On-Line Anomaly Detection in Medical Documents
Publication Date: 2010-Jun-08
The IP.com Prior Art Database
The invention described herein is aimed at presenting a method for on-line detection of errors when a decision is made by the doctor at the point of care, based on comparison to previous, similar cases.
Millions of digital clinical documents are daily created by doctors all over the world. Such documents typically summarize care and services given to a patient or describe the health status of that patient, (e.g., a description of symptoms, blood test results, diagnosis and treatment). Usually, the treatment (medication, dose, period) is determined after all relevant information is collected. Cases of wrong treatment due to medical abbreviations, or human error are not very rare. For example, a recent study showed that 5% of medication errors (sometimes fatal) are due to abbreviations in notes .
In many cases, errors can be easy to detect if comparison to other cases shows significantly different results. For example, in a case where the medication's dose quantity given for a patient with a weight of 60kg is 200% higher than past prescriptions of the same medication for patients with similar weight.
Another example is a
medication that is given for a certain diagnosis and was never given before. These cases do not point directly on errors (a new medication was approved), but a sanity check that would alert suspicious treatment is essential. Data sets of healthcare centers can be used to design algorithms that would alert unusual treatment.
A reasonable assumption made is that the data is collected using a predefined template
consisting of fields that their value is provided by the Doctors. Height,
pressure, presence of certain diseases to name but a few. This setting is general and flexible since fields can easily be added, and there might be missing fields when data
was either not available,
or not relevant.
This patent is aimed at presenting a method for on-line detection of errors when a decision is made by the doctor at the point of care, based on comparison to previous similar cases. There are no current solutions for this problem, and the task is hard because on-line search over past documents is expensive, time consuming, and requires on-line access to data bases.
It should be emphasized that the learning process to be described requires occasionally updates, and misdetections or false alarms may occur (for example when using a new medication, or when data is not available for the given case).
Two types of comparison are possible:
when large sets of data for a given diagnosis are
available, features are compared to ranges of previously seen data. When there are only few examples available, comparison is held against all of them.
The first type requires a training phase. It is performed when the number of available examples of a given diagnosis is large enough (parameter that can be defined by the user). The process starts by going over all the documents that are related to that diagnosis. Every possible input is correlated with every possible output.
dependency criteria can be used (for example mutual information instead of linear correlation). This data is saved, and a representation of the connection be...