Abstract: Systems and methods for efficient information retrieval for clinical data are provided. A document including one or more patient health records can be obtained and tokenized to generate a sequence of tokens, each corresponding to a word or sub-word within the document. A plurality of segments each containing a token sequence comprising a sub-set of the tokens are then generated. Next, a transformer model is used to process each of the plurality of segments to generate a corresponding word-level encoding for each segment. The word-level encodings for each of the segments are combined and fused to obtain document-level contextual data. The combined and fused word-level encodings can then be analyzed, such as to identify named entities and relationships between them.

Abstract: Systems and methods for data normalization are disclosed. The disclosed systems and methods train one or more machine learning model based on a plurality of annotated records and apply these trained models to new or updated records. The models are configured to analyze the records and generate classifications or concepts from the records and corresponding confidence scores. If the confidence scores exceed a threshold, the corresponding classifications or concepts are appended to the appropriate record.

Abstract: Systems and methods for de-identifying patient data are disclosed herein. In some embodiments, a method for de-identifying patient data includes receiving a patient record including one or more identifiers. The method can include generating a first de-identified record from the patient record using a first de-identification process. The first de-identification process can be configured to produce a first re-identification risk score. The method can further include receiving a request from a data recipient to access the first de-identified record. The method can also include generating a second de-identified record from the first de-identified record by using a second de-identification process. The second de-identification process can be configured to produce a second re-identification risk score lower than the first re-identification risk score.