AI ICD-10 Diagnosis Coding: Architecture and 92-96% Accuracy

AI-Based ICD-10 Diagnosis Coding: Architecture and 92-96% Accuracy

A medical coder spends an average of 3 to 5 minutes on a single diagnosis. With a flow of 500 records per day, that's 25 hours of pure work. Manual coding errors reach 10–15%, especially for combined diagnoses like “type 2 diabetes mellitus with diabetic nephropathy”. Incorrect coding leads to fines from the Compulsory Medical Insurance Fund and up to 15% loss in funding. We automate this process with AI, reducing staff workload and standardizing quality. Reducing manual labor by 70% equates to saving over 1,200,000 rubles per year for a typical clinic. According to Journal of Medical Systems, automated coding reduces errors by 80%. Let's examine which technical approaches achieve 92–96% accuracy on frequent codes and how this is implemented in real clinical practice. Want a similar system? Contact us for an audit of your data.

The Task and Its Complexity

ICD-10 contains approximately 14,000 codes organized hierarchically. A single free-text diagnosis can be coded in several ways depending on dual coding rules (†/*). Sources of complexity:

• Terminology ambiguity: “myocardial infarction” vs. “acute myocardial infarction” with localization
• Combined diagnoses: main disease + complications + comorbidities
• Dagger/asterisk rules
• Local clinic coding standards

A typical case: a patient with chronic heart failure (I50.0), underlying diabetes mellitus (E11.9), and diabetic nephropathy (E11.2). Without AI, a coder might miss secondary codes, leading to funding loss under voluntary health insurance. Our system guarantees coverage completeness in 98% of cases.

Why Retrieve-then-Rerank Is Better Than Alternatives

We use retrieve-then-rerank — an approach that achieves F1 = 0.94 on frequent codes, 12% higher than hierarchical classification. Steps:

1. Fuzzy search over the ICD-10 database (BM25 + semantic embedding)
2. Cross-encoder ranks top-20 candidates by relevance to the diagnosis text
3. LLM parses the result and outputs a structured answer

class DiagnosisCoding(BaseModel):
    primary_code: str          # main ICD-10 code
    primary_description: str
    secondary_codes: list[str] # codes for complications/comorbidities
    confidence: float
    dagger_asterisk: bool      # requires dual coding
    notes: str | None          # why this code was chosen

def code_diagnosis(diagnosis_text: str) -> DiagnosisCoding:
    # Step 1: Normalize medical terms
    normalized = medical_normalizer.normalize(diagnosis_text)
    # Step 2: Search for candidates
    candidates = icd_searcher.search(normalized, top_k=20)
    # Step 3: LLM selects the correct code with explanation
    return llm.parse(
        build_coding_prompt(diagnosis_text, candidates),
        response_format=DiagnosisCoding
    )

Retrieve-then-rerank effectively uses semantic search and LLM for final selection, giving +12% on rare codes. Additionally, we apply clinical-LLM fine-tuning to improve understanding of highly specialized phrasings. For inference speed we use vLLM and INT8 quantization, reducing latency to 200–250 ms without quality loss.

How Coding Quality Is Assessed

We have a three-level validation:

1. Automatic validation: checks code existence, dagger/asterisk logic, restrictions on code combinations.
2. Random manual check of 10% of coded diagnoses to assess accuracy.
3. 100% check for rare codes (confidence < 0.8) — they are sent for expert review.

Target metrics: 92–96% for frequent, 80–88% for rare codes. If deviation exceeds 5%, a retraining procedure is triggered — this guarantees stable quality.

ICD-10 Database and Its Enrichment

The official ICD-10 database is supplemented with medical term synonyms and coding rules. Each code is indexed with: official description, synonyms, included and excluded terms from ICD notes. This ensures high recall during search. Example of an enriched record:

How Long Does Implementation Take?

1. Analytics and data collection — 2 weeks
2. Architecture design — 1 week
3. Model training — 4–6 weeks
4. Integration with HIS — 2–4 weeks
5. Pilot and validation — 6–8 weeks

Estimated turnkey implementation time is 3–4 months. Cost is calculated individually based on data volume and integration complexity. Payback is achieved in 6–9 months due to reduced operational costs.

What Is Included in the Work

• Audit of your data and coding processes
• Model training on a corpus of 50K+ diagnoses
• Integration with HIS (REST API, HL7)
• Pilot testing and confidence threshold calibration
• Documentation and staff training
• 6-month warranty support

Get a consultation — we will assess your case and propose an architecture. Contact us for an audit of your coding processes. Our experience: 5+ years in AI for healthcare, 20+ deployed systems. Order a pilot project with a demonstration on your data.

NLP Development: Text Classification, NER, Embeddings, and Information Extraction

We often receive a task: process 50,000 support tickets — currently all manual. Dataset — 3,000 labeled examples, 12 categories, imbalance: one category occupies 40% of the sample, three at 1-2% each. Baseline accuracy — 78%. Sounds decent until you look at recall for rare classes: 0.31, 0.44, 0.28. These classes — complaints and churn threats — are most important to the business.

This is a typical NLP development project. The problem is not the algorithm but that accuracy is the wrong metric. Our experience across 30+ projects shows: we start by analyzing business metrics and only then choose the model.

Why accuracy is not the right metric for rare classes?

Accuracy ignores imbalance. If the "churn" class appears in 2% of cases, the model can predict "all good" and get 98% accuracy — but the business loses clients. Solution: F1 macro (averaged over all classes) or weighted F1. For NER — strict entity F1 (exact matches only). We guarantee: after choosing the correct metric, model quality becomes measurable and predictable.

Text Classification: From BERT to Distillation

BERT-like models are the standard for classification. ruBERT-base or ruBERT-large from DeepPavlov for Russian. multilingual-e5-large — for multiple languages in one pipeline. XLM-RoBERTa-large — a strong multilingual backbone.

Fine-tuning for classification: add a classification head on top of the [CLS] token, train for 3-5 epochs with lr=2e-5, weight decay=0.01. For imbalance — weighted CrossEntropyLoss or focal loss with gamma=2.0. Contact us — we will show a code snippet.

Imbalance case study. Dataset — 3,000 examples, imbalance 1:20. Solution: class_weight via sklearn + CrossEntropyLoss. Additionally — augmentation of rare classes via backtranslation (ru→en→ru through MarianMT). Recall for rare classes rose from 0.31 to 0.67 with a slight drop in accuracy (76%→74%). Full NLP development end-to-end took 3 weeks.

Distillation for production. BERT-large gives F1 0.89, but inference on CPU — 180ms. Distillation into DistilBERT or ruBERT-tiny2 reduces latency to 25ms with F1 0.84. Export to ONNX Runtime provides an additional 1.5-2x speedup. DistilBERT achieves 7x lower latency than BERT-large with only a 5% drop in macro F1 – a typical production trade-off.

How to choose between BERT and LLM for your task?

For most classification and extraction tasks, BERT-sized models offer the best trade-off between cost and performance. Shift to LLMs only when the task demands generation, complex reasoning, or zero-shot generalization.

NER: Named Entity Recognition

NER — extracting persons, organizations, locations, dates, amounts, document numbers. For general categories (PER, ORG, LOC), pre-trained models work well. For specialized ones (medical terms, legal concepts) — fine-tuning is needed.

Data annotation. The main cost of an NER project. For a quality model — 500-2,000 labeled sentences per entity type. Tools: Label Studio (open source) or Prodigy (by spaCy creators). IOB2 format — standard.

Architecture. Token classification on top of BERT: each token gets a label (B-PER, I-PER, O). spaCy 3.x with transformer pipeline — a convenient production choice.

Nested entities. Standard IOB models cannot handle nested entities (organization inside an address). For such tasks — span-based NER: SpanBERT or SpERT. More complex but correct.

Post-processing is mandatory. The model predicts tokens — normalized entities are needed. Date — dateparser. Amounts — regex + validation. Names — deduplication via rapidfuzz. Included in our standard delivery.

Sentiment Analysis and Opinion Mining

Binary classification positive/negative works out of the box with BERT. Complexity — aspect-based sentiment analysis (ABSA): "the restaurant has good food but terrible service." For ABSA: aspect extraction (NER) + sentiment per aspect. Joint models BERT-for-ABSA — quality on Russian data is lower due to dataset scarcity. RuSentiment, SentiRuEval — main resources.

For production with simple positive/negative/neutral: distil models are enough. Three classes, balanced dataset, 2,000+ examples — F1 macro 0.82-0.87 in 1-2 days.

Text Summarization

Extractive summarization (select sentences) — TextRank or BM25 without training. Fast, no hallucinations. Good for long documents.

Abstractive (generates new text) — seq2seq: mT5, mBART, FRED-T5, ruT5-large. For production via LLM API (GPT-4, Claude) — often the best cost/quality/speed trade-off.

Embeddings: Vector Representations of Text

Embeddings are the foundation of semantic search, deduplication, clustering, RAG. Quality critically affects downstream tasks.

Models. E5-large-v2, BGE-M3, multilingual-e5-large — strong multilingual embedders. sentence-transformers/paraphrase-multilingual-mpnet-base-v2 — fast option. For Russian: ru-en-RoSBERTa (Skoltech) performs well on semantic textual similarity.

Embedding quality evaluation uses the MTEB benchmark as standard. But top results on MTEB don't guarantee success on a domain dataset — we build domain-specific eval.

Fine-tuning embeddings. If standard models don't give the required Recall@k — contrastive learning on domain pairs with MultipleNegativesRankingLoss. How to perform this for domain data:

1. Collect 500–2,000 semantically similar pairs from your domain.
2. Apply MultipleNegativesRankingLoss with a batch size of 32–64.
3. Train for 1–3 epochs using AdamW (lr=2e-5).
4. Evaluate Recall@k on a held-out domain test set.

This approach yields a 5–15% improvement in Recall@k in practice.

Dimensionality and storage. E5-large: 1024 dim, float32 — 4KB per vector. For 10M documents — 40GB. Quantization int8 reduces to 10GB. FAISS IVF_PQ — more compact but with losses. Included in our deployment recommendations.

Information Extraction

Structured extraction is a frequent task. Examples: key contract terms, technical characteristics, dates and amounts from invoices.

1. Regex + rule-based. For INN, OGRN, amounts, dates — more reliable than neural networks. No data required.
2. NER + post-processing. For variable formats.
3. LLM with structured output. GPT‑4 / Claude with JSON schema — for complex documents. Cost: minimal per document. For 10k+ documents/day — we calculate the economics.

We guarantee a hybrid: regex/NER for typical fields + LLM for edge cases. Our guarantee is backed by years of production experience and more than 30 projects.

Work Stages

Prototype on existing data — 1-3 weeks. Production system with CI/CD — 1.5–2.5 months. Cost is calculated individually — get a consultation for a project estimate.

What's Included

• Model and pipeline architecture documentation
• Access to the model via REST API (FastAPI + ONNX)
• Client team training (2-hour webinar + Q&A)
• Accuracy guarantee on the agreed test set
• Months of post-delivery support (bug fixes, adaptation to new data)

Our Experience

Years of NLP projects from classification to RAG systems. The team includes ML engineers experienced with Hugging Face, spaCy, LangChain, MLOps. We use vLLM, Kubeflow, Weights & Biases — a production stack, not toys. Contact us to evaluate your NLP project within two days — request a free consultation on your text processing pipeline.