PHI Detection via Segmented NER and Random Forest
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing models for detecting Protected Health Information (PHI) in data loss prevention fail to effectively incorporate both medical terms and personally identifiable information (PII), leading to missed predictions of combinations that strongly correlate with PHI in documents.
Innovation Solution
A named-entity recognition (NER) model is trained to identify both PII and medical terms, combined with a random forest classifier that uses confidence values from the NER model to predict the presence of PHI in documents, ensuring robust detection and interpretable results.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing PHI detection models focus on single entity types (either PII or medical terms), then model complexity is reduced, but detection accuracy deteriorates due to failure to predict combinations of PII and medical terms that strongly correlate with PHI
Solution Approach 1:
The detection system is segmented into two specialized components: a PII detector trained specifically on personally identifiable information and a medical term detector trained specifically on medical terminology. Each detector focuses on its specialized domain, reducing individual model complexity while collectively achieving comprehensive PHI detection through the combination of both detectors' outputs
Solution Approach 2:
The system merges the outputs of the PII detector and medical term detector by combining their confidence scores and named entity results. This integration allows the system to detect PHI by identifying combinations of PII and medical terms, thereby improving detection accuracy without requiring a single complex model to learn all patterns simultaneously
2Measurement precision
If a comprehensive PHI detection model is trained to recognize both PII and medical terms together, then detection accuracy improves, but training data requirements and model complexity increase
Solution Approach 1:
The training process is segmented into two separate training phases: one for the PII detector using PII-labeled data and another for the medical term detector using medical term-labeled data. This segmentation allows each detector to be trained on specialized datasets, reducing the overall training data requirements compared to training a single comprehensive model on all PHI patterns simultaneously
3Measurement precision
If NER models are used to identify named entities in documents, then detection of PHI combinations improves, but processing time increases due to tokenization and confidence value calculations
Solution Approach 1:
The system performs preliminary tokenization and named entity recognition before the actual PHI detection classification. By pre-processing documents to identify and tag named entities with confidence values in advance, the system reduces the computational burden during the final classification stage, thereby reducing overall processing time while maintaining high detection accuracy
Data Source
AI summary
A named-entity recognition (NER) model detects named entities with types that correspond to protected health information (PHI) in potentially sensitive documents. The NER model is trained to detect named entities corresponding to both personally identifiable information (PII) and medical terms. Output of the NER model is preprocessed as input to a random forest classifier that outputs a verdict that documents comprise sensitive data. The verdict is interpretable via high confidence named entities detected by the NER model that led to the verdict.