ATC and ICD-10 are complementary, not interchangeable, healthcare classification systems. ATC codes describe medicines by anatomical, therapeutic and chemical properties, while ICD-10 codes describe diseases, symptoms and health conditions. An LLM can help connect information expressed in clinical notes, prescriptions and claims—but it should not be treated as an authority that automatically determines the correct code.
For Indian hospitals, health-tech companies and research teams, the useful problem is therefore not “one model that converts everything.” It is a governed workflow that combines terminology standards, retrieval, deterministic checks, human review and audit trails.
ATC and ICD-10: what each system represents
The Anatomical Therapeutic Chemical (ATC) system, maintained by the WHO Collaborating Centre for Drug Statistics Methodology, groups medicines into five levels:
- Level 1: anatomical main group, such as the alimentary tract or cardiovascular system
- Level 2: therapeutic subgroup
- Level 3: pharmacological subgroup
- Level 4: chemical subgroup
- Level 5: chemical substance, usually the most specific level
ATC classification is generally based on active ingredients and therapeutic use. A brand name, strength, route and combination product may require additional mapping before an ATC code can be assigned reliably.
ICD-10 classifies diagnoses and other health-related conditions. The exact code set depends on the implementation: WHO ICD-10, a national modification, payer requirements or a local hospital dictionary. India-facing systems must also account for clinical terminology used in English and Indian languages, abbreviations, spelling variation and documentation practices.
A key distinction: an ATC code does not prove why a medicine was prescribed, and an ICD-10 code does not fully describe the treatment. The relationship between them is contextual.
What does “ATC ICD10 codes LLM” mean in practice?
The phrase can refer to an LLM-assisted system that extracts medicines and diagnoses from unstructured text, maps them to standardized codes, and links the two across a patient timeline. It is not a universally defined coding standard or a specific model architecture.
A robust pipeline may include:
1. Text extraction: identify drug names, diagnoses, symptoms, negations, dates and clinical status.
2. Normalization: map brand names to generic ingredients and resolve spelling or abbreviation variants.
3. Candidate generation: retrieve possible ATC and ICD-10 concepts from an approved terminology service.
4. Context analysis: determine whether a diagnosis is confirmed, historical, suspected, ruled out or family history.
5. Relationship linking: connect a medication to a diagnosis only when the encounter, timing and clinical evidence support that link.
6. Validation and review: apply rules, confidence thresholds and coder or clinician approval.
7. Audit storage: preserve the source text, selected code, alternatives, model version and reviewer decision.
This approach is more dependable than asking an LLM to invent codes from memory. Teams building such systems can learn from broader machine learning applications in healthcare in India, particularly around deployment constraints, evaluation and clinical workflows.
Where LLMs add value
LLMs are strongest at handling language and variation. They can assist with:
- extracting medicines from discharge summaries, prescriptions and progress notes;
- recognizing generic names, brands, misspellings and dosage expressions;
- separating active conditions from negated or historical conditions;
- suggesting likely code candidates with evidence spans from the source note;
- summarizing medication changes across longitudinal records;
- identifying missing information, such as an unclear route or unspecified diagnosis;
- translating or normalizing multilingual documentation before coding review.
The model should produce a ranked suggestion with rationale, not an unreviewed final code. For example, “hypertension” may map to a broad or more specific ICD-10 option depending on documentation. Similarly, a combination medicine may need ingredient-level analysis before an ATC mapping is selected.
For sensitive clinical applications, explainable AI models for integrative healthcare offer useful design principles: show supporting text, expose uncertainty, and make it possible for a reviewer to correct the system.
A practical architecture for Indian healthcare teams
A production design should separate the language model from the source of truth:
- Terminology layer: versioned ATC and ICD-10 dictionaries, local synonyms, formulary mappings and code-status metadata.
- Retrieval layer: searches the approved terminology set rather than relying only on model memory.
- LLM layer: extracts entities, resolves context and ranks candidates.
- Rules engine: checks negation, age or sex constraints, duplicate codes, incompatible combinations and required specificity.
- Human review interface: displays the original text, proposed code, confidence, alternatives and reason for selection.
- Data layer: stores provenance, timestamps, encounter IDs, code-set versions and reviewer actions.
- Security layer: enforces role-based access, encryption, retention controls and environment isolation.
India-specific implementation should align with applicable health-data governance, institutional ethics processes, consent requirements and security controls. If records are processed through external APIs, review data residency, retention, training-use terms and contractual safeguards before sending identifiable information. Open-source healthcare AI projects in India can be a useful starting point for teams that need greater control over deployment and data handling.
Evaluation: measure more than coding accuracy
A credible evaluation set should be created from representative Indian clinical documents, with independent review by qualified coders or clinicians. Report performance separately for medicines, diagnoses and medication–diagnosis links.
Useful measures include:
- Exact-match accuracy: whether the selected code matches the approved reference.
- Top-k recall: whether the correct code appears among the model’s suggestions.
- Entity-level precision and recall: whether medicines and conditions were extracted correctly.
- Abstention quality: whether the system declines uncertain cases instead of guessing.
- Calibration: whether confidence scores correspond to actual correctness.
- Review time and correction rate: whether the tool improves workflow without creating hidden burden.
- Subgroup performance: variation across specialties, languages, facilities and documentation styles.
Test difficult cases deliberately: brand-to-generic conversion, polypharmacy, adverse drug reactions, provisional diagnoses, chronic conditions copied forward, dose changes and multiple encounters close together. A model that performs well on clean notes may fail on real outpatient documentation.
Common failure modes and safeguards
Hallucinated or obsolete codes: use retrieval from a versioned terminology source and reject codes outside the approved release.
False medication–diagnosis links: require temporal and textual evidence; do not infer indication solely from a drug’s common use.
Negation errors: detect phrases such as “no history of,” “rule out” and “denies,” with mandatory review for high-impact cases.
Over-specific coding: never infer severity, complication, laterality or chronicity unless documented.
Privacy leakage: de-identify development data, minimise prompts and monitor logs. For a wider view of operational constraints, see AI solutions for rural healthcare in India, where connectivity, staffing and infrastructure shape system design.
Automation bias: present the model as an assistant, require confirmation for billing or clinical decisions, and track overrides as a continuous improvement signal.
A sensible rollout plan
Start with a narrow, low-risk use case such as coder assistance for retrospective analytics. Establish a gold-standard sample, terminology ownership and an escalation process. Then pilot in one department, compare time and quality against the current workflow, and expand only after monitoring drift.
Keep a human in the loop for billing, pharmacovigilance, clinical decision support and any action that can affect patient care. Build feedback capture into the interface, update terminology on a controlled schedule, and revalidate after model, code-set or workflow changes.
The best ATC–ICD-10 LLM systems are not the most autonomous. They are the ones that make coding work faster, traceable and safer, while preserving the authority of qualified healthcare professionals.