NHI CQL Generator

Generate CQL Library (.cql) and FHIR ValueSet (.json) files from Taiwan NHI reimbursement rules text.

Overview

This skill generates standardized clinical decision support artifacts for Taiwan's National Health Insurance system:

1. CQL Library - Clinical Quality Language for payment審查邏輯
2. FHIR ValueSet - Code sets (適應症, 禁忌症, 藥品, etc.)
3. Dependency Libraries - Required CQL libraries that must accompany every generated CQL file

Critical: All patterns are learned from 50+ GT examples. Do NOT deviate from documented patterns.

Required Dependency Libraries

Every generated CQL file depends on two shared libraries located in .github/skills/nhi-cql-generator/assets/:

1. FHIRHelpers.cql (version 4.0.1)

   • Provides FHIR data model interaction helpers
   • Converts FHIR types to CQL types
   • Required by all FHIR-based CQL libraries

2. CDSConnectCommonsForFHIRv401.cql (version 2.1.0, aliased as C3F)

   • Provides common clinical decision support functions that can be used directly
   • Frequently used functions:
     • C3F.Confirmed(Conditions) - Filter confirmed conditions
     • C3F.Verified(Observations) - Filter verified observations
     • C3F.MostRecent(Observations) - Get most recent observation
     • C3F.QuantityValue(Observation) - Extract quantity value
     • C3F.ObservationLookBack(Observations, LookBack) - Observations in time window
     • C3F.MostRecentProcedure(Procedures) - Get most recent procedure
     • C3F.ActiveCondition(Conditions) - Filter active conditions
     • C3F.PeriodToInterval(Period) - Convert FHIR Period to CQL Interval
   • Usage: Directly use C3F functions with the C3F. prefix (e.g., C3F.Confirmed([Condition: "code"]))
   • Note: GT examples redundantly redefine these functions, but this is unnecessary - you should use C3F functions directly

IMPORTANT: When generating CQL artifacts, you MUST copy these two files to the output directory alongside the generated .cql files. Without them, the CQL will fail with error:

Could not load source for library CDSConnectCommonsForFHIRv401, version 2.1.0, namespace uri null.

CQL Structure Patterns

Library Declaration

library "[規則編號]" version '0.0.1'

• Library name: Exact rule code (e.g., "08111B", "12216C", "69041B")
• Version: Always '0.0.1'

Standard Headers (Fixed Order)

library "[規則編號]" version '0.0.1'

using FHIR version '4.0.1'

include "FHIRHelpers" version '4.0.1' called FHIRHelpers
include "CDSConnectCommonsForFHIRv401" version '2.1.0' called C3F

Never change: FHIR version, FHIRHelpers version, C3F version, or aliases.

ValueSet Declarations

After include statements, before codesystem declarations:

valueset "[中文名稱] valueset": '[ValueSet URL]'

URL Pattern: https://example.org/fhir/ValueSet/2.16.840.1.113762.1.4.1287.[sequential-number]

Examples:

• valueset "17022B 呼氣一氧化氮監測(FeNO) valueset": 'https://example.org/fhir/ValueSet/2.16.840.1.113762.1.4.1287.6'
• valueset "糖尿病合併慢性腎病變 valueset": 'https://example.org/fhir/ValueSet/2.16.840.1.113762.1.4.1287.46'

CodeSystem URLs (Complete Approved List)

Use ONLY these codesystem URLs:

codesystem "ICD-10-CM": 'http://hl7.org/fhir/sid/icd-10-cm'
codesystem "ICD-10-PCS": 'http://www.cms.gov/Medicare/Coding/ICD10'
codesystem "LOINC": 'http://loinc.org'
codesystem "SCT": 'http://snomed.info/sct'
codesystem "TWMedicalServicePayment": 'https://twcore.mohw.gov.tw/ig/twcore/CodeSystem/medical-service-payment-tw'
codesystem "CONDVERSTATUS": 'http://terminology.hl7.org/CodeSystem/condition-ver-status'
codesystem "CONDCLINSTATUS": 'http://terminology.hl7.org/CodeSystem/condition-clinical'
codesystem "ActCode": 'http://terminology.hl7.org/CodeSystem/v3-ActCode'

Context Declaration

context Patient

Always use Patient context.

Define Statements Pattern

Standard structure observed in all examples:

define "[適應症或條件名稱]":
  exists ( C3F.Confirmed([Condition: "[valueset名稱]"]) )

define "MeetsInclusionCriteria":
  "[主要適應症]"
  
define "InPopulation":
  "MeetsInclusionCriteria"

define "Recommendation":
  if "InPopulation" then '健保給付[點數]點' 
    else null

define "Rationale":
  null

define "Links":
  null

define "Suggestions":
  null

define "Errors":
  null

Using C3F Helper Functions

Direct Usage - Use C3F library functions directly without redefining them:

// ✅ Correct: Use C3F functions directly
define "適應症":
  exists ( C3F.Confirmed([Condition: "適應症 valueset"]) )

define "Recent Observations":
  C3F.ObservationLookBack([Observation: "code"], 1 year)

define "Lab Value":
  C3F.QuantityValue( C3F.MostRecent([Observation: "code"]) )

Common C3F Functions:

• C3F.Confirmed(Conditions) - Filters to confirmed conditions (verificationStatus = 'confirmed')
• C3F.Verified(Observations) - Filters to verified observations (status = 'final', 'corrected', or 'amended')
• C3F.ObservationLookBack(Observations, LookBack) - Observations within time window
• C3F.MostRecent(Observations) - Most recent observation by effective date
• C3F.QuantityValue(Observation) - Extract quantity value from observation
• C3F.MostRecentProcedure(Procedures) - Most recent procedure
• C3F.ActiveCondition(Conditions) - Active conditions (clinicalStatus = 'active', no abatement)
• C3F.PeriodToInterval(Period) - Convert FHIR Period to CQL Interval

Note: GT examples redundantly redefine these functions at the end of each file, but this is unnecessary duplication. Simply use C3F.FunctionName() directly.

Common Logic Patterns

Condition Checking

exists ( C3F.Confirmed([Condition: "[valueset名稱]"]) )

Multiple Conditions (OR)

exists ( C3F.Confirmed([Condition: "valueset1"])
    union C3F.Confirmed([Condition: "valueset2"])
)

Multiple Conditions (AND)

"condition1" and "condition2"

Observation with Time Window

Count(C3F.ObservationLookBack([Observation: "[code]"], 1 year)) < 3

Age Checking

AgeInYears() >= 6 and AgeInYears() < 13

Encounter with Specialty

[Encounter: "[valueset]"] E
  where ( E.serviceType ~ "[specialty code]" )

Procedure Lookback (using C3F)

C3F.ProcedureLookBack([Procedure: "[code]"], 6 months)

Nested Conditionals (for複雜給付條件)

define "Recommendation":
  if "condition1" then if "condition2" then '說明A'
    else '說明B'
  else if "condition3" then '說明C'
  else null

Subpopulations Pattern

For rules with multiple適應症 with different頻率:

define "subpopulation 1":
  if "InPopulation" is not true then null 
    else [specific condition for subpop1]

define "subpopulation 2":
  if "InPopulation" is not true then null 
    else [specific condition for subpop2]
    
define "Recommendation":
  if "subpopulation 1" then '健保給付[點數]點，[規範1]'
  else if "subpopulation 2" then '健保給付[點數]點，[規範2]'
  else null

Medical Knowledge Mapping

Spinal Disorders Coding

When rules mention "脊髓損傷或病變", include ALL three categories:

Category	ICD-10-CM	Examples
Trauma	S14., S24., S34., S44.	Spinal cord injury
Disease	G95., G82.	Spinal cord disorders, paraplegia
Degeneration	M43.3, M47.1, M50.0, M51.0	Spondylolisthesis, myelopathy, disc disorders

Critical: "病變" includes both functional disease (G-codes) AND degenerative structural changes (M-codes).

Stroke Sequelae Coding

For "腦中風後遺症" or "中風導致之肢體偏癱":

Scope	ICD-10-CM	Purpose
General stroke	I69.*	All stroke sequelae
Hemiplegia types	I69.33-I69.35	Dominant/non-dominant/unspecified side
Specific sequelae	I69.3, I69.33, I69.34, I69.35, I69.39	Cerebral infarction aftereffects

Pattern: Use I69.* for broad inclusion, add specific I69.3X codes when hemiplegia is explicitly mentioned.

Peripheral Neuropathy

For "周邊神經病變":

• Diabetic: E11.4* (Type 2 DM with neurological complications)
• General: G62.* (polyneuropathy), G63.* (neuropathy in diseases)
• Specific: G62.9 (unspecified polyneuropathy)

Observation Value Checking Pattern

For laboratory or measurement constraints:

define "低密度脂蛋白膽固醇異常":
  exists (
    [Observation: "LOINC code 18262-6"] O
      where O.value as Quantity > 100 'mg/dL'
        and O.status = 'final'
  )

Complex Age Range Checks

For rules specifying age ranges with measurement timing:

define "符合年齡條件":
  ( AgeInYears() >= 10 and AgeInYears() <= 17 )
    or ( AgeInYears() >= 6 and AgeInYears() < 10 and "已測量過基線值" )

Code Style Rules

• Indentation: 2 spaces (NO tabs)
• String quotes: Single quotes '...' for all strings
• Line breaks: Blank line between major sections (valuesets, codesystems, context, defines)
• Comments: Use // sparingly, written in Chinese when explaining業務邏輯
• Naming: Chinese variable names permitted (e.g., define "氣喘":, define "懷孕糖尿病人、妊娠糖尿病人":)
• Boolean checks: Use "variable" not "variable" = true

NHI Rule Text Parsing

Rule ID Format

Pattern: [數字]{2,5}[字母] (e.g., 08111B, 12216C, 69041B, 83102K)

Common Phrases to Extract

適應症 (Indications):

• "適應症：", "適應症包括："
• "限...使用", "限...申報"
• "適用於...", "符合..."

禁忌症 (Contraindications):

• "禁忌症：", "不得用於..."
• "排除...", "不適用"

支付規範 (Payment Rules):

• "支付規範：", "執行頻率："
• "每[時間]限申報[次數]次"
• "一年限申報[數字]次"
• "不得同時申報..."

點數 (Points):

• "[數字]點" (e.g., "800點", "17350點")

專科限制 (Specialty):

• "限...專科醫師", "限由...執行"

Diagnosis Code Extraction

From text patterns like:

• "多發性硬化症、運動神經元疾病、脊髓損傷..." → Create valueset for ICD-10-CM codes
• "四肢自體動脈粥樣硬化（I70.2～I70.7）" → Code range with all subcodes
• "糖尿病酮酸中毒" → Look up ICD-10-CM code

Frequency Constraints Encoding

CRITICAL: Always filter by the specific NHI procedure code using TWMedicalServicePayment codesystem. NEVER use bare [Procedure] (this would match ALL procedures, not the one being regulated).

Required: Declare procedure code at top of file

codesystem "TWMedicalServicePayment": 'https://twcore.mohw.gov.tw/ig/twcore/CodeSystem/medical-service-payment-tw'

code "[代號] 健保給付代碼": '[代號]' from "TWMedicalServicePayment" display '[項目名稱]'

Frequency check pattern

// ✅ Correct: counts only THIS specific procedure
define "頻率符合":
  Count(C3F.ProcedureLookBack([Procedure: "[代號] 健保給付代碼"], 1 year)) < N

// ❌ WRONG: counts ALL procedures in the patient's history
define "頻率符合":
  Count(C3F.ProcedureLookBack([Procedure], 1 year)) < N

Rule Text	CQL Encoding
"一年限申報三次"	Count(ObservationLookBack([Observation: ...], 1 year)) < 3
"每四年限申報一次"	Count(ObservationLookBack([Observation: ...], 4 years)) < 1
"每人以申報一次為原則"	Track across patient lifetime
"每次就診...每次治療應間隔至少一週"	Date diff checking
"一年限申報三次"	Count(C3F.ProcedureLookBack([Procedure: "[代號] 健保給付代碼"], 1 year)) < 3
"每四年限申報一次"	Count(C3F.ProcedureLookBack([Procedure: "[代號] 健保給付代碼"], 4 years)) < 1
"每人以申報一次為原則"	not exists(C3F.ProcedureLookBack([Procedure: "[代號] 健保給付代碼"], 100 years))
"每週最多N次"	Count(C3F.ProcedureLookBack([Procedure: "[代號] 健保給付代碼"], 7 days)) < N
"每月限一次"	Count(C3F.ProcedureLookBack([Procedure: "[代號] 健保給付代碼"], 30 days)) < 1

Per-Encounter / Per-Hospitalization check

When the rule says "限住院期間申報一次" or "同一就診申報一次", use Encounter reference instead of a time window. A time window (e.g., "1 year") would incorrectly block re-admission for a new event.

// 取得目前就醫 Encounter（最近一次進行中或完成的）
define "目前住院Encounter":
  Last(
    [Encounter] E
      where E.status = 'in-progress' or E.status = 'finished'
      sort by start of period
  )

// 本次住院期間尚未申報過本項目
define "本次住院未申報過":
  "目前住院Encounter" is not null
    and not exists(
      [Procedure: "[代號] 健保給付代碼"] P
        where P.encounter.reference = 'Encounter/' + "目前住院Encounter".id
    )

"同一治療期間" (Episode of Care) check

When rule says "同一治療期間至多N次", the ideal boundary is EpisodeOfCare. As a pragmatic approximation, use Encounter-reference for the per-visit part, and ProcedureLookBack with a reasonable lookback window for the total-count part:

define "本次就診未申報":
  "目前住院Encounter" is not null
    and not exists(
      [Procedure: "[代號] 健保給付代碼"] P
        where P.encounter.reference = 'Encounter/' + "目前住院Encounter".id
    )

// 治療期間總次數：以 1 年作為最長治療期間的保守估計
define "治療期間總次數符合":
  Count(C3F.ProcedureLookBack([Procedure: "[代號] 健保給付代碼"], 1 year)) < N

Note: The ideal approach for "同一治療期間" is to use EpisodeOfCare, but this requires the FHIR server to populate EpisodeOfCare resources. The time-window approximation is the practical fallback.

Other Regulation Handling (其他規範處理)

Critical: Distinguish between pre-authorization criteria (事前資格檢查) and post-payment rules (事後支付規則).

Type 1: Pre-Authorization Criteria → Convert to CQL Logic

These MUST be encoded as executable logic in MeetsInclusionCriteria:

Pattern: Contraindication ("禁忌症：...")

When a rule lists contraindications, create a separate contraindication ValueSet and check for absence:

// 1. Declare a separate contraindication valueset
valueset "[代號] 禁忌症": 'https://example.org/fhir/ValueSet/2.16.840.1.113762.1.4.1287.[n]'

// 2. Define the contraindication check
define "有禁忌症":
  exists ( C3F.Confirmed([Condition: "[代號] 禁忌症"]) )

// 3. Incorporate into MeetsInclusionCriteria
define "MeetsInclusionCriteria":
  "Has Indication" and not "有禁忌症"

// 4. Recommendation branches (always check indication first → then contraindication)
define "Recommendation":
  if not "Has Indication" then null
    else if "有禁忌症" then '不得申報。病患符合禁忌症：[詳細原因]。'
    else '健保給付。...'

Important: The contraindication JSON file uses the same structure as the indication ValueSet. Give it a distinct OID number (e.g., indication = .9, contraindication = .200).

Pattern: Specialist Restriction ("限...專科醫師執行")

codesystem "SCT": 'http://snomed.info/sct'

code "Specialist code": '[SNOMED-code]' from "SCT" display '[specialist-type]'

define "performer_id":
  Last(Split(First("RecentProcedure".performer).actor.reference, '/'))

define "Performer_ref":
  [Practitioner] P where P.id = "performer_id"

define "Qualified_Practitioner":
  exists ( "Performer_ref".qualification q
    where q.code ~ "Specialist code" )

define "MeetsInclusionCriteria":
  "[適應症]" and "Qualified_Practitioner"

Examples:

• "限外科專科醫師執行" → SNOMED CT: 224555008 (Vascular surgeon)
• "限兒科專科" → SNOMED CT: 24251000087109 (General pediatric specialty)
• "限心臟血管外科" → SNOMED CT: 309296001 (Cardiac surgeon)

Pattern: Procedure Type Restriction

When rule specifies specific surgical procedures:

codesystem "ICD-10-PCS": 'http://www.cms.gov/Medicare/Coding/ICD10'

code "Procedure1 code": '[code]' from "ICD-10-PCS" display '[description]'
code "Procedure2 code": '[code]' from "ICD-10-PCS" display '[description]'

define "ApplicableProcedure":
  C3F.MostRecentProcedure ( [Procedure: "Procedure1 code"]
    union [Procedure: "Procedure2 code"]
  )

define "MeetsInclusionCriteria":
  "[適應症]" and exists("ApplicableProcedure")

Pattern: Encounter Specialty Check

For rules limiting to specific clinical departments:

code "Department specialty code": '[SNOMED-code]' from "SCT" display '[specialty]'

define "SpecialtyEncounter":
  [Encounter] E where E.serviceType ~ "Department specialty code"

define "encounter check":
  exists("SpecialtyEncounter")

define "Recommendation":
  if "InPopulation" is not true then null
    else if "encounter check" then '給付'
    else '非[科別]申報，不予給付'

Type 2: Post-Payment Rules → Include in Recommendation Text Only

These should appear in Recommendation string but NOT in MeetsInclusionCriteria:

Examples:

• "一般材料費得另加計百分之五十" → Include in Recommendation only
• "應事先申請" → Include in Recommendation only
• "需檢附...報告" → Include in Recommendation only
• "不得同時申報..." → Include in Recommendation as note

Pattern:

define "Recommendation":
  if "InPopulation" then '健保給付[點數]點，[描述]，一般材料費得另加計百分之五十'
  else null

Decision Logic: When to Convert to CQL vs Text Only

Phrase Pattern	Type	Action
限...專科醫師執行	Pre-auth	✅ Convert to Practitioner.qualification check
限...科別申報	Pre-auth	✅ Convert to Encounter.serviceType check
限特定手術類型	Pre-auth	✅ Convert to Procedure code check
材料費加計...%	Post-payment	❌ Text only in Recommendation
應事先申請	Administrative	❌ Text only in Recommendation
需檢附報告	Administrative	❌ Text only in Recommendation

Critical Rule: If the regulation affects whether payment is approved, it's pre-authorization → convert to CQL. If it only affects how much or when to pay, it's post-payment → text only.

ValueSet Generation

ValueSet JSON Template

Based on GT examples, use this complete structure:

{
  "resourceType": "ValueSet",
  "id": "2.16.840.1.113762.1.4.1287.[number]",
  "meta": {
    "versionId": "1",
    "lastUpdated": "[timestamp]",
    "profile": [
      "http://hl7.org/fhir/StructureDefinition/shareablevalueset",
      "http://hl7.org/fhir/us/cqfmeasures/StructureDefinition/computable-valueset-cqfm",
      "http://hl7.org/fhir/us/cqfmeasures/StructureDefinition/publishable-valueset-cqfm"
    ]
  },
  "extension": [
    {
      "url": "http://hl7.org/fhir/StructureDefinition/valueset-author",
      "valueContactDetail": {
        "name": "Industrial Technology Research Institute Author"
      }
    },
    {
      "url": "http://hl7.org/fhir/StructureDefinition/resource-lastReviewDate",
      "valueDate": "[YYYY-MM-DD]"
    },
    {
      "url": "http://hl7.org/fhir/StructureDefinition/valueset-effectiveDate",
      "valueDate": "[YYYY-MM-DD]"
    }
  ],
  "url": "http://example.org/fhir/ValueSet/2.16.840.1.113762.1.4.1287.[number]",
  "identifier": [
    {
      "system": "urn:ietf:rfc:3986",
      "value": "urn:oid:2.16.840.1.113762.1.4.1287.[number]"
    }
  ],
  "version": "[YYYYMMDD]",
  "name": "[EnglishCamelCaseName]",
  "title": "[English title or Chinese title]",
  "status": "active",
  "date": "[timestamp]",
  "publisher": "Industrial Technology Research Institute Steward",
  "description": "[規則編號，可多個用逗號分隔如 63007B, 63008B]",
  "jurisdiction": [
    {
      "extension": [
        {
          "url": "http://hl7.org/fhir/StructureDefinition/data-absent-reason",
          "valueCode": "unknown"
        }
      ]
    }
  ],
  "purpose": "(Clinical Focus: ...),(Data Element Scope: ...),(Inclusion Criteria: ...),(Exclusion Criteria: ...)",
  "compose": {
    "include": [
      {
        "system": "[CodeSystem URL]",
        "version": "2025",
        "filter": [
          {
            "property": "concept",
            "op": "[operator]",
            "value": "[code]"
          }
        ]
      }
    ]
  }
}

ID Naming Convention

Pattern: 2.16.840.1.113762.1.4.1287.[sequential-number]

The ID is the full OID, not a custom name. File name matches the ID:

• File: 2.16.840.1.113762.1.4.1287.6.json
• ID: "2.16.840.1.113762.1.4.1287.6"

URL Pattern

Critical: Use http:// (NOT https://) in url field:

"url": "http://example.org/fhir/ValueSet/2.16.840.1.113762.1.4.1287.[number]"

In CQL: Use https:// in valueset declarations:

valueset "name": 'https://example.org/fhir/ValueSet/2.16.840.1.113762.1.4.1287.[number]'

Filter Operators (from GT examples)

Use filter instead of explicit concept array:

Common operators:

• "op": "descendent-of" - Include code and all descendants (most common)
• "op": "is-a" - Include code and descendants
• "op": "=" - Exact match single code
• "op": "in" - Multiple codes (comma-separated in value)

Examples:

// All J45.* asthma codes
{"property": "concept", "op": "descendent-of", "value": "J45"}

// Specific codes
{"property": "concept", "op": "=", "value": "C79.81"}

// Multiple specific codes
{"property": "concept", "op": "in", "value": "E10.22,E13.22,E08.22"}

Filter Operator Selection Strategy

Choose operator based on code specificity and clinical intent:

Operator	Use When	ICD-10 Code Pattern	Example
=	Exact leaf code needed	6-7 digits (e.g., G62.9, I69.398)	Single specific diagnosis
is-a	Mid-level category	4-5 digits (e.g., J45.5, E11.4)	Specific subtype with few children
descendent-of	Broad family inclusion	3-4 digits (e.g., J45, I69, G62)	All variants needed
in	Multiple unrelated codes	Mixed	Enumerate disparate codes

Code Granularity Guidelines:

• Preferred: Mid-level codes (5-6 digits) - balance precision and coverage
• Avoid: Too broad (3 digits: J45 vs J45.*) or too narrow (7+ digits)
• Quick judgment: Code length 3-4 → descendent-of; 5-6 → is-a; 7+ → =

Real Examples from GT:

// G62.9 (7 digits, leaf code) - use exact match
{"property": "concept", "op": "=", "value": "G62.9"}

// E11.4 (5 digits, mid-level) - use is-a
{"property": "concept", "op": "is-a", "value": "E11.4"}

// I69 (3 digits, broad family) - use descendent-of
{"property": "concept", "op": "descendent-of", "value": "I69"}

// Multiple unrelated codes - use in
{"property": "concept", "op": "in", "value": "M43.3,M47.1,M50.0,M51.0"}

Decision Logic:

1. If code explicitly ends with .0, .9, or has 6+ digits → =
2. If code has 4-5 digits and represents specific subtype → is-a
3. If code is 3-4 digits and needs all variants → descendent-of
4. If multiple disparate codes listed → in

Purpose Field Format

Standard 4-part structure (observed in all GT examples):

(Clinical Focus: [臨床焦點]),(Data Element Scope: [資料元素範圍]),(Inclusion Criteria: [納入條件]),(Exclusion Criteria: [排除條件，若無則寫 No])

Critical: ValueSet.url in JSON must exactly match the valueset declaration URL in CQL (except http vs https).

Common ValueSet Types

1. Diagnosis codes (ICD-10-CM) - Most common
   • System: http://hl7.org/fhir/sid/icd-10-cm
   • Version: "2025"
   • Use filters, not explicit concept lists
2. Procedure codes (ICD-10-PCS or TWMedicalServicePayment)
   • System: http://www.cms.gov/Medicare/Coding/ICD10 or https://twcore.mohw.gov.tw/ig/twcore/CodeSystem/medical-service-payment-tw
3. Lab tests (LOINC)
   • System: http://loinc.org
4. Clinical findings (SNOMED CT)
   • System: http://snomed.info/sct

Multiple Rules Sharing Same ValueSet

From GT examples, one ValueSet may support multiple rules:

• "description": "63007B, 63008B" - Modified radical mastectomy (single/bilateral)
• "description": "13030B,13031B,13032B" - Different H. pylori tests

When rules share indications, use the same ValueSet ID.

Workflow

Step 1: Collect Input

Ask user for:

• 規則編號 (Rule ID) - e.g., "08111B"
• 完整給付規則文字 (Full rule text including適應症, 禁忌症, 支付規範, 點數)

Step 2: Parse Rule Text

Extract and identify:

• Rule ID
• 點數 (reimbursement points)
• 適應症 (indications) - list of diagnoses/conditions
• 禁忌症 (contraindications) - if any
• 執行頻率限制 (frequency limits)
• 不得同時申報項目 (cannot claim with)
• 專科限制 (specialty restrictions)
• 其他支付規範 (other payment rules)

Step 3: Generate ValueSets

For each distinct code set identified:

1. Create JSON file: 2.16.840.1.113762.1.4.1287.[number].json
2. Use complete template structure from GT examples
3. Assign sequential OID number
4. Set url with http:// (NOT https)
5. Populate filter operators based on code patterns:
   • Use descendent-of for code families (e.g., J45 for all asthma)
   • Use in for multiple specific codes
   • Use = for single specific codes
6. Fill purpose field with 4-part structure
7. Save with UTF-8 encoding

Important:

• File name = full OID (e.g., 2.16.840.1.113762.1.4.1287.6.json)
• url field uses http://
• CQL valueset declaration uses https://

Step 4: Generate CQL Library

Following exact structure:

1. Library declaration with rule ID
2. using FHIR version
3. include statements (FHIRHelpers, C3F)
4. valueset declarations (URLs must match JSON files)
5. codesystem declarations
6. Standard code/concept for Confirmed
7. context Patient
8. define statements (適應症, conditions, InPopulation, Recommendation, etc.)
9. Standard define statements (Rationale, Links, Suggestions, Errors as null)
10. All five helper functions

Step 5: Validate Generated Files

Checklist:

• ✓ CQL valueset declaration URLs exactly match ValueSet JSON url fields
• ✓ Only codesystem URLs from approved list used
• ✓ All five helper functions included at end
• ✓ Structure follows examples (headers, blank lines, 2-space indentation)
• ✓ Recommendation includes 點數 and key給付規範
• ✓ File encoding is UTF-8
• ✓ No tabs, only spaces

Step 6: Provide Output

Deliver to user:

1. CQL file: [規則編號].cql (plain text)
2. ValueSet file(s): One or more 2.16.840.1.113762.1.4.1287.[number].json files
3. Dependency libraries: Copy from .github/skills/nhi-cql-generator/assets/ to output directory:
   • FHIRHelpers.cql
   • CDSConnectCommonsForFHIRv401.cql
4. Mapping table:

| CQL ValueSet Declaration | ValueSet File | URL (CQL uses https, JSON uses http) |
|--------------------------|---------------|---------------------------------------|
| [中文名稱] valueset | 2.16.840.1.113762.1.4.1287.[N].json | https://example.org/fhir/ValueSet/2.16.840.1.113762.1.4.1287.[N] |

5. Validation report: Confirm all patterns followed ✓

Critical: The CQL file will NOT execute without the dependency libraries in the same directory. Always include them in the output.

Examples

Example 1: Simple Rule (08111B)

Input:

• Rule ID: 08111B
• 點數: 800點
• 適應症: 凝血異常，疑有Von-willebrands disease者

Output:

• 08111B.cql - With Von Willebrand disease codes
• 2.16.840.1.113762.1.4.1287.1.json - ICD-10-CM codes with filter for VWD codes

Example 2: Rule with Frequency (17022B)

Input:

• Rule ID: 17022B
• 點數: 748點
• 適應症: 六歲以上至未滿十三歲確診氣喘患者追蹤使用
• 頻率: 一年最多申報三次
• 專科: 兒科

Output:

• 17022B.cql - With age check, frequency check, and pediatric specialty check
• 2.16.840.1.113762.1.4.1287.6.json - Uses descendent-of filter for J45 (asthma)

Example 3: Multiple Subpopulations (12216C)

Input:

• Rule ID: 12216C
• 點數: 900點
• 適應症1: C型肝炎抗體陽性之HIV感染者、靜脈注射藥癮者 (每一年限一次)
• 適應症2: C型肝炎抗體陽性之慢性血液透析病人 (每四年限一次)

Output:

• 12216C.cql - With subpopulation 1 and 2, different frequency rules
• 2.16.840.1.113762.1.4.1287.[N].json - HIV codes with appropriate filters
• 2.16.840.1.113762.1.4.1287.[N+1].json - Dialysis dependency codes

Important Rules

1. Never create new codesystem URLs - Only use the eight approved URLs
2. Never deviate from standard structure - Follow examples exactly
3. Always include all helper functions - Even if not directly used
4. CQL and ValueSet URLs must match - This is absolutely critical
5. Use patterns, don't innovate - Combine existing patterns for new scenarios
6. Chinese in variable names is OK - This is standard practice
7. Preserve exact indentation - 2 spaces, no tabs

Triggering Keywords

Activate when user mentions:

• NHI CQL, 健保 CQL
• 健保給付規則, 給付審查, 藥品審查
• CQL Library, 健保 ValueSet, 健保代碼
• Taiwan health insurance reimbursement
• Generate CQL for [rule code]

Notes

• GT/ folder was used for learning only - not available at runtime
• All knowledge must come from patterns in this SKILL.md
• When uncertain, combine documented patterns rather than improvising
• 中文 and English 混用 is standard in NHI CQL (both in variable names and comments)

Critical ValueSet URL Pattern Difference

In ValueSet JSON (url field):

"url": "http://example.org/fhir/ValueSet/2.16.840.1.113762.1.4.1287.6"

Uses http:// (no 's')

In CQL (valueset declaration):

valueset "name": 'https://example.org/fhir/ValueSet/2.16.840.1.113762.1.4.1287.6'

Uses https:// (with 's')

This is intentional and follows GT examples exactly.

ValueSet File Naming

• File name = Full OID: 2.16.840.1.113762.1.4.1287.6.json
• NOT descriptive names like 17022B-indications.json
• id field in JSON = file name without .json extension