Redacting PHI from Scanned Medical Records & Faxes

Unlike digitally generated PDFs, scanned records are pure pixel grids. Common challenges include: - **Resolution Degradation**: Standard faxes are often transmitted at 200×100 DPI (fine) or 200×200 DPI, far below the 300 DPI ideal for reliable OCR. - **Skew and Distortion**: Pages fed through automatic document feeders frequently rotate 1° to 5°, degrading baseline character recognition. - **Artifacts and Bleed-Through**: Ink stamps, carbon paper bleed, hole punches, and thermal paper discoloration create false positive and false negative character matches.

To maximize detection recall on degraded scans without sending images to third-party cloud vision APIs, PHIPass runs local image preprocessing before OCR: 1. **Adaptive Binarization**: Converts color and grayscale scans to high-contrast binary representations using localized thresholding. 2. **Skew Correction**: Detects document orientation and baseline text angles, rotating the image matrix to a true horizontal axis. 3. **Multi-Scale Tesseract OCR**: Identifies character bounding boxes alongside word-level confidence metrics.

When an identifier (such as a date of birth or physician signature) is confirmed for redaction, the redaction must alter the underlying image array: - The exact coordinate polygon of the finding is mapped back to the original image dimensions. - All pixels within the bounding box are set to a solid mask color (typically solid black or white) in memory. - The output image is re-encoded into the destination document. Forensics cannot recover the obliterated pixels because the data is permanently replaced.

No automated OCR system achieves 100% recall on handwritten marginalia or faint rubber stamps. For this reason, PHIPass incorporates a strict **Human Review Gate**: - Any OCR detection with low character confidence or ambiguous clinical context produces a `REVIEW_REQUIRED` status. - Clinicians or privacy reviewers visually inspect the highlighted image regions on their local screen. - Export is blocked until the reviewer explicitly signs off on ambiguous items or marks uninspected areas as verified.

Healthcare staff can significantly improve automated de-identification accuracy by standardizing ingestion protocols: - Scan documents at **300 DPI minimum** in 24-bit color or 8-bit grayscale. - Ensure scanner glass is free of adhesive residue and dust. - Flatten creased or folded records before feeding.