Determining patient condition from unstructured text data

Non-transitory computer-readable media having computer-executable instructions embodied thereon that when executed, facilitate a method for determining one or more likely clinical conditions for a human patient using a structured topic model (STM), the method comprising:

• acquire patient information comprising a set of historical text documents related to care of human patients, each document having metadata labels ascribing one or more conditions of a patient identified at a time the document was created;
• receive a structured topic modeling (STM) model associating terms with metadata labels;
• receive one or more topic-clusters for a candidate condition, each topic-cluster associated with a condition and comprising an association of one or more terms and one or more metadata labels;
• perform K-means clustering on the patient information, using terms in the set of historical text documents and the STM thereby forming a set of patient clusters;
• calculate a lexical distance between each patient cluster and the one or more topic clusters to determine statistical similarity of the patient cluster to the one or more topic clusters;
• for each patient cluster statistically similar to a particular topic cluster, determining that the patient likely has the condition associated with the topic cluster;
• and
• determining a set of candidate conditions associated with each patient cluster based on the statistical similarity of the patient cluster to the one or more topic clusters;
• updating the patient information to indicate that the patient has a likelihood of having the condition;
• receiving unstructured clinical narratives associated with a particular patient;
• determining, using the STM model and the received unstructured clinical narratives, a likely cluster membership of the particular patient into one or more of the topic clusters, wherein the likely cluster membership is determined by calculating a quantitative lexical distance between the unstructured clinical narratives associated with the patient and the set of candidate conditions;
• and
• updating the unstructured clinical narrative to indicate the likely cluster membership of the particular patient.

A method for determining one or more patient conditions from unstructured text data, the method comprising:

• acquiring a set of historical text documents related to care of human patients, each document having metadata labels ascribing one or more conditions of a patient associated with the document, the one or more conditions identified at a time the document was created;
• receiving a structured topic modeling (STM) model associating terms with metadata labels;
• receiving a set of clusters comprising an association of one or more terms and one or more metadata labels;
• determining, for each cluster, a listing of conditions, based on the association of one or more terms and one or more metadata labels of the cluster;
• receiving unstructured clinical narratives associated with a particular patient;
• determining using the STM model and the received unstructured clinical narratives, a likely cluster membership is determined by calculating a quantitative lexical distance between the unstructured clinical narratives associated with the particular patient and the listing of conditions;
• and
• storing an electronic record of the likely cluster membership of the particular patient in the data store.