DiSMed

DiSMed performs de-identification of Spanish medical texts by combining named entity recognition and entity randomization to anonymize patient-identifying information for research use.

Key Features:

• Named Entity Recognition (NER): Tailored to Spanish medical texts and identifies and categorizes sensitive entities such as patient names, locations, and other identifiable information.
• Neural Network Integration: Utilizes four distinct neural network architectures tested on a dataset of radiology reports, achieving a 97.18% recall for entity identification.
• Randomization Algorithm: Substitutes identified entities with new ones from the same category so that real and synthetic data are indistinguishable, reducing re-identification risk.
• Cross-Language Applicability: Methodology is translatable to other languages for international medical text de-identification.
• Efficiency with Limited Data: Performs effectively without requiring an extensive training corpus, enabling extension to various languages and medical text types including EHRs.
• Validation: Validated on proprietary radiology reports and the MEDDOCAN challenge dataset for EHRs, with recall rates of 97.18% (radiology) and 69.18% (MEDDOCAN).

Scientific Applications:

• Radiology report de-identification: Applied and validated on proprietary radiology reports with 97.18% recall for entity detection.
• EHR de-identification: Evaluated on the MEDDOCAN challenge dataset for electronic health records with 69.18% recall.
• Creation of de-identified corpora for research: Facilitates secondary use of clinical text while protecting patient privacy.

Methodology:

Combines named entity recognition with entity randomization using four neural network architectures integrated with spaCy and TensorFlow 1; implemented in Python with curated datasets available under a research use agreement.