In medicine we repeatedly need answers to the following kinds of questions:
Based on a set of documents (like an H&P, a few progress notes, perhaps the results of a few lab tests), does a patient have or not have a clinical condition, eg, things like significant cardiac disease, risk factors for seizure, etc?
Based on a set of documents (like drug labels, perhaps the results of a couple of additional studies), has a given drug been associated with clinical conditions, eg, things like significant cardiac risk, risk of renal effects, etc?
This is a time-consuming problem. Usually the clinician ultimately must carefully read the chart, labels, or study to make such an assessment. But an intelligent search of such documents could assist the clinician’s evaluation.
Unfortunately, it’s a hard problem for software to provide any kind of definitive guidance. A first naive pass might include searching the relevant documents for phrases like “myocardial infarction” and “conduction disorder”, or “head injury” and “epilepsy”.
It helps to use standardized synonyms for various conditions and automatically search documents for all such synonyms. Examples of synonym sources include the UMLS Metathesaurus and National Library of Medicine. A second approach is then to search using all the synonyms for a clinical condition.
Here is an experimental search of drug labels using medical synonyms from the National Library of Medicine.
Example:
- Enter “stribild” as drug
- Press “Fetch label(s)” button (or just press enter)
- Enter “pain” as search term
- Press “Get synonyms for search terms” button (or just press enter)
- Press “Search with synonyms”
Ultimately what is needed is semantic search, which may be amenable to hybrid approaches using these kinds of simple synonym searches combined with machine learning. For example, biologically-oriented BERT models provide some semantic search capability and are the subject of active research.