Sentedel

Redact what's
private.

Purpose-built models for detecting and redacting PHI in healthcare EDI -- where general-purpose tools fall short.

View Benchmarks Try v2.1 demo
EDI-PHI demo Browser demo
Relaxed Recall (50% Overlap)
100%
Strict Accuracy (Exact Boundary)
91.0%
Performance

PHI Detection on
Synthetic EDI Benchmarks

General-purpose PII models struggle with X12 delimiters and qualifiers. On our synthetic 837P test set (n=500), Sentedel EDI-PHI reaches 91.0% strict boundary recall and 100% relaxed recall (≥50% overlap).

Swipe table →

Model Strict Recall Relaxed Recall (50%) Strict F1
Sentedel EDI-PHI v1
91.0% 100.0% 76.4%
GLiNER PII Base 49.0% 54.2% 50.3%
NVIDIA GLiNER PII 35.9% 40.9% 39.8%
OpenAI Privacy Filter 3.0% 64.4% 1.2%

Relaxed vs strict

On 500 held-out synthetic transactions, relaxed matching finds 11,000 of 11,002 labeled elements (99.98%); strict exact-boundary recall is 91.0%.

Boundary alignment

S-tag label mapping reduced subword boundary errors, improving strict boundary recall from 0.8% (baseline) to 91.0% on the same benchmark.

Training efficiency

Dynamic sequence packing and LoRA cut training time by ~62% compared to our baseline training setup.

Latest Research

Technical Report NTE New

Closing the NTE Gap in EDI PHI Detection

by Sentedel Research June 2026 v2.1

The first model trained to handle PHI detection across both structured EDI segments and embedded NTE free text — achieving 93% NTE recall where competitors manage 29–67%.

93.0% NTE
99.7% relaxed F1
Read the report
Technical Blog Benchmark

PHI Detection for EDI Transactions

by Sentedel Research June 2026 v1

Building the first PHI detection benchmark for payer-side healthcare EDI — and why general-purpose models fail dramatically on structured transaction data.

91.0% STRICT
100% relaxed recall
Read the post