HighCostDrugsDemo/progress.txt

# Progress Log - Direct SNOMED Indication Mapping

## Project Context

This project extends the existing HCD Pathway Analysis application with direct SNOMED code matching from GP records. The previous project (Phases 1-5) established the pre-computed pathway architecture and modern UI. This phase adds:

1. **Diagnosis-based directorate assignment** - Primary method using GP SNOMED codes
2. **Indication-based icicle chart** - New chart type showing Trust → Search_Term → Drug → Pathway

## Key Files Reference

**Existing (reuse these):**
- `data_processing/schema.py` - SQLite schema (add new table)
- `data_processing/diagnosis_lookup.py` - Existing cluster-based lookup (extend with direct SNOMED)
- `data_processing/pathway_pipeline.py` - Pathway processing (add indication type)
- `cli/refresh_pathways.py` - CLI refresh command (add chart type support)
- `pathways_app/pathways_app.py` - Reflex app (add chart type toggle)
- `tools/data.py` - Data transformations including department_identification()

**New data:**
- `data/drug_snomed_mapping_enriched.csv` - 163K rows, 187 Search_Terms, 364 drugs

## Known Patterns

### SNOMED Mapping Structure
The enriched mapping CSV has columns:
- Drug, Indication, TA_ID (from NICE TAs)
- Search_Term (simplified grouping, 187 unique values)
- SNOMEDCode, SNOMEDDescription
- CleanedDrugName, PrimaryDirectorate, AllDirectorates

### Direct SNOMED Lookup Logic
For a patient on drug X:
1. Get all SNOMED codes for that drug from ref_drug_snomed_mapping
2. Query PrimaryCareClinicalCoding for those codes (patient's GP record)
3. If match found → use Search_Term and PrimaryDirectorate from matched row
4. If no match → fall back to department_identification()
5. Use most recent SNOMED code by EventDateTime if multiple matches

### Chart Type Architecture
- `chart_type` column in pathway_nodes: "directory" or "indication"
- 12 total pathway datasets: 6 date filters × 2 chart types
- Indication chart: mixed labels (Search_Term for matched, Directorate for unmatched)

### Date Filter Combinations
| ID | Initiated | Last Seen | Default |
|----|-----------|-----------|---------|
| `all_6mo` | All years | Last 6 months | Yes |
| `all_12mo` | All years | Last 12 months | No |
| `1yr_6mo` | Last 1 year | Last 6 months | No |
| `1yr_12mo` | Last 1 year | Last 12 months | No |
| `2yr_6mo` | Last 2 years | Last 6 months | No |
| `2yr_12mo` | Last 2 years | Last 12 months | No |

### Expected Volumes
- SNOMED mapping: 163K rows
- Search_Terms: 187 unique
- Pathway nodes per date filter: ~300 (directory), ~400-600 (indication)

---

## Iteration Log

## Iteration 1 — 2026-02-05
### Task: 1.1 Create SQLite Table for SNOMED Mapping
### Why this task:
- First task in Phase 1 (Data Infrastructure) — all other phases depend on having the data layer in place
- No external dependencies — pure schema definition work
- Follows "data infrastructure first" principle
### Status: COMPLETE
### What was done:
- Added `REF_DRUG_SNOMED_MAPPING_SCHEMA` to `data_processing/schema.py` with 11 columns:
  - id, drug_name, indication, ta_id, search_term, snomed_code, snomed_description
  - cleaned_drug_name, primary_directorate, all_directorates, created_at
- Added 5 custom indexes for lookup performance:
  - idx_ref_drug_snomed_mapping_drug (drug_name)
  - idx_ref_drug_snomed_mapping_cleaned (cleaned_drug_name)
  - idx_ref_drug_snomed_mapping_snomed (snomed_code)
  - idx_ref_drug_snomed_mapping_search_term (search_term)
  - idx_ref_drug_snomed_mapping_drug_snomed (composite: cleaned_drug_name, snomed_code)
- Added `create_drug_snomed_mapping_table()` helper function
- Added schema to `REFERENCE_TABLES_SCHEMA` (included in `ALL_TABLES_SCHEMA`)
- Updated helper functions to include new table:
  - `drop_reference_tables()` — drops new table
  - `get_reference_table_counts()` — counts new table (with try/except for safety)
  - `verify_reference_tables_exist()` — checks for new table
### Validation results:
- Tier 1 (Code): `python -m py_compile data_processing/schema.py` — PASSED
- Tier 1 (Code): Import check — PASSED
- Tier 2 (Data): Migration created table with 0 rows — PASSED
- Tier 2 (Data): All 11 columns and 6 indexes present — PASSED
### Files changed:
- `data_processing/schema.py` — added schema, updated helpers
### Committed: cf35937 "feat: add ref_drug_snomed_mapping schema (Task 1.1)"
### Patterns discovered:
- Existing schema pattern: CREATE TABLE + CREATE INDEX in single SQL string
- Helper functions follow consistent pattern: create, drop, get_counts, verify_exists
- `get_reference_table_counts` needed try/except for tables that might not exist yet
### Next iteration should:
- Start Task 1.2: Create `data_processing/load_snomed_mapping.py` script to load the CSV
- The CSV has columns: Drug, Indication, TA_ID, Search_Term, SNOMEDCode, SNOMEDDescription, CleanedDrugName, PrimaryDirectorate, AllDirectorates
- Note: SNOMEDCode in CSV has `.0` suffix (e.g., "156370009.0") — may need cleanup
### Blocked items:
- None

## Iteration 2 — 2026-02-05
### Task: 1.2 Load Enriched Mapping Data
### Why this task:
- Task 1.1 (schema creation) is complete — table `ref_drug_snomed_mapping` exists
- Task 1.2 is the next logical step in Phase 1 (Data Infrastructure)
- Task 1.3 (Extend Diagnosis Lookup) depends on having data in the table to query
- Following the "data infrastructure first" principle
### Status: COMPLETE
### What was done:
- Created `data_processing/load_snomed_mapping.py` with:
  - `migrate_drug_snomed_mapping()` - reads CSV, cleans SNOMED codes, inserts to SQLite
  - `get_drug_snomed_mapping_counts()` - statistics function
  - `verify_drug_snomed_mapping_migration()` - validation function
  - `clean_snomed_code()` - removes trailing .0 from SNOMED codes (e.g., "156370009.0" → "156370009")
  - CLI interface with `--verify-only` and `-v` options
- Added CLI entry point: `python -m data_processing.load_snomed_mapping`
- Fixed Unicode encoding issue on Windows console (emojis → ASCII text)
### Validation results:
- Tier 1 (Code): `python -m py_compile data_processing/load_snomed_mapping.py` — PASSED
- Tier 1 (Code): Import check — PASSED
- Tier 2 (Data): Migration loaded 144,056 rows from 163,937 CSV rows (19,881 duplicates skipped)
- Tier 2 (Data): Verified: 707 unique drugs, 187 unique search terms, 21,265 SNOMED codes
- Tier 2 (Data): SNOMED codes cleaned — 0 rows with .0 suffix remaining
### Files changed:
- `data_processing/load_snomed_mapping.py` — new file
### Committed: 6d68b5e "feat: add SNOMED mapping loader script (Task 1.2)"
### Patterns discovered:
- CSV has 163,937 rows but only 144,056 unique (drug_name, indication, snomed_code) combinations
- SNOMED codes in CSV have `.0` suffix from float formatting — need `clean_snomed_code()` function
- Windows console can't print Unicode emojis — use ASCII text like `[OK]` instead of ✅
### Next iteration should:
- Start Task 1.3: Extend Diagnosis Lookup Module
- Add `get_drug_snomed_codes(drug_name)` to `data_processing/diagnosis_lookup.py`
- Add `patient_has_indication_direct(patient_pseudonym, snomed_codes, connector)` function
- Look at existing `diagnosis_lookup.py` for patterns to follow
### Blocked items:
- None

## Iteration 3 — 2026-02-05
### Task: 1.3 Extend Diagnosis Lookup Module
### Why this task:
- Tasks 1.1 and 1.2 are complete — the data infrastructure (schema + loaded data) is in place
- Task 1.3 is the next logical step — functions to query the data before using it in the pipeline
- Phase 2 (Pathway Processing) depends on having these lookup functions
- Following the "data infrastructure first" principle
### Status: COMPLETE
### What was done:
- Added two new dataclasses to `data_processing/diagnosis_lookup.py`:
  - `DrugSnomedMapping`: Holds SNOMED code mapping with snomed_code, snomed_description, search_term, primary_directorate, indication, ta_id
  - `DirectSnomedMatchResult`: Result of direct SNOMED lookup with matched flag, snomed_code, search_term, primary_directorate, event_date, source
- Added `get_drug_snomed_codes(drug_name)` function:
  - Queries `ref_drug_snomed_mapping` table for all SNOMED codes for a drug
  - Case-insensitive matching on both `cleaned_drug_name` and `drug_name` columns
  - Returns list of DrugSnomedMapping dataclass instances
- Added `patient_has_indication_direct(patient_pseudonym, drug_snomed_mappings, connector)` function:
  - Queries `PrimaryCareClinicalCoding` directly for exact SNOMED code matches
  - Returns most recent match by EventDateTime (ORDER BY DESC LIMIT 1)
  - Handles Snowflake unavailability gracefully
- Updated `__all__` exports to include new dataclasses and functions
### Validation results:
- Tier 1 (Code): `python -m py_compile data_processing/diagnosis_lookup.py` — PASSED
- Tier 1 (Code): Import check — PASSED
- Tier 2 (Data): ADALIMUMAB returns 1320 SNOMED mappings across 10 Search_Terms
- Tier 2 (Data): RANIBIZUMAB returns 104 SNOMED mappings
- Tier 2 (Data): Case insensitivity verified (upper/lower/mixed all return same results)
- Tier 2 (Data): Empty mappings returns unmatched result correctly
### Files changed:
- `data_processing/diagnosis_lookup.py` — added 2 dataclasses, 2 functions, updated __all__
- `IMPLEMENTATION_PLAN.md` — marked Task 1.3 complete
### Committed: b44d22d "feat: add direct SNOMED lookup functions (Task 1.3)"
### Patterns discovered:
- ADALIMUMAB has 10 unique Search_Terms with varying SNOMED code counts:
  - rheumatoid arthritis: 867 codes (largest)
  - ankylosing spondylitis: 96 codes
  - uveitis: 124 codes
  - ulcerative colitis: 78 codes
  - juvenile idiopathic arthritis: 57 codes
  - crohn's disease: 50 codes
  - psoriatic arthritis: 17 codes
  - plaque psoriasis: 16 codes
  - hidradenitis suppurativa: 8 codes
  - axial spondyloarthritis: 7 codes (smallest)
- The query uses DISTINCT to avoid duplicate rows
- Function handles edge cases: empty mappings, Snowflake unavailable
### Next iteration should:
- Phase 1 is now complete! Start Phase 2: Pathway Processing Updates
- Begin with Task 2.1: Update Directorate Assignment Logic
  - Create `get_directorate_from_diagnosis(upid, drug_name, connector)` function
  - Logic: Try diagnosis-based first (using get_drug_snomed_codes + patient_has_indication_direct)
  - Fallback to department_identification() if no match
  - Return: (directorate, source) where source is "DIAGNOSIS" or "FALLBACK"
- Key consideration: This will need to integrate with the existing data flow in `fetch_and_transform_data()`
### Blocked items:
- None

## Iteration 4 — 2026-02-05
### Task: 2.1 Update Directorate Assignment Logic
### Why this task:
- Phase 1 (Data Infrastructure) is complete — all 3 tasks marked [x]
- Task 2.1 is the first task in Phase 2 (Pathway Processing Updates)
- Previous iteration explicitly recommended starting with Task 2.1
- Tasks 2.2 (schema) and 2.3 (indication processing) logically depend on having the directorate assignment logic
- Following "data infrastructure first, then pipeline" principle
### Status: COMPLETE
### What was done:
- Added `DirectorateAssignment` dataclass to `data_processing/diagnosis_lookup.py`:
  - Fields: upid, drug_name, directorate, search_term, source, snomed_code, event_date
  - source is "DIAGNOSIS" (SNOMED match found) or "FALLBACK" (no match)
- Added `get_directorate_from_diagnosis(upid, drug_name, connector, db_manager, before_date)` function:
  - Gets all SNOMED codes for drug from ref_drug_snomed_mapping
  - Queries patient's GP records (via patient_has_indication_direct)
  - Returns diagnosis-based directorate and search_term if match found
  - Returns FALLBACK result if no match (caller handles fallback logic)
  - Extracts PatientPseudonym from UPID by removing first 3 characters (ProviderCode)
- Updated `__all__` exports to include new dataclass and function
### Validation results:
- Tier 1 (Code): `python -m py_compile data_processing/diagnosis_lookup.py` — PASSED
- Tier 1 (Code): Import check for DirectorateAssignment, get_directorate_from_diagnosis — PASSED
- Tier 2 (Data): Test with ADALIMUMAB returns 1320 SNOMED mappings across 10 search_terms — PASSED
- Tier 2 (Data): Test with unknown drug returns FALLBACK source — PASSED
- Tier 2 (Data): Test without Snowflake returns FALLBACK source correctly — PASSED
### Files changed:
- `data_processing/diagnosis_lookup.py` — added DirectorateAssignment dataclass, get_directorate_from_diagnosis function, updated __all__
- `IMPLEMENTATION_PLAN.md` — marked Task 2.1 complete
### Committed: 5067694 "feat: add get_directorate_from_diagnosis() function (Task 2.1)"
### Patterns discovered:
- UPID format: Provider Code (3 chars) + PersonKey, where PersonKey = PatientPseudonym
- The function is designed to be called at the DataFrame level during pipeline processing
- For batch processing, the caller will need to iterate over rows and collect DIAGNOSIS vs FALLBACK statistics
- The function handles edge cases: no SNOMED mappings, Snowflake unavailable, no GP record match
### Next iteration should:
- Start Task 2.2: Add Chart Type Support to Schema
  - Add `chart_type` column to `pathway_nodes` table (values: "directory", "indication")
  - Update schema in `data_processing/schema.py`
  - Consider: may need ALTER TABLE migration for existing data
  - Alternative: add to pathway_date_filters or create pathway_chart_types reference table
- Key consideration: The indication chart will group by Search_Term (from SNOMED match) or Directorate (fallback)
- The chart_type column allows filtering pathway_nodes by chart type when user toggles in UI
### Blocked items:
- None

## Iteration 5 — 2026-02-05
### Task: 2.2 Add Chart Type Support to Schema
### Why this task:
- Task 2.1 (directorate assignment logic) is complete
- Task 2.2 is the next task in Phase 2 — schema changes before pipeline changes
- Task 2.3 (Indication Pathway Processing) requires the chart_type column to store indication pathways
- Following "data infrastructure first" principle — schema changes before pipeline logic
### Status: COMPLETE
### What was done:
- Modified `PATHWAY_NODES_SCHEMA` in `data_processing/schema.py`:
  - Added `chart_type TEXT NOT NULL DEFAULT 'directory'` column
  - Updated UNIQUE constraint from `(date_filter_id, ids)` to `(date_filter_id, chart_type, ids)`
  - Added `idx_pathway_nodes_chart_type` index on `(date_filter_id, chart_type)`
  - Updated `idx_pathway_nodes_filter_composite` index to include `chart_type`
- Added `migrate_pathway_nodes_chart_type(conn)` function:
  - Checks if chart_type column exists (idempotent)
  - Adds column with ALTER TABLE if missing
  - Creates/updates indexes
  - Returns (success: bool, message: str)
- Updated `data_processing/migrate.py`:
  - Added import for `migrate_pathway_nodes_chart_type`
  - Called migration in `initialize_database()` after creating tables
- Ran migration on existing database: 293 rows updated with chart_type='directory'
### Validation results:
- Tier 1 (Code): `python -m py_compile data_processing/schema.py` — PASSED
- Tier 1 (Code): `python -m py_compile data_processing/migrate.py` — PASSED
- Tier 1 (Code): Import check for migrate_pathway_nodes_chart_type — PASSED
- Tier 2 (Data): Migration ran successfully on existing database — PASSED
- Tier 2 (Data): All 293 existing rows have chart_type='directory' — PASSED
- Tier 2 (Data): Both indexes created (idx_pathway_nodes_chart_type, idx_pathway_nodes_filter_composite) — PASSED
- Tier 2 (Data): Database status check shows all tables intact — PASSED
### Files changed:
- `data_processing/schema.py` — added chart_type column, updated UNIQUE constraint, added indexes, added migration function
- `data_processing/migrate.py` — import and call migration function
- `IMPLEMENTATION_PLAN.md` — marked Task 2.2 complete with full checklist
### Committed: 19607d7 "feat: add chart_type column to pathway_nodes schema (Task 2.2)"
### Patterns discovered:
- SQLite ALTER TABLE ADD COLUMN works with DEFAULT values — existing rows get the default
- SQLite doesn't allow modifying UNIQUE constraints after table creation, but since all existing rows have the same chart_type value, the new constraint is satisfied
- Migration function is idempotent — safe to run multiple times
- Composite indexes including chart_type allow efficient filtering by (date_filter_id, chart_type)
### Next iteration should:
- Start Task 2.3: Create Indication Pathway Processing
  - Add `process_indication_pathways()` to `pathway_pipeline.py`
  - Group by: Trust → Search_Term → Drug → Pathway
  - For unmatched patients: use directorate name as Search_Term fallback
  - Add `extract_indication_fields()` for denormalized columns
- Key consideration: Need to modify `generate_icicle_chart()` or create parallel function for indication hierarchy
- The indication chart uses Search_Term (from SNOMED match) or Directorate (from fallback) as the level-2 grouping
- Output structure should match directory chart but with different grouping labels
### Blocked items:
- None

## Iteration 6 — 2026-02-05
### Task: 2.3 Create Indication Pathway Processing
### Why this task:
- Tasks 2.1 and 2.2 are complete — directorate assignment logic and schema are in place
- Task 2.3 is the next logical step in Phase 2 — processing logic before Phase 3 (CLI updates)
- Previous iteration explicitly recommended starting Task 2.3
- Phase 3 depends on having the indication pathway processing functions
- Following "pipeline before UI" principle
### Status: COMPLETE
### What was done:
- Added `generate_icicle_chart_indication()` to `analysis/pathway_analyzer.py`:
  - Variant of generate_icicle_chart() that uses indication_df instead of directory_df
  - Takes `indication_df` parameter mapping UPID → Indication_Group
  - The indication_df must have 'Directory' column (renamed from Indication_Group for compatibility)
  - Hierarchy: Trust → Indication_Group → Drug → Pathway

- Added `process_indication_pathway_for_date_filter()` to `data_processing/pathway_pipeline.py`:
  - Wrapper function that calls generate_icicle_chart_indication()
  - Takes indication_df parameter (UPID → Indication_Group mapping)
  - Computes date ranges and passes to the chart generator

- Added `extract_indication_fields()` to `data_processing/pathway_pipeline.py`:
  - Similar to extract_denormalized_fields() but for indication charts
  - Extracts: trust_name, directory (stores search_term), drug_sequence
  - Uses 'directory' column for schema compatibility

- Updated `convert_to_records()` with `chart_type` parameter:
  - Added chart_type to the record dictionary
  - Supports "directory" and "indication" values
  - Logs chart_type in output message

- Added `ChartType` type alias: `Literal["directory", "indication"]`

- Updated `__all__` exports to include new functions and type

### Validation results:
- Tier 1 (Code): `python -m py_compile data_processing/pathway_pipeline.py` — PASSED
- Tier 1 (Code): `python -m py_compile analysis/pathway_analyzer.py` — PASSED
- Tier 1 (Code): Import check for all new functions — PASSED
  - ChartType, process_indication_pathway_for_date_filter, extract_indication_fields all exported
  - generate_icicle_chart_indication importable from pathway_analyzer
### Files changed:
- `analysis/pathway_analyzer.py` — added generate_icicle_chart_indication() function
- `data_processing/pathway_pipeline.py` — added ChartType, process_indication_pathway_for_date_filter(), extract_indication_fields(), updated convert_to_records()
- `IMPLEMENTATION_PLAN.md` — marked Task 2.3 complete with full checklist
### Committed: 7cbc648 "feat: add indication pathway processing functions (Task 2.3)"
### Patterns discovered:
- The build_hierarchy() function uses directory_df to map UPID → Directory for level-2 grouping
- For indication charts, we pass indication_df with 'Directory' column (renamed from Indication_Group) to build_hierarchy()
- The indication_df must be indexed by UPID (same as directory_df)
- Schema compatibility is maintained by storing search_term in the 'directory' column
- The chart_type column allows filtering pathway_nodes by chart type in Reflex UI
### Next iteration should:
- Phase 2 is now complete! Start Phase 3: CLI & Data Refresh Updates
- Begin with Task 3.1: Update Refresh Command for Dual Chart Types
  - Modify `cli/refresh_pathways.py` to process both "directory" and "indication" chart types
  - Add `--chart-type` argument: "all" (default), "directory", "indication"
  - For each of 6 date filters: generate 2 chart datasets
  - Total: 12 pathway datasets (6 dates × 2 chart types)
- Key consideration: Need to create indication_df by looking up GP diagnoses for all patients
  - This involves batching Snowflake queries for performance
  - Use get_directorate_from_diagnosis() function from diagnosis_lookup.py
  - For unmatched patients: indication_group = "Directory (no GP dx)"
### Blocked items:
- None

## Iteration 7 — 2026-02-05
### Task: 3.1 Update Refresh Command for Dual Chart Types
### Why this task:
- Phase 2 complete — all data infrastructure and processing functions ready
- Task 3.1 is first in Phase 3 — CLI command is the entry point for the pipeline
- Previous iteration recommended starting Task 3.1
- Processing functions exist but weren't integrated into refresh workflow yet
### Status: COMPLETE
### What was done:
- Modified `cli/refresh_pathways.py`:
  - Added `--chart-type` argument with choices: "directory", "indication", "all"
  - Default is "directory" to maintain backward compatibility
  - Updated `insert_pathway_records` to include `chart_type` column (required for new schema)
  - Added `chart_type` parameter to `refresh_pathways()` function
  - Updated Step 2 to loop through chart_types_to_process list
  - For "directory": uses existing `process_all_date_filters()`
  - For "indication": placeholder with warning (requires Task 3.2 for GP diagnosis)
  - Updated logging to show per-chart-type counts
  - Updated help text with new examples
- Added imports for new pathway_pipeline functions:
  - `ChartType`, `DATE_FILTER_CONFIGS`, `process_indication_pathway_for_date_filter`
  - `extract_indication_fields`, `convert_to_records`
### Validation results:
- Tier 1 (Code): `python -m py_compile cli/refresh_pathways.py` — PASSED
- Tier 1 (Code): Import check for `refresh_pathways`, `main` — PASSED
- Tier 1 (Code): `python -m cli.refresh_pathways --help` — Shows new argument — PASSED
- Tier 2 (Data): Not fully testable until Task 3.2 (GP diagnosis integration)
### Files changed:
- `cli/refresh_pathways.py` — added chart_type support
- `IMPLEMENTATION_PLAN.md` — marked Task 3.1 items complete
### Committed: 593d14c "feat: add chart_type argument to refresh command (Task 3.1)"
### Patterns discovered:
- The `convert_to_records()` function already has `chart_type` parameter with default "directory"
- Results dictionary now keyed by "date_filter_id:chart_type" (e.g., "all_6mo:directory")
- Stats now include `chart_type_counts` for summary by chart type
### Next iteration should:
- Start Task 3.2: Integrate Diagnosis-Based Directorate in Pipeline
  - This is the key task that enables indication chart processing
  - Need to add batch GP diagnosis lookup during `fetch_and_transform_data()`
  - Create `indication_df` mapping UPID → Indication_Group (Search_Term or fallback)
  - Call `process_indication_pathway_for_date_filter()` with the indication_df
  - Key consideration: Batch Snowflake queries for performance (don't query per patient)
  - Expected: ~35K patients, need to query in batches of ~1000
  - Coverage logging: "X% diagnosis-matched, Y% fallback"
### Blocked items:
- None

## Iteration 8 — 2026-02-05
### Task: 3.2 Integrate Diagnosis-Based Directorate in Pipeline
### Why this task:
- Task 3.1 complete — CLI argument added but indication processing was placeholder
- Task 3.2 is the key task that enables actual indication chart processing
- Previous iteration explicitly recommended starting Task 3.2
- Task 3.3 (full pipeline test) and Phase 4 (UI) depend on this being complete
- Following "pipeline before UI" principle
### Status: COMPLETE
### What was done:
- Added `batch_lookup_indication_groups()` to `data_processing/diagnosis_lookup.py`:
  - Efficient batch function to look up GP diagnoses for all patients
  - Queries Snowflake in batches of 500 patients (configurable batch_size)
  - Gets all SNOMED codes for drugs from local SQLite (fast)
  - Builds single query per batch checking all patient-SNOMED combinations
  - Returns DataFrame with: UPID, Indication_Group, Source
  - Indication_Group is Search_Term (if matched) or "Directory (no GP dx)" (if fallback)
  - Source is "DIAGNOSIS" or "FALLBACK"
  - Logs coverage statistics: X% diagnosis-matched, Y% fallback
- Updated `cli/refresh_pathways.py` indication chart processing:
  - Import batch_lookup_indication_groups
  - When processing indication chart type:
    1. Call batch_lookup_indication_groups(df) to create indication_df
    2. Log coverage statistics to stats dict
    3. Rename Indication_Group → Directory for compatibility with generate_icicle_chart_indication
    4. Set index to UPID for lookup during chart generation
    5. Process all 6 date filters with process_indication_pathway_for_date_filter()
    6. Extract indication fields and convert to records with chart_type="indication"
  - Added error handling with fallback to empty results if GP lookup fails
- Added TYPE_CHECKING import for pandas type hints
### Validation results:
- Tier 1 (Code): `python -m py_compile data_processing/diagnosis_lookup.py` — PASSED
- Tier 1 (Code): `python -m py_compile cli/refresh_pathways.py` — PASSED
- Tier 1 (Code): Import check for batch_lookup_indication_groups — PASSED
- Tier 1 (Code): `python -m cli.refresh_pathways --help` — Shows all arguments — PASSED
- Tier 2 (Data): Not fully testable without Snowflake connection (requires --dry-run with SSO)
### Files changed:
- `data_processing/diagnosis_lookup.py` — added batch_lookup_indication_groups(), TYPE_CHECKING import
- `cli/refresh_pathways.py` — integrated batch lookup, added full indication processing flow
- `IMPLEMENTATION_PLAN.md` — marked Task 3.2 items complete
### Committed: 8952156 "feat: integrate batch GP diagnosis lookup for indication charts (Task 3.2)"
### Patterns discovered:
- Batch Snowflake queries: Build one query with IN clauses for both patients AND SNOMED codes
- ORDER BY EventDateTime DESC in query lets us pick first result = most recent in Python
- PersonKey column = PatientPseudonym (used directly for Snowflake lookup)
- indication_df must be indexed by UPID and have 'Directory' column (renamed from Indication_Group)
- Fallback label format: "Directory (no GP dx)" distinguishes matched vs unmatched in chart
### Next iteration should:
- Start Task 3.3: Test Full Refresh Pipeline
  - Run `python -m cli.refresh_pathways --chart-type all` with real data (requires Snowflake SSO)
  - Verify pathway_nodes table has both chart_type="directory" and chart_type="indication"
  - Verify indication chart hierarchy: Trust → Search_Term → Drug → Pathway
  - Verify unmatched patients show with "Directory (no GP dx)" labels
  - Document: Processing time, record counts, coverage percentages
  - If no Snowflake access, skip to Phase 4 (UI) and note as blocked
### Blocked items:
- Task 3.3 verification requires Snowflake connection (NHS SSO)

## Iteration 9 — 2026-02-05
### Task: 3.3 Test Full Refresh Pipeline
### Why this task:
- Previous iteration completed Task 3.2 but noted verification was pending
- Task 3.3 is the final task in Phase 3 to validate all pipeline work
- Phase 4 (UI) depends on having working indication chart data
- Following "validate before moving forward" principle
### Status: IN PROGRESS
### What was done:
1. **First refresh attempt**: Ran `python -m cli.refresh_pathways --chart-type all --verbose`
   - Directory chart: Processed successfully (293 nodes for all_6mo)
   - Indication chart: **0% diagnosis matches** - all 37,257 patients fell back to FALLBACK
   - This resulted in no indication pathway data (charts empty)

2. **Diagnosed root cause #1**: SNOMED codes stored in scientific notation
   - CSV has codes like "1.0629311000119108e+16" due to pandas/Excel export
   - The `clean_snomed_code()` function only handled ".0" suffix removal
   - Codes were stored as "1.06e+16" which never match Snowflake data
   - **Fix**: Updated `clean_snomed_code()` to convert scientific notation to integers
   - Reloaded 144,056 SNOMED mappings with properly formatted codes

3. **Diagnosed root cause #2**: Wrong patient identifier used for GP lookup
   - `batch_lookup_indication_groups()` was using `PersonKey` column
   - `PersonKey` = `LocalPatientID` (provider-specific like "J188448")
   - GP records use `PatientPseudonym` which matches `PseudoNHSNoLinked` (SHA-256 hash)
   - **Fix**: Changed to use `PseudoNHSNoLinked` column for GP record matching
   - Test showed ~20% match rate for ADALIMUMAB patients with correct identifier

4. **Committed fixes**: `5b1569e` "fix: correct patient identifier for GP diagnosis lookup (Task 3.3)"

5. **Started second refresh**: Running in background (task ID: be9b9e7)
   - Processing time expected: ~15-20 minutes total
   - Should now show non-zero GP matches

### Validation results:
- Tier 1 (Code): Syntax check passed for both modified files
- Tier 1 (Code): Import check passed
- Tier 2 (Data): SNOMED codes now properly formatted (0 scientific notation entries)
- Tier 2 (Data): GP record matching test: 20 matches found in 100 ADALIMUMAB patients
- Tier 2 (Data): Full refresh still running (started 15:XX) - pending final verification
### Files changed:
- `data_processing/load_snomed_mapping.py` — fixed clean_snomed_code() for scientific notation
- `data_processing/diagnosis_lookup.py` — changed to use PseudoNHSNoLinked for GP lookup
- `IMPLEMENTATION_PLAN.md` — marked Task 3.3 as in progress
### Committed: 5b1569e "fix: correct patient identifier for GP diagnosis lookup (Task 3.3)"
### Patterns discovered:
- **Critical**: PersonKey ≠ PatientPseudonym. HCD data has two patient identifiers:
  - `LocalPatientID` (aliased as PersonKey) — provider-specific, NOT in GP records
  - `PseudoNHSNoLinked` — pseudonymised NHS number, matches `PatientPseudonym` in GP records
- SNOMED codes can have 15-16 digits, causing float precision issues in pandas/Excel exports
- Scientific notation must be converted back to integers for string matching
### Next iteration should:
1. **Check refresh completion**: Read output from task be9b9e7
   - Look for "DIAGNOSIS matches: X%" line in batch lookup output
   - Should now show non-zero percentage (expected 10-30% based on ADALIMUMAB test)
   - Look for "indication: X nodes total" confirming indication charts generated

2. **If refresh succeeded**: Verify database state
   - `SELECT chart_type, COUNT(*) FROM pathway_nodes GROUP BY chart_type`
   - Should show both "directory" (293) and "indication" (expected 300-600) rows
   - `SELECT DISTINCT directory FROM pathway_nodes WHERE chart_type='indication' LIMIT 20`
   - Should show Search_Term values like "rheumatoid arthritis", "macular degeneration"

3. **Mark Task 3.3 complete** with validation evidence:
   - Processing time
   - Record counts per chart type
   - Coverage percentage (diagnosis vs fallback)

4. **If refresh still running**: Wait or check `tail -50` of output file

5. **Start Phase 4**: If 3.3 passes, begin Task 4.1 (Add Chart Type State to Reflex)
### Blocked items:
- None (Snowflake connection established)