Machine learning model for extracting diagnoses, treatments, and key dates

Inventors

Rich, Alexander • Leybovich, Barry • IRVINE, BENJAMIN • Singh, Nisha • Birnbaum, Benjamin

Assignees

Flatiron Health Inc

Abstract

A model-assisted system for determining a patient event date may include a processor. The processor may be programmed to access a database storing a medical record associated with a patient, the medical record comprising unstructured data; analyze the unstructured data to identify a plurality of snippets of information in the medical record associated with a patient event; determine a date associated with each of the plurality of snippets, identify a plurality of query periods associated with the patient event; and generate, for each of the query periods, a probability of whether the patient event occurred during the query period based on the plurality of snippets and the associated dates.

Core Innovation

The document describes a model-assisted system for determining a patient event date from a medical record that includes unstructured data. The system accesses a database storing the medical record, analyzes the unstructured data to identify a plurality of snippets of information associated with a patient event, and determines a date associated with each snippet. The system generates a plurality of snippet vectors based on the snippets by replacing at least one term within the snippets with a tokenized representation of the term.

After snippet vector generation, the system identifies a plurality of query periods associated with the patient event. For each query period, the system applies a plurality of aggregation functions to the plurality of snippet vectors and the associated dates to generate a plurality of query outputs. The aggregation functions include a first function representing a sum of the snippet vectors relative to at least one time window, a second function representing an average of the snippet vectors relative to the at least one time window, and a third function relative to the at least one time window that is different from the first and second functions and includes a smooth approximation function.

For each query period, the system generates a probability of whether the patient event occurred during the query period based on the query outputs. The described workflow supports model-assisted event date determination by combining snippet extraction, date association, tokenized snippet vector generation, time-window aggregation, and probability generation for each query period.

Claims Coverage

The partial content includes three independent claim sets (a system, a method, and a non-transitory computer-readable medium). Across these independent claims, the inventive coverage is centered on extracting event-related snippets from unstructured medical-record data, assigning dates to snippets, creating tokenized snippet vectors, aggregating snippet-vector information over time windows using multiple functions including a smooth approximation, and generating per-query-period probabilities that the patient event occurred.

The independent claim coverage consistently requires snippet extraction and date association from unstructured medical-record data, tokenized representation to form snippet vectors, multiple aggregation functions over time windows including a smooth approximation function, and probability generation per query period that the patient event occurred during that period.

Stated Advantages

Documented Applications