96% accuracy in 8 days: how LabelFort structured 140,000 fields across 20,000 patient records

The client builds healthcare analytics and metabolic risk prediction pipelines that depend on structured clinical outputs - not just raw patient text. Their models require extracted blood pressure values, rule-based BP categories, obesity grouping, metabolic risk counts, and metabolic syndrome flags derived consistently from 20,000 patient records.

LabelFort processed 20,000 patient records and delivered 140,000 validated labeled fields within 8 days at 96% annotation accuracy, using a maker-checker AI workflow across 41 annotators.

The challenge

Healthcare datasets require exceptional precision - even small extraction errors impact downstream clinical risk analysis. This project combined direct extraction tasks (systolic and diastolic values) with rule-based classification tasks (BP category, obesity group, metabolic risk count, metabolic syndrome flag).

The client needed 140,000 outputs held to strict clinical thresholds across a large annotator team.

• Correct extraction of blood pressure values from varied formats such as “132/78”
• Consistent application of clinical rules for BP categories and obesity groups
• Correct prioritization when systolic and diastolic fall into different categories
• Accurate computation of metabolic risk count across multiple abnormal conditions
• Consistent metabolic syndrome flag decisions at scale
• Low error rate across 140,000 outputs under strict clinical thresholds

What we delivered

Each patient record was structured into seven required outputs, combining extraction and derived classifications:

1. Systolic pressure - extract the upper BP value
2. Diastolic pressure - extract the lower BP value
3. BP category - Normal, Elevated, Stage 1, or Stage 2 per clinical thresholds; higher category used when systolic and diastolic disagree
4. Obesity group - Underweight, Normal, Overweight, or Obese based on BMI thresholds
5. Metabolic risk count - total categories outside normal range
6. Metabolic syndrome flag - Yes when metabolic risk count ≥ 2; No when 0 or 1
7. Structured output validation - completeness and consistent formatting across all fields

Our approach

LabelFort ran the engagement as AI-assisted primary extraction followed by maker validation and checker QA review on every record.

• AI-assisted tools generated initial predictions for extraction and derived classifications.
• Makers confirmed correct BP extraction and applied defined rules for categories and risk indicators.
• Checkers ensured no extraction errors, correct conflict-rule category mapping, and uniform formatting.
• Clinical threshold rules were enforced consistently across all 20,000 records.

Results

The engagement delivered a clinical-grade structured dataset on schedule - maker-checker validation minimized misclassification risk across 140,000 outputs.

One reporting note: accuracy is checker-phase field agreement against the clinical threshold guideline, not an inter-annotator agreement (kappa) score.

• 20,000 patient records processed
• 140,000 structured outputs delivered
• 96% annotation accuracy across the validated field set
• 414 hours 27 minutes of total agent effort
• 00:01:15 average handle time per record · 00:00:11 per labeled field
• 41 annotators on a maker-checker AI workflow

Key insights

• Rule-based classification tasks require strict QA to avoid drift in clinical thresholds.
• Combining extraction and derived metrics increases downstream dataset value significantly.
• Maker-checker review ensures reliable categorization when values fall into borderline ranges.
• AI-assisted workflows accelerate delivery while maintaining annotation consistency.

Impact

• Metabolic risk prediction models
• Patient risk segmentation and classification systems
• Healthcare analytics pipelines for BP and obesity grouping
• Early detection systems for metabolic syndrome
• Training and evaluation of clinical decision-support ML models

Key takeaways

• Delivered 140,000 structured outputs across 20,000 patient records in 8 days at 96% annotation accuracy.
• Seven-field framework combined direct extraction with rule-based derived classifications.
• Maker-checker QA held clinical threshold rules consistent across borderline BP and obesity ranges.
• AI-assisted extraction accelerated throughput while checkers protected field-level precision at scale.