Clinical Advisor Review Guide¶

A step-by-step guide for Dr. Shrikanth Naidu (Clinical Advisor) to review and correct the AI's clinical extractions. Your corrections directly improve the AI — every fix makes the system smarter for future patients.

Why Your Reviews Matter¶

When a patient uploads a blood work report, the AI (Clinical Context Agent) extracts: - Conditions (e.g., "Fatty liver", ICD K76.0) - Lab values (e.g., Hemoglobin 13.5 g/dL) - Comorbidities (e.g., pre-diabetes from HbA1c 6.2%)

The AI gets it right ~85% of the time. Your corrections fix the remaining 15% and teach the AI to get it right next time through automated prompt improvement.

Your correction → Feedback store → Pattern detected (3+ similar corrections)
→ New prompt example generated → AI tested → Accuracy improves

How to Review a Case¶

Step 1: Pick a case to review¶

Go to the API docs: services.curaway.ai/docs

Or use curl to list cases with EHR data:

curl -s "https://services.curaway.ai/api/v1/cases" \
  -H "X-Tenant-ID: tenant-apollo-001" \
  -H "X-Patient-ID: demo-patient" | python3 -m json.tool

Step 2: View the AI's extraction¶

Get the EHR for a specific case:

curl -s "https://services.curaway.ai/api/v1/cases/{CASE_ID}/ehr" \
  -H "X-Tenant-ID: tenant-apollo-001" | python3 -m json.tool

This returns:

{
  "medical_history": {
    "conditions": [
      {"name": "Fatty (change of) liver, not elsewhere classified", "icd10": "K76.0", "source": "agent"},
      {"name": "Bradycardia, unspecified", "icd10": "R00.1", "source": "agent"},
      {"name": "Impaired fasting glucose", "icd10": "R73.01", "source": "agent"}
    ],
    "observations": [
      {"parameter": "Hemoglobin", "value": 13.5, "unit": "g/dL"},
      {"parameter": "HbA1c", "value": 5.8, "unit": "%"},
      ...
    ]
  },
  "comorbidities": ["fatty_liver", "bradycardia", "impaired_glucose"]
}

Step 3: Review against the source document¶

View the uploaded documents:

curl -s "https://services.curaway.ai/api/v1/cases/{CASE_ID}/document-checklist" \
  -H "X-Tenant-ID: tenant-apollo-001" | python3 -m json.tool

Step 4: Submit your corrections¶

Use the provider feedback endpoint:

curl -X POST "https://services.curaway.ai/api/v1/cases/{CASE_ID}/provider-feedback" \
  -H "X-Tenant-ID: tenant-apollo-001" \
  -H "Content-Type: application/json" \
  -d '{
    "provider_id": "clinical-advisor",
    "reviewer_name": "Dr. Shrikanth Naidu",
    "conditions_confirmed": ["K76.0", "R00.1"],
    "conditions_added": [
      {"icd10": "E11.9", "name": "Type 2 Diabetes Mellitus", "note": "HbA1c 6.2% indicates pre-diabetic/diabetic"}
    ],
    "conditions_removed": [],
    "observations_corrected": [
      {"parameter": "HbA1c", "ai_value": 5.8, "correct_value": 6.2}
    ],
    "overall_accuracy": 0.85,
    "notes": "Good extraction. Missed pre-diabetic indication from HbA1c trend."
  }'

What to Look For¶

Conditions (ICD-10 codes)¶

Check	What to do
AI correctly identified a condition	Add ICD code to `conditions_confirmed`
AI missed a condition that's in the report	Add to `conditions_added` with ICD code + note
AI identified something that isn't there	Add to `conditions_removed`
AI used wrong ICD code	Add correct code to `conditions_added`, wrong to `conditions_removed`

Common AI Misses¶

Based on the automated review, these are the most common patterns:

Pattern	What the AI misses	What to look for in the report
Pre-diabetes	HbA1c 5.7-6.4% not flagged	HbA1c value in the pre-diabetic range
Mild CKD	eGFR 60-89 not flagged	eGFR in Stage 2 range
Subclinical hypothyroid	TSH 4.5-10 not flagged	Elevated TSH with normal T3/T4
Iron deficiency	Low ferritin with normal Hb	Check ferritin if available
Vitamin D deficiency	25-OH Vitamin D < 20	Often in lab panels but not flagged

Lab Values (Observations)¶

Check	What to do
Value extracted correctly	No action needed
Wrong value (misread from PDF)	Add to `observations_corrected` with `ai_value` and `correct_value`
Missing value (in report but not extracted)	Note in the `notes` field
Wrong unit	Note in `observations_corrected`

Overall Accuracy Rating¶

Rate the AI's extraction on a 1-5 scale:

Rating	Meaning
5	Perfect — all conditions and values correct
4	Good — minor issues (one missed condition or slightly off value)
3	Acceptable — correct on the main diagnosis, missed some comorbidities
2	Poor — missed significant conditions or multiple wrong values
1	Unusable — major errors that would mislead a provider

Review Workflow¶

Quick review (5 minutes per case)¶

Open the EHR (/cases/{id}/ehr)
Scan the conditions list — any obvious misses?
Check HbA1c, eGFR, TSH values specifically (most common misses)
Submit corrections
Move to next case

Thorough review (10 minutes per case)¶

Open the EHR
Open the source document (from document checklist)
Compare every condition against the report
Compare every lab value against the report
Check for conditions implied by lab combinations (e.g., metabolic syndrome)
Submit corrections with detailed notes

Recommended cadence¶

Frequency	Cases	Time	Impact
Daily (10 min)	1-2 cases	10 min	Steady improvement
Weekly batch	5-10 cases	1 hour	Good pattern detection
Monthly deep dive	All new cases	2-3 hours	Comprehensive accuracy tracking

How Your Corrections Improve the AI¶

Week 1: You correct 5 cases — "AI keeps missing pre-diabetes from HbA1c 6.0-6.4"
Week 2: Pattern detector sees 3+ similar corrections
Week 2: System generates a new prompt example:
         "When HbA1c is 6.0-6.4, flag pre-diabetic state (R73.03)"
Week 3: New prompt A/B tested against old prompt
Week 3: Extraction accuracy for pre-diabetes: 40% → 85%
Week 4: Promoted to production — AI now catches pre-diabetes reliably

Tracking your impact¶

Check the eval dashboard:

curl -s "https://services.curaway.ai/api/v1/internal/eval/summary" \
  -H "X-Tenant-ID: tenant-apollo-001" | python3 -m json.tool

This shows: - Total corrections submitted - Average accuracy score - Corrections applied to prompts - Corrections pending review

Quick Reference¶

Endpoints¶

Action	Method	URL
List cases	GET	`/api/v1/cases`
View EHR	GET	`/api/v1/cases/{id}/ehr`
View documents	GET	`/api/v1/cases/{id}/document-checklist`
Submit corrections	POST	`/api/v1/cases/{id}/provider-feedback`
View auto-review	POST	`/api/v1/cases/{id}/auto-review`
Eval summary	GET	`/api/v1/internal/eval/summary`

Required Headers¶

X-Tenant-ID: tenant-apollo-001
Content-Type: application/json

Common ICD-10 Codes¶

Code	Condition
M17.11	Primary osteoarthritis, right knee
M17.12	Primary osteoarthritis, left knee
E11.9	Type 2 diabetes mellitus without complications
R73.03	Prediabetes
K76.0	Fatty liver, not elsewhere classified
R00.1	Bradycardia, unspecified
E78.5	Hyperlipidemia, unspecified
N18.3	Chronic kidney disease, stage 3
E03.9	Hypothyroidism, unspecified
D50.9	Iron deficiency anemia, unspecified
R73.01	Impaired fasting glucose
E78.0	Pure hypercholesterolemia

Questions?¶

Contact SD (Srikanth Donthi) — this guide will be updated as the review process evolves.