further by reassembling multi-organ disease structure. One example was a set of diabetic complications linked through a shared glucose signature. Clinically, that is exactly the sort of pattern one would expect if a common metabolic disturbance is driving several downstream diagnoses rather than each code representing a biologically isolated entity.
A glucose-centered cluster also makes physiologic sense because hyperglycemia leaves a broad laboratory footprint. It affects glycation, renal handling, inflammatory pathways, and vascular injury, all of which can appear across different clinical endpoints while still reflecting a single upstream metabolic process.
Ascertainment Bias Was Treated as a Biological Confounder, Not an Afterthought
The sharpest contribution of the analysis is not simply that diagnoses clustered together. It is that the investigators explicitly tested whether a laboratory difference reflected disease biology or differential testing behavior. Using Cohen’s d analysis, they assessed 46 ICD-10 code clusters and flagged measurements in which the case-control difference in testing frequency suggested ascertainment bias rather than a true difference in measured value.
This matters because EHR laboratory data are never neutral. Sicker patients are tested more often, follow-up testing is ordered more aggressively in certain conditions, and some diseases trigger broader panels than others. A code can appear to have a strong laboratory signature simply because clinicians order that assay more often in that population. By separating clusters into 19 value-driven groups and clusters reflecting ascertainment bias, the analysis drew a line between reproducible biology and the downstream habits of care delivery.
One sentence captures the practical consequence: a laboratory signature is only useful if it survives the testing behavior that produced it.
Nonlinear Models Captured Structure Conventional Summaries Miss
The use of gradient-boosted decision trees and SHAP feature importance allowed the model to capture nonlinear relationships among labs, demographics, and diagnosis codes. That matters because disease signatures in real-world EHRs are rarely linear. Renal dysfunction may shift several analytes at once, inflammatory states can distort multiple markers in parallel, and endocrine disorders can alter routine chemistry in ways that are highly context dependent.
By clustering diagnosis codes on the basis of those feature-importance profiles, the study moved beyond simple pairwise correlations. Instead of asking whether one lab differed by disease status, the approach asked which combinations of laboratory behavior best characterized each diagnosis code. In practice, this creates a more flexible taxonomy of disease than standard code-based grouping, one that is closer to the way clinicians reason through overlapping syndromes, complications, and comorbidities.
The scale of the cohort strengthens the case for the approach. With nearly 212,000 participants and roughly two decades of linked data, the analysis had enough breadth to detect recurrent patterns across common diagnoses, while still confronting the unevenness that defines routine care data.
Why This Matters for Biobanks and Clinical Informatics
Biobank-linked EHR data are often described as a rich resource, but richness alone is not enough. The challenge is converting disparate observations into structured clinical meaning without washing out the signal that makes the data valuable. This study shows one path forward: use routine laboratory measurements to infer disease signatures, then validate those signatures against known biology and against the quirks of healthcare utilization.
The findings also suggest a future in which ICD-10 codes are not treated as the final word on disease identity. A code can be a convenient administrative label while still sitting inside a broader biological family. In the setting of metabolic disease, multi-organ complications, and common laboratory abnormalities, a signature-based regrouping may be more informative than rigid code buckets. It could also help researchers build cleaner phenotype definitions for downstream association studies, especially when the same diagnosis is documented across multiple specialties and care settings.
The preprint stops short of claiming clinical deployment, and it should. A laboratory-signature map built from one national biobank still needs external validation in other health systems, other laboratory platforms, and more diverse populations. Missingness, local ordering practice, and code assignment behavior can all reshape the clusters. Even so, the analytic framework is strong enough to change how EHR data are viewed: not as a messy archive to be cleaned into compliance, but as a record of biology filtered through the realities of medical care.
The remaining question is not whether EHR laboratory data contain disease information. They do. The harder task is deciding how much of the signal belongs to pathophysiology, how much belongs to testing behavior, and how much survives once the same method is applied outside Estonia.