PDF Evidence Table Extraction Workflow Design

Project Goal

Develop a repeatable workflow to extract structured evidence tables from ARS (American Radium Society) and similar clinical guideline PDFs, outputting the data to Excel format.

Initial Analysis (December 20, 2025)

Source Files Analyzed

File	Size	Pages	Description
`2021_ARS_ReRT_Full_2021_09_1.pdf`	1.3 MB	64	Full ARS guideline on head/neck cancer re-irradiation
`ReRT_Evidence_Table.pdf`	121 KB	7	Extracted "Supplemental Table 1: Evidence Table"

Evidence Table Structure

The evidence table contains clinical study data with the following columns:

Column	Data Type	Notes
Reference	Text	Author, Year + superscript reference number
Study Type	Text	RCT, MA, SR, SAT, RMI, RSI
Topic/Objective	Text	Multi-line possible
Disease	Text	Cancer type/site
Arm(s)/Cohort(s)	Text	Treatment descriptions
N	Text/Number	Sample size, sometimes "X studies, Y patients"
Median FU (Mo.)	Text	Follow-up duration
Results	Text	Outcomes - longest field
Study Quality	Number	1-4 scale

Extraction Challenges Identified

Multi-line cells that wrap across rows
Superscript reference numbers embedded in text
Text wrapping across page boundaries
Variable PDF text quality depending on generator
Column alignment lost in raw text extraction

Output Format Decision

Target: Excel (.xlsx)

Rationale:

Easy to review and validate extracted data
Supports filtering and sorting for analysis
Can be further processed or imported to other systems

Approach Options Considered

Option 1: Claude Chat (Current Session)

Upload PDFs to conversation
Claude extracts and parses data
Generate Excel file for download
Limitation: Manual process each time

Option 2: Claude Code (Included with Max Plan)

Command-line tool included with Pro/Max subscription
Can interact with local files
Could script a workflow for PDF → Excel
Uses same usage allocation as chat

Option 3: Claude API (Pay-as-you-go)

Separate account at console.anthropic.com
Pay per token used
Build fully automated C# tool
Additional cost beyond Max subscription

API Cost Analysis

Pricing (per million tokens)

Model	Input	Output	Best For
Haiku 3	$0.25	$1.25	Simple tasks, high volume
Sonnet 4.5	$3.00	$15.00	Balanced, good for parsing
Opus 4.5	$5.00	$25.00	Most capable

Estimated Cost Per Extraction

Based on the sample ARS document (~146,000 characters ≈ 37,000 tokens input):

Model	Input Cost	Output Cost	Total Per Document
Haiku 3	$0.01	$0.006	~$0.02
Sonnet 4.5	$0.11	$0.08	~$0.19
Opus 4.5	$0.19	$0.13	~$0.32

Volume Projections

Volume	Haiku	Sonnet	Opus
10 documents	$0.20	$1.90	$3.20
100 documents	$2.00	$19.00	$32.00
1,000 documents	$20.00	$190.00	$320.00

Recommendation: Sonnet 4.5 (~$0.19/document) offers the best balance of capability and cost for complex medical table parsing.

Important Note: Max Plan vs API

The Claude Max subscription ($100-200/month) does not include API access. The API requires a separate Console account with pay-per-token billing. These are independent systems:

Max Plan: Fixed monthly fee, chat-based access via claude.ai and Claude Desktop
API: Pay-per-token, requires separate account at console.anthropic.com

Next Steps

Current Phase: Sample Collection

Before building the extraction workflow, collect 50+ sample PDF files to:

Categorize formats - Group similar table structures together
Identify common patterns - What's consistent across all files
Document variations - What differs and how to handle each case
Estimate complexity - Some may need simple extraction, others may need AI parsing
Refine cost estimates - Based on actual file sizes and complexity
Design flexible workflow - That handles variations gracefully

Questions to Answer During Collection

Are these all from ARS, or multiple organizations?
Are they all "Evidence Tables" or other table types too?
Date range of publications? (Older PDFs may have different formatting)
Are they all text-based, or some image-based (requiring OCR)?

Proposed Workflow (Pending Sample Analysis)

┌─────────────────┐     ┌──────────────────┐     ┌─────────────────┐
│  Source PDF     │────▶│  Text Extraction │────▶│  AI Parsing     │
│  (Evidence      │     │  (Python/C#)     │     │  (Claude API)   │
│   Table)        │     │                  │     │                 │
└─────────────────┘     └──────────────────┘     └────────┬────────┘
                                                          │
                                                          ▼
┌─────────────────┐     ┌──────────────────┐     ┌─────────────────┐
│  Validation     │◀────│  Excel Output    │◀────│  Structured     │
│  & Review       │     │  (.xlsx)         │     │  JSON Data      │
└─────────────────┘     └──────────────────┘     └─────────────────┘

Decision Point

Action Required: Collect 50+ sample PDFs, then resume design work.

When samples are ready, Claude can:

Analyze format variations across all samples
Recommend extraction strategy
Build proof-of-concept
Develop production workflow

Document Created: December 20, 2025
Last Updated: December 20, 2025