AyurAssist

AI-powered clinical decision support that bridges modern medical terminology (SNOMED CT, UMLS) with traditional Ayurvedic knowledge using the WHO International Terminologies for Ayurveda (ITA).

Architecture

AyurAssist runs on Modal as a CPU-only serverless system. A terminology bridge (NER + UMLS + WHO-ITA) provides Ayurvedic vocabulary context to Qwen3-32B (via Groq API), which performs clinical reasoning across 13 structured assessment categories.

Browser (static site on GitHub Pages)
  |
  |─ GET /warmup ──> [CPU Container] ──> {"status": "ready"}
  |
  |─ POST / ──────> [CPU Container: ASGI + NER + UMLS + ITA + Groq]
                      |
                      ├─ 1. Input cleaning (strip conversational filler)
                      ├─ 2. NER extraction (scispacy en_core_sci_lg)
                      ├─ 3. Terminology bridge: NER → UMLS → ITA fuzzy matching
                      ├─ 4. UMLS lookup (CUI, SNOMED, ICD-10)
                      ├─ 5. Three-pass diagnosis via Groq (sequential):
                      │     Pass 1: Modern medical diagnosis
                      │     Pass 2: Ayurvedic diagnosis (+ ITA vocab + few-shot)
                      │     Pass 3: Ayurvedic treatment (+ ITA vocab + few-shot)
                      ├─ 6. 13 detail questions via Groq (parallel, semaphore=5)
                      └─ 7. Assemble JSON response with codes + treatment

Terminology Bridge

The bridge provides context to the LLM, not the diagnosis itself:

• Biomedical NER -- scispacy en_core_sci_lg extracts medical entities with stopword filtering to reduce noise.
• UMLS mapping -- Two-step API lookup: keyword → CUI, then CUI atoms → SNOMED CT + ICD-10 codes.
• WHO-ITA vocabulary -- 3,550 Ayurvedic terms from who-ita/ita_terms_ascii.csv matched via fuzzy English matching (threshold 0.80). Matched terms are injected into LLM prompts as a constrained vocabulary dictionary.

LLM (Qwen3-32B via Groq)

Three sequential diagnostic passes followed by 13 parallel detail questions:

1. Modern diagnosis -- NER entities + patient narrative → modern condition names
2. Ayurvedic diagnosis -- Modern diagnosis + ITA vocabulary + few-shot examples → Sanskrit condition names
3. Ayurvedic treatment -- Ayurvedic diagnosis + ITA vocabulary + few-shot examples → treatment plan
4. 13 detail questions -- Run in parallel (semaphore=5) covering dosha analysis, symptoms, single remedies, formulations, panchakarma, diet, yoga, prognosis, modern correlation, differential diagnosis, investigations, prevention, and psychotherapy

Data Flow

Patient narrative
    → Clean input (strip conversational filler)
    → NER: scispacy extracts biomedical entities
    → Bridge: entities → UMLS preferred terms → ITA fuzzy match
    → UMLS: cleaned phrase → CUI, SNOMED, ICD-10
    → Groq: 3 sequential passes (modern dx → ayur dx → ayur tx)
    → Groq: 13 parallel detail questions with ITA context
    → Response: structured JSON with codes + 13 treatment sections

Project Structure

AyurAssist/
├── main.py                        # Modal backend (CPU-only, Groq + ITA bridge)
├── config.py                      # All tuneable constants (no secrets)
├── who-ita/
│   └── ita_terms_ascii.csv        # 3,550 WHO-ITA Ayurvedic terms
├── ayurveda_snomed_mapping.csv    # 226 WHO ITA conditions → SNOMED (reference)
├── docs/                          # Frontend (GitHub Pages)
│   ├── index.html
│   ├── style.css
│   └── app.js
└── README.md

Configuration

All tuneable constants are in config.py. Edit this file to change model names, timeouts, thresholds, etc.

Modal Secrets

Secrets are not stored in code. They are managed through Modal's secret manager. You need to create two secrets in your Modal dashboard:

Modal secret name	Required env vars	How to get it
groq-secret	GROQ_API_KEY	Groq console
my-umls-secret	UMLS_API_KEY	UMLS license

Create them with the Modal CLI:

modal secret create groq-secret GROQ_API_KEY=gsk_xxxxxxxxxxxxx
modal secret create my-umls-secret UMLS_API_KEY=xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxx

Frontend

Update API in docs/app.js to match your Modal deployment URL:

var API='https://<your-modal-username>--ayurparam-service-fastapi-app.modal.run';

Deployment

Prerequisites

• Python 3.11+
• Modal CLI installed and authenticated
• Modal secrets created (see above)
• A Groq API key with access to qwen/qwen3-32b

Deploy to Modal

modal deploy main.py

This builds the CPU container image, deploys the ASGI web endpoint, and outputs your deployment URL.

Local development

modal serve main.py

This runs the app locally with hot-reload. The terminal will show the temporary URL.

API

GET /warmup

Returns immediately to confirm the service is ready.

Response:

{"status": "ready"}

POST /

Analyzes a patient narrative and returns Ayurvedic treatment recommendations across 13 categories.

Request:

{"text": "patient complains of severe headache and nausea"}

Response:

{
  "input_text": "...",
  "clinical_entities": [{"word": "headache", "score": 1.0, "entity_group": "ENTITY"}],
  "umls_cui": "C0018681",
  "snomed_code": "25064002",
  "snomed_name": "Headache",
  "icd10_code": "R51",
  "csv_match": {"ita_id": "...", "ayurveda_term": "...", "sanskrit_iast": "...", "sanskrit": "...", "description": ""},
  "results": [{
    "ayurveda_term": "Shiroroga",
    "snomed_code": "25064002",
    "icd10_code": "R51",
    "treatment_info": {
      "condition_name": "Shiroroga",
      "sanskrit_name": "...",
      "disclaimer": "...",
      "ayurparam_responses": {
        "overview_dosha_causes": "...",
        "symptoms": "...",
        "single_drug_remedies": "...",
        "classical_formulations": "...",
        "panchakarma": "...",
        "diet_lifestyle": "...",
        "yoga": "...",
        "prognosis": "...",
        "modern_correlation_warnings": "...",
        "differential_diagnosis": "...",
        "investigations_labs": "...",
        "prevention_recurrence": "...",
        "psychotherapy_satvavajaya": "..."
      }
    }
  }]
}

Cost

Component	When running	When idle
CPU container	~$0.04/hr	$0 (scales to zero after 5 min)
Groq API	Pay per token (pricing)	$0

No GPU costs — all LLM inference is handled by Groq's hosted API.

Key Dependencies

Package	Purpose
scispacy + en_core_sci_lg	Biomedical NER entity extraction
groq	Qwen3-32B API client
modal	Serverless CPU container
fastapi	ASGI web framework
requests	UMLS API calls

Experiments

Validation experiments are on the experiments branch:

• BhashaBench-Ayur (BBA) -- Ayurvedic medical knowledge benchmark with 9,348 English and 5,615 Hindi multiple-choice questions. AyurParam-2.9B scores 40.0% overall (best in the <3B class), while Qwen3-32B achieves 54.2% (best overall across all model sizes).
• Ablation Study -- Compares three pipeline configurations (direct LLM, full pipeline with ITA bridge, bridge-only) across 80 gold-standard clinical vignettes. The full pipeline achieves 80% diagnosis accuracy vs 75% for direct LLM.
• Inter-Rater Reliability -- IRR analysis of four clinicians (2 MDs, 2 Ayurveda practitioners) rating 80 patient narratives, with LLM-augmented annotation for missing labels.

Citation

If you use AyurAssist in your research, please cite:

@inproceedings{kuruvikkattil2026ayurassist,
  title={AyurAssist: Bridging Ayurvedic and Biomedical Clinical Knowledge Through Terminology-Grounded LLM Reasoning},
  author={Kuruvikkattil, Aravind V. and Pradeep, Anagha and Kumar, Veena S. and Purkayastha, Saptarshi},
  year={2026}
}