Technical Summary: Real-world drug use in ATC and ICD-10

Problem Statement

The increasing utilization of Electronic Health Records (EHRs) for drug repositioning and large-scale biomedical research has highlighted a critical gap: the lack of a well-curated resource describing real-world drug-diagnosis relationships. Existing resources, such as PrimeKG or ClinGraph, often rely on mechanistic disease ontologies (e.g., HPO, MONDO) or literature-based indications (e.g., MEDI, DrugBank). These ontologies do not align well with the International Classification of Diseases (ICD) and Anatomical Therapeutic Chemical (ATC) systems, which remain the standard for clinical documentation, billing, and administrative coding in most healthcare settings. Furthermore, statistical co-occurrence of drugs and diagnoses in EHRs does not necessarily imply direct therapeutic intent; associations may arise from treating comorbidities, prophylaxis, symptom management, or side effects. There is a need for a data-driven resource that maps ATC drug codes to ICD diagnosis codes, distinguishing between direct treatment and indirect associations, while accounting for heterogeneity across different healthcare systems (primary vs. secondary care) and geographic regions.

Methodology

The study generated a data-driven mapping of drugs to disease diagnoses using real-world data from two distinct healthcare systems:

1. Data Sources:

   • Denmark (Secondary Care): Data from the Danish National Patient Registry (DNPR) and prescription records from 12 public hospitals (2008–2016), covering approximately 1.3 million patients. Diagnoses were recorded as ICD-10 codes, and medications as ATC codes.
   • United Kingdom (Primary Care): Data from the UK Biobank (UKBB), linking approximately 225,000 participants to General Practitioner (GP) records. Diagnoses were originally in Read codes and prescriptions in dm+d/VMP codes, which were mapped to ICD-10 and ATC codes using official NHS mappings.

2. Data Processing and Filtering:

   • ICD-10 chapters describing external events (XIX–XXII) were excluded.
   • Drug-diagnosis pairs were defined as co-occurring if the prescription occurred within a five-day window of the diagnosis.
   • Pairs were filtered for prevalence (observed in at least 50 unique patients) to ensure statistical robustness.
   • Association Testing: Fisher's exact test was applied to all pairs, followed by False Discovery Rate (Benjamini-Hochberg) correction. Only pairs with a corrected p-value < 0.05 and an Odds Ratio (OR) > 3.5 were retained.

3. Expert Curation and Annotation:

   • Six independent clinicians (three from Denmark, three from the UK) reviewed the significant pairs specific to their respective healthcare systems.
   • Labels: Clinicians assigned "yes" (direct treatment), "no" (indirect association, e.g., comorbidity, symptom, side effect), or "unsure."
   • Scoring: To handle inter-rater variability and uncertainty, the authors employed a probabilistic label-aggregation framework based on the Dawid-Skene model using an Expectation-Maximization (EM) algorithm. This estimated the latent "true" class for each pair while modeling clinician-specific labeling patterns.
   • Output: A signed confidence score ranging from -1 to +1 was generated for each pair. Positive scores indicate direct treatment; negative scores indicate indirect associations; the magnitude reflects confidence.

Key Results

• Resource Scale: The final resource comprises 7,763 unique, significantly associated drug-diagnosis pairs (6,434 from Denmark, 1,615 from the UK) with OR > 3.5.
• Distribution Differences:
  • Denmark (Hospital): Dominated by ICD Chapter II (Neoplasms, 22.32%) and ATC Chapter N (Nervous system, 18.53%), reflecting surgical and oncological care.
  • UK (Primary Care): Dominated by ICD Chapters I (Infectious), V (Mental/Behavioral), and X (Respiratory), and ATC Chapters C (Cardiovascular) and R (Respiratory), reflecting chronic disease management.
• Annotation Findings:
  • The majority of statistically significant associations were not direct treatments. In Denmark, ~65% of pairs were unanimously labeled "no," whereas in the UK, ~20% were labeled "yes" by all three clinicians.
  • Inter-rater Agreement: Fleiss' kappa indicated moderate agreement in Denmark (0.545) and fair agreement in the UK (0.283). The lower UK agreement highlighted significant variability in clinician interpretation and the complexity of primary care prescribing.
  • Overlap: Only 268 pairs (4.16% of Danish, 16.78% of UK) were common to both datasets. These shared pairs showed higher consensus on direct treatment, particularly for cardiovascular and respiratory conditions.
• Confidence Scores: The distribution of scores revealed that Danish pairs had a higher density of negative scores (indirect associations), while UK pairs showed a higher density of positive scores but also greater uncertainty (scores clustered near zero).

Significance and Claims

The authors claim that this resource provides a practical, empirically grounded framework for studying medication use in historical EHR data, specifically addressing the gap between ontology-based research tools and clinical coding realities (ICD/ATC).

Key claims regarding the resource's utility include:

1. Distinction of Direct vs. Indirect Use: The resource explicitly demonstrates that statistical co-occurrence does not equal therapeutic indication. It allows researchers to differentiate between drugs prescribed for a specific diagnosis versus those used for comorbidities, prophylaxis, or side-effect management.
2. Handling Heterogeneity: By integrating data from primary and secondary care across two countries, the resource captures setting-specific variations in prescribing and coding, offering a more comprehensive view of real-world drug usage than single-system studies.
3. Probabilistic Ground Truth: The use of the Dawid-Skene model to generate signed confidence scores offers a nuanced alternative to binary labels. This approach accounts for annotator disagreement and uncertainty, providing a continuous target variable that can improve the development and evaluation of data-driven predictive models and large language models (LLMs) in healthcare.
4. Complementarity: The authors position this resource as complementary to literature-based databases (like MEDI), enabling the discovery of patterns (such as symptom treatment) that are not captured in official drug indications.

The study concludes that while the majority of drug-diagnosis associations in EHRs are indirect, the curated confidence scores provide a necessary reference for validating models and interpreting drug usage patterns in large-scale translational research.