Systems and methods for computing with private healthcare data
Inventors
Ardhanari, Sankar • MURUGADOSS, Karthik • Aravamudan, Murali • Rajasekharan, Ajit
Assignees
Interested in licensing this patent?
MTEC can help explore whether this patent might be available for licensing for your application.
Publication Number
US-12216799-B2
Publication Date
2025-02-04
Expiration Date
Abstract
Techniques are provided for computing with private healthcare data. The techniques include a de-identification method including receiving a text sequence; providing the text sequence to a plurality of entity tagging models, each of the plurality of entity tagging models being trained to tag one or more portions of the text sequence having a corresponding entity type; tagging one or more entities in the text sequence using the plurality of entity tagging models; and obfuscating each entity among the one or more tagged entities by replacing the entity with a surrogate, the surrogate being selected based on one or more attributes of the entity and maintaining characteristics similar to the entity being replaced.
Core Innovation
The invention addresses de-identification of private healthcare data by receiving a plurality of data sets, including a labeled data set for one or more entity types and an unlabeled data set for the one or more entity types. It determines one machine-learning model from a plurality of machine-learning models for each of the one or more entity types and fine-tunes the determined machine-learning model for each entity type using training data sets that include the labeled data set and the unlabeled data set.
Fine-tuning includes training the determined machine-learning model using the first training data set and validating the trained machine-learning model by generating a recall score for each entity type and comparing the recall score to a threshold for the recall score for each entity type. After validation, the trained machine-learning model is updated using the second training data set as a function of the validation, and the second data set is obfuscated using the fine-tuned machine-learning model.
In the disclosed implementation, the de-identification is performed on entity-based masking of unstructured EHR text using entity tagging models configured to identify entity types and obfuscate detected content. Obfuscation includes information masking and additional information obfuscation using an aggregator and DReG filters, with rule-based templates generated from syntax templates inferred via NER/token patterns. The system further uses surrogate substitutions selected to maintain similar attributes and includes optional consistent per-patient surrogate mapping for obfuscating personal identifiers.
Claims Coverage
The independent claim covers a de-identification method that trains and fine-tunes machine-learning models per entity type using labeled and unlabeled data, validates the models via recall-score thresholds per entity type, updates the model based on validation, and obfuscates the second data set using the fine-tuned models.
De-identification method with labeled and unlabeled entity-type training
A de-identification method comprising receiving a plurality of data sets including a first data set comprising a labeled data set for one or more entity types and a second data set comprising an unlabeled data set for the one or more entity types.
Per-entity-type model determination and fine-tuning
Determining one machine-learning model from a plurality of machine-learning models for each of one or more entity types and fine-tuning the determined machine-learning model for each of the one or more entity types, where fine-tuning comprises creating training data sets including the first data set and the second data set and training the determined machine-learning model using the first training data set.
Validation by recall-score threshold per entity type and model update
Validating the trained machine-learning model by generating a recall score for each entity type of the one or more entity types, comparing the recall score to a threshold for the recall score for each entity type, and updating the trained machine-learning model using the second training data set as a function of the validation.
Obfuscating unlabeled data using the fine-tuned model
Obfuscating the second data set using the fine-tuned machine-learning model.
Overall, the claim coverage centers on per-entity-type machine-learning model selection and fine-tuning using labeled and unlabeled data, validation using recall-score thresholds per entity type, updating based on that validation, and obfuscating the unlabeled data set using the fine-tuned model.
Stated Advantages
Mitigate runtime leaks and preserve provenance using a secure enclave and federated chain of trust execution.
Preserve de-identification probability and inscrutability through encrypted I/O/proofs.
Prevent PHI false negatives using DReG-style filtering and rule-based templates.
Maintain similarity of attributes through information obfuscation using surrogates selected to preserve similar attributes.
Documented Applications
Information-retrieval over heterogeneous/biomedical corpora, including masked corpora, using fragment-based search, knowledge-base query expansion, statistical significance scoring, and neural inference of relationships (e.g., drug-disease adverse event).
Downstream augmented curation and temporal discrimination of de-identified health records, including enriched patient phenotypes over a temporal window.
Interested in licensing this patent?