System to identify size and location information from unstructured inputs
A method including receiving an input of unstructured data; extracting at least one size entity and at least one location entity from the unstructured data using a named entity recognition model; parsing dimensions of the at least one size entity; and outputting structured size data and structured location data.
1 . A method comprising:
receiving an input of unstructured data comprising a plurality of health records of an individual, wherein each health record of the individual comprises a free-text radiography report from a point in time;
extracting at least one size entity and at least one location entity from the unstructured data using a named entity recognition model, wherein the named entity recognition model is a pre-trained BERT-based named entity recognition model trained using a dataset of free-text radiography reports which are manually annotated to highlight size entities and location entities current to a particular radiography report, wherein extracting the at least one size entity and the at least one location entity comprises:
extracting from a first free-text radiography report of the individual a first size entity and a first location entity of a first lesion while excluding results of any past radiography reports, wherein the first radiography report is obtained at a first time, wherein the first size entity comprises a first size of the first lesion, wherein the first location entity comprises a first anatomical location of the first lesion; and
extracting from a second free-text radiography report of the individual a second size entity and a second location entity of a second lesion while excluding results of any past radiography reports including the first radiography report, wherein the second radiography report is obtained at a second time after the first time;
identifying that the second lesion is a later version of the first lesion by confirming that the second location entity comprises the same anatomical location as the first anatomical location, wherein the second size entity comprises an updated size of the first lesion;
parsing dimensions and values of the at least one size entity including the first size entity and the second size entity; and
outputting structured size data, from the parsed dimensions and values, and structured location data by displaying the structured size data and the structured location data as a function of time, wherein the display of the structured size data reports changes in size of the same lesion at the same anatomical location between the first time and the second time allowing a user to view lesion progression and evaluate treatment without having to search within the unstructured data.
2 . The method of claim 1 , further comprising linking the at least one size entity and the at least one location entity.
3 . The method of claim 2 , wherein the at least one location entity is in a sentence with exactly one size entity and that location entity is linked with that size entity.
4 . The method of claim 2 , wherein the at least one location entity is in a sentence with more than one size entity and that location entity is linked with a size entity with a closest distance metric.
5 . The method of claim 4 , wherein the distance metric is a number of words.
6 . The method of claim 4 , wherein the distance metric is a path in a dependency tree.
7 . The method of claim 2 , wherein the at least one location entity is in a sentence without a size entity and is linked to a size entity in an adjacent sentence.
8 . The method of claim 1 , further comprising determining an anatomical location for the at least one size entity.
9 . The method of claim 8 , wherein determining the anatomical location comprises using metadata of the input.
10 . The method of claim 8 , wherein determining the anatomical location comprises using a dictionary mapping types of health records to anatomical locations.
11 . The method of claim 8 , wherein determining the anatomical location comprises using a location entity linked with the at least one size entity.
12 . The method of claim 1 , comprising extracting more than one size entity and further comprising identifying any size entities that refer to a common lesion.
13 . The method of claim 12 , wherein identifying any size entities that refer to a common lesion comprises comparing dimensions of two or more size entities.
14 . The method of claim 12 , wherein identifying any size entities that refer to a common lesion comprises identifying a section of the input where the size entities are located.
15 . The method of claim 12 , wherein identifying any size entities that refer to a common lesion comprises identifying an estimation keyword.
16 . The method of claim 1 , wherein the named entity recognition model is trained using an annotated unstructured dataset.
17 . The method of claim 1 , wherein the named entity recognition model is trained using a dataset of manually annotated health records.
18 . The method of claim 1 , wherein the named entity recognition model extracts size entities and location entities from a current observation without extracting size entities and location entities from a prior observation.
19 . The method of claim 1 , wherein parsing the dimensions comprises using regular expressions.
20 . The method of claim 1 , wherein parsing the dimensions of each of the size entities comprises extracting a vector of numeric values and a vector of units.
21 . The method of claim 1 , wherein parsing the dimensions of each of the size entities comprises calculating a maximum dimension of each parsed size entity.
22 . A system comprising:
a non-transitory memory; and
one or more hardware processors configured to read instructions from the non-transitory memory that, when executed cause one or more of the hardware processors to perform operations comprising:
receiving an input of unstructured data comprising a plurality of health records of an individual, wherein each health record of the individual comprises a free-text radiography report from a point in time;
extracting at least one size entity and at least one location entity from the unstructured data using a named entity recognition model, wherein the named entity recognition model is a pre-trained BERT-based named entity recognition model trained using a dataset of free-text radiography reports which are manually annotated to highlight size entities and location entities current to a particular radiography report, wherein extracting the at least one size entity and the at least one location entity comprises:
extracting from a first free-text radiography report of the individual a first size entity and a first location entity of a first lesion while excluding results of any past radiography reports, wherein the first radiography report is obtained at a first time, wherein the first size entity comprises a first size of the first lesion, wherein the first location entity comprises a first anatomical location of the first lesion; and
extracting from a second free-text radiography report of the individual a second size entity and a second location entity of a second lesion while excluding results of any past radiography reports including the first radiography report, wherein the second radiography report is obtained at a second time after the first time;
identifying that the second lesion is a later version of the first lesion by confirming that the second location entity comprises the same anatomical location as the first anatomical location, wherein the second size entity comprises an updated size of the first lesion;
parsing dimensions and values of the at least one size entity including the first size entity and the second size entity; and
outputting structured size data, from the parsed dimensions and values, and structured location data by displaying the structured size data and the structured location data as a function of time, wherein the display of the structured size data reports changes in size of the same lesion at the same anatomical location between the first time and the second time allowing a user to view lesion progression and evaluate treatment without having to search within the unstructured data.
23 . The system of claim 22 , wherein the operations further comprise linking the at least one size entity and the at least one location entity.
24 . The system of claim 23 , wherein the at least one location entity is in a sentence with exactly one size entity and that location entity is linked with that size entity.
25 . The system of claim 23 , wherein the at least one location entity is in a sentence with more than one size entity and that location entity is linked with a size entity with a closest distance metric.
26 . The system of claim 23 , wherein the at least one location entity is in a sentence without a size entity and is linked to a size entity in an adjacent sentence.
27 . The system of claim 22 , wherein the operations further comprise determining an anatomical location for the at least one size entity.
28 . The system of claim 27 , wherein determining the anatomical location comprises using metadata of the input.
29 . The system of claim 27 , wherein determining the anatomical location comprises using a location entity linked with the at least one size entity.
30 . The system of claim 22 , wherein the operations further comprise extracting more than one size entity and further comprising identifying any size entities that refer to a common lesion.
31 . The system of claim 30 , wherein identifying any size entities that refer to a common lesion comprises comparing dimensions of two or more size entities.
32 . The system of claim 22 , wherein the named entity recognition model is trained using an annotated unstructured dataset.
33 . The system of claim 22 , wherein the named entity recognition model extracts only size entities and location entities from a current observation without extracting size entities from a prior observation.
34 . A non-transitory computer-readable medium storing instructions that, when executed by one or more hardware processors, cause the one or more hardware processors to perform operations comprising:
receiving an input of unstructured data comprising a plurality of health records of an individual, wherein each health record of the individual comprises a free-text radiography report from a point in time;
extracting at least one size entity and at least one location entity from the unstructured data using a named entity recognition model, wherein the named entity recognition model is a pre-trained BERT-based named entity recognition model trained using a dataset of free-text radiography reports which are manually annotated to highlight size entities and location entities current to a particular radiography report, wherein extracting the at least one size entity and the at least one location entity comprises:
extracting from a first free-text radiography report of the individual a first size entity and a first location entity of a first lesion while excluding results of any past radiography reports, wherein the first radiography report is obtained at a first time, wherein the first size entity comprises a first size of the first lesion, wherein the first location entity comprises a first anatomical location of the first lesion; and
extracting from a second free-text radiography report of the individual a second size entity and a second location entity of a second lesion while excluding results of any past radiography reports including the first radiography report, wherein the second radiography report is obtained at a second time after the first time;
identifying that the second lesion is a later version of the first lesion by confirming that the second location entity comprises the same anatomical location as the first anatomical location, wherein the second size entity comprises an updated size of the first lesion;
parsing dimensions and values of the at least one size entity including the first size entity and the second size entity; and
outputting structured size data, from the parsed dimensions and values, and structured location data by displaying the structured size data and the structured location data as a function of time, wherein the display of the structured size data reports changes in size of the same lesion at the same anatomical location between the first time and the second time allowing a user to view lesion progression and evaluate treatment without having to search within the unstructured data.
35 . The non-transitory computer-readable medium of claim 34 , wherein the operations further comprise linking the at least one size entity and the at least one location entity.
36 . The non-transitory computer-readable medium of claim 35 , wherein the at least one location entity is in a sentence with exactly one size entity and that location entity is linked with that size entity.
37 . The non-transitory computer-readable medium of claim 35 , wherein the at least one location entity is in a sentence with more than one size entity and that location entity is linked with a size entity with a closest distance metric.
38 . The non-transitory computer-readable medium of claim 35 , wherein the at least one location entity is in a sentence without a size entity and is linked to a size entity in an adjacent sentence.
39 . The non-transitory computer-readable medium of claim 34 , wherein the operations further comprise determining an anatomical location for the at least one size entity.
40 . The non-transitory computer-readable medium of claim 39 , wherein determining the anatomical location comprises using metadata of the input.
41 . The non-transitory computer-readable medium of claim 39 , wherein determining the anatomical location comprises using a location entity linked with the at least one size entity.
42 . The non-transitory computer-readable medium of claim 34 , wherein the operations further comprise extracting more than one size entity and further comprising identifying any size entities that refer to a common lesion.
43 . The non-transitory computer-readable medium of claim 42 , wherein identifying any size entities that refer to a common lesion comprises comparing dimensions of two or more size entities.
44 . The non-transitory computer-readable medium of claim 34 , wherein the named entity recognition model is trained using an annotated unstructured dataset.
45 . The non-transitory computer-readable medium of claim 34 , wherein the named entity recognition model extracts only size entities and location entities from a current observation without extracting size entities from a prior observation.