System and methods for generating and leveraging a disease-agnostic model to predict chronic disease onset
Methods, systems, and computer-readable media are disclosed herein for generating a disease-agnostic data model that can be used to predict the onset of multiple chronic diseases in individual patients. In an aspect, the data model is made by autonomously selecting features from longitudinal medical records of patients having chronic diseases that will become predictors for the onset of a chronic disease. The features are vectorized around a disease onset date and processed through a recurrent neural network to produce the data model. Then, the data model may leveraged to predict, for new longitudinal medical records that are input, a future time period when another patient is predicted to experience the onset of the chronic disease. The same data model may utilized to make predictions for multiple chronic diseases.
1 . A computerized method performed by one or more hardware processors of a medical-information computing system, the computerized method comprising:
accessing electronic data that includes a plurality of encoded longitudinal medical histories, the plurality of encoded longitudinal medical histories being created from clinical data extracted from a corresponding plurality of electronic health records, and each record of the plurality of electronic health records including (a) a defined time span associated with the record and (b) comprehensive record-specific information for the defined time span associated with the record;
receiving, via the one or more hardware processors, an indication for a chronic disease associated with at least a portion of: the defined time spans and the comprehensive record-specific information from which the plurality of encoded longitudinal medical histories are created;
in response to the receiving of the indication via the one or more hardware processors, automatically performing by the one or more hardware processors and without user input:
(a) generating a plurality of distributions for a plurality of features in the plurality of encoded longitudinal medical histories; and
(b) selecting, via the one or more hardware processors and without user input, two or more of the plurality of features from the plurality of distributions, wherein the one or more hardware processors are associated with an electronic memory of the medical-information computing system, and wherein the two or more of the plurality of features are selected as corresponding to the chronic disease;
generating an onset time period for the chronic disease, the onset time period for the chronic disease being generated (a) for a particular longitudinal medical history of the plurality of encoded longitudinal medical histories associated with the chronic disease, (b) based on a first plurality of vectors for the plurality of encoded longitudinal medical histories, and (c) based further on a recurrent neural network, wherein: the first plurality of vectors includes the two or more of the plurality of features selected from the plurality of distributions as corresponding to the chronic disease,
the recurrent neural network is configured using input data that is based on input vectors of data corresponding to instances of information selected from a group comprising longitudinal medical history data, the first plurality of vectors, and chronic disease indication data,
the recurrent neural network is further configured based on aligning sequential observation time periods of longitudinal medical histories relative to known dates of onset of the chronic disease for two or more of the plurality of encoded longitudinal medical histories, and
an instance of the input data associated with the particular longitudinal medical history is input to the recurrent neural network to produce output prediction data indicating the onset time period for the chronic disease; and
presenting, via an electronic interface and based on at least the output prediction data, the onset time period for the chronic disease, wherein presenting the onset time period for the chronic disease includes communicating via the one or more hardware processors and via the electronic interface a date or a range of dates associated with the onset time period for the chronic disease.
2 . The computerized method of claim 1 , wherein the onset time period is generated based further on a second plurality of vectors for a plurality of medical concepts identified in the electronic data, by concatenating each of the first plurality of vectors with one of the second plurality of vectors to generate a specific vector for each longitudinal medical history of the plurality of encoded longitudinal medical histories.
3 . The computerized method of claim 1 , wherein the computerized method is performed via multiple hardware processors at different geographic locations associated with a distributed memory of the medical-information computing system, and wherein generating the plurality of distributions and selecting the two or more of the plurality of features are performed autonomously by the one or more hardware processors.
4 . The computerized method of claim 1 , wherein generating the onset time period is performed via the one or more hardware processors at:
an embedding layer of the medical-information computing system,
in combination with a self-attention layer of the medical-information computing system,
in combination with a recurrent neural network layer of the medical-information computing system.
5 . One or more non-transitory media having instructions which, when executed by one or more hardware processors of a medical-information computing system, cause the one or more hardware processors to perform a plurality of operations, the operations comprising:
accessing electronic data that includes a plurality of encoded longitudinal medical histories, the plurality of encoded longitudinal medical histories being created from clinical data extracted from a corresponding plurality of electronic health records, and each record of the plurality of electronic health records including (a) a defined time span associated with the record; and (b) comprehensive record-specific information for the defined time span associated with the record;
receiving, via the one or more hardware processors, an indication for a chronic disease associated with at least a portion of: the defined time spans and the comprehensive record-specific information from which the plurality of encoded longitudinal medical histories are created;
in response to the receiving of the indication via the one or more hardware processors, automatically performing by the one or more hardware processors and without user input:
(a) generating a plurality of distributions for a plurality of features in the plurality of encoded longitudinal medical histories; and
(b) selecting, via the one or more hardware processors and without user input, two or more of the plurality of features from the plurality of distributions, wherein the one or more hardware processors are associated with an electronic memory of the medical-information computing system, and wherein the two or more of the plurality of features are selected as corresponding to the chronic disease;
generating an onset time period for the chronic disease, the onset time period for the chronic disease being generated (a) for a particular longitudinal medical history of the plurality of encoded longitudinal medical histories associated with the chronic disease, (b) based on a first plurality of vectors for the plurality of encoded longitudinal medical histories, and (c) based further on a recurrent neural network, wherein: the first plurality of vectors includes the two or more of the plurality of features selected from the plurality of distributions as corresponding to the chronic disease,
the recurrent neural network is configured using input data that is based on input vectors of data corresponding to instances of information selected from a group comprising longitudinal medical history data, the first plurality of vectors, and chronic disease indication data,
the recurrent neural network is further configured based on aligning sequential observation time periods of longitudinal medical histories relative to known dates of onset of the chronic disease for two or more of the plurality of encoded longitudinal medical histories, and
an instance of the input data associated with the particular longitudinal medical history is input to the recurrent neural network to produce output prediction data indicating the onset time period for the chronic disease; and
presenting, via an electronic interface and based on at least the output prediction data, the onset time period for the chronic disease, wherein presenting the onset time period for the chronic disease includes communicating via the one or more hardware processors and via the electronic interface a date or a range of dates associated with the onset time period for the chronic disease.
6 . The one or more non-transitory media of claim 5 , wherein each of the first plurality of vectors corresponds to the electronic data for a different time period for one of the plurality of encoded longitudinal medical histories.
7 . The one or more non-transitory media of claim 5 , wherein operations further comprise, for each individual:
comparing two or more features in the first plurality of vectors to two or more features temporally associated with onset for the chronic disease in a disease-agnostic onset prediction model;
determining a future time period for the at least one of the plurality of encoded longitudinal medical histories, the future time period being associated with a greatest likelihood of onset of the chronic disease for the at least one of the plurality of encoded longitudinal medical histories based on the comparing; and
identifying the future time period as the onset time period for the chronic disease for the at least one of the plurality of encoded longitudinal medical histories.
8 . The one or more non-transitory media of claim 7 , wherein the operations further comprise electronically writing the onset time period to a memory associated with the medical-information computing system, and wherein the onset time period includes a date and a range of dates.
9 . The one or more non-transitory media of claim 5 , wherein the operations further comprise updating the recurrent neural network based on information associated with additional instances of the data.
10 . The one or more non-transitory media of claim 5 , wherein configuring the recurrent neural network is based at least in part on a process selected from a group comprising supervised machine learning, reinforcement machine learning, and unsupervised machine learning.
11 . The one or more non-transitory media of claim 5 , wherein the operations further comprise automatically selecting, via the one or more hardware processors, two of the plurality of features from the plurality of distributions.
12 . The one or more non-transitory media of claim 5 , wherein the recurrent neural network includes at least one of a long short-term memory (LSTM) type of recurrent neural network or a bidirectional gated recurrent units (GRU) type recurrent neural network layer.
13 . The one or more non-transitory media of claim 5 , wherein the recurrent neural network includes a bi-directional neural network architecture having a plurality of nodes and configured to allow an output from a first node of the plurality of nodes to affect a subsequent input to the first node of the plurality of nodes.
14 . The one or more non-transitory media of claim 5 , wherein for each of the plurality of electronic health records, the comprehensive record-specific information for the defined time span comprises:
a diagnosis code and/or procedure code, that is associated with the chronic disease; and
descriptions of two or more medical encounters occurring within the defined time span relative to the diagnosis code and/or procedure code associated with the chronic disease.
15 . The one or more non-transitory media of claim 5 , wherein the operations further comprise: embedding the first plurality of vectors for the plurality of encoded longitudinal medical histories; and prior to the presenting, processing the first plurality of vectors using the recurrent neural network.
16 . The one or more non-transitory media of claim 5 , wherein the operations further comprise:
embedding the first plurality of vectors for the plurality of encoded longitudinal medical histories;
embedding a second plurality of vectors for a plurality of medical concepts identified in the electronic data; and
prior to the presenting:
processing the first plurality of vectors using the recurrent neural network; and
processing the second plurality of vectors using the recurrent neural network.
17 . The one or more non-transitory media of claim 16 , wherein:
embedding the first plurality of vectors comprises:
flattening two or more features from one or more sequential observation time periods in the longitudinal medical histories, and
embedding the second plurality of vectors comprises:
flattening one or more medical concepts from the longitudinal medical histories.
18 . The one or more non-transitory media of claim 5 , wherein the operations further comprise embedding the first plurality of vectors for the plurality of encoded longitudinal medical histories, the first plurality of vectors including one or both of diagnosis codes and procedure codes that are present in the electronic data, and producing a particular disease-agnostic onset prediction model based on the embedding.
19 . The one or more non-transitory media of claim 18 , wherein the operations further comprise embedding a second plurality of vectors for a plurality of medical concepts identified in the electronic data, the second plurality of vectors including information selected from a group comprising medications, allergies, care plans, provider appointments, questionnaire information, and a frequency for one or more diagnosis codes and procedure codes that are present in the electronic data, and wherein the producing of the particular disease-agnostic onset prediction model is based on the embedding of the second plurality of vectors.
20 . The one or more non-transitory media of claim 5 , wherein a first vector of the first plurality of vectors includes:
a gender, an age, and a race that are present in the electronic data and that are specific to a first longitudinal medical history of the plurality of encoded longitudinal medical histories; and
a frequency of one or both of diagnosis codes and procedure codes that are (a) present in the electronic data and (b) specific to the first longitudinal medical history.
21 . The one or more non-transitory media of claim 5 , wherein the two or more of the plurality of features are selected, as corresponding to the chronic disease, based on a heat map analysis performed automatically via the one or more hardware processors.
22 . A medical-information computing system having one or more hardware processors configured to perform a plurality of operations, the operations comprising:
accessing electronic data that includes a plurality of encoded longitudinal medical histories, the plurality of encoded longitudinal medical histories being created from clinical data extracted from a corresponding plurality of electronic health records, and each record of the plurality of electronic health records including (a) a defined time span associated with the record and (b) comprehensive record-specific information for the defined time span associated with the record;
receiving, via the one or more hardware processors, an indication for a chronic disease associated with at least a portion of: the defined time spans and the comprehensive record-specific information from which the plurality of encoded longitudinal medical histories are created;
in response to the receiving of the indication via the one or more hardware processors, automatically performing by the one or more hardware processors and without user input:
(a) generating a plurality of distributions for a plurality of features in the plurality of encoded longitudinal medical histories; and
(b) selecting, via the one or more hardware processors and without user input, two or more of the plurality of features from the plurality of distributions, wherein the one or more hardware processors are associated with an electronic memory of a the medical-information computing system, and wherein the two or more of the plurality of features are selected as corresponding to the chronic disease;
generating an onset time period for the chronic disease, the onset time period for the chronic disease being generated (a) for a particular longitudinal medical history of the plurality of encoded longitudinal medical histories associated with the chronic disease, (b) based on a first plurality of vectors for the plurality of encoded longitudinal medical histories, and (c) based further on a recurrent neural network, wherein: the first plurality of vectors includes the two or more of the plurality of features selected from the plurality of distributions as corresponding to the chronic disease,
the recurrent neural network is configured using input data that is based on input vectors of data corresponding to instances of information selected from a group comprising longitudinal medical history data, the first plurality of vectors, and chronic disease indication data,
the recurrent neural network is further configured based on aligning sequential observation time periods of longitudinal medical histories relative to known dates of onset of the chronic disease for two or more of the plurality of encoded longitudinal medical histories, and
an instance of the input data associated with the particular longitudinal medical history is input to the recurrent neural network to produce output prediction data indicating the onset time period for the chronic disease; and
presenting, via an electronic interface and based on at least the output prediction data, the onset time period for the chronic disease, wherein presenting the onset time period for the chronic disease includes communicating via the one or more hardware processors and via the electronic interface a date or a range of dates associated with the onset time period for the chronic disease.