Artificial Intelligence (AI)

CEBI Miner
Artificial intelligence (AI) is increasingly utilized in healthcare to help doctors and patients make better health-care decisions.

We utilize AI applications for clinical decision making and clinical decision support (CDS). Mostly focused in diagnostic imaging, we develop applications utilizing rule-based and machine learning algorithms.  These include applications in predictive modeling and natural language processing that are integrated into CDS systems and evidence-based imaging initiatives. We also utilize machine learning for augmenting evidence that is disseminated to the medical community. Ultimately, the goal is to enhance CDS systems that are available in clinical practice using evidence-based medicine.

We address challenges that relate to applying AI into medical practice include improving the quality of the data that we feed into AI systems, developing ways to evaluate whether an AI system performs adequately for specific tasks, and ensuring minimal workflow disruption for providers. Machine learning using supervised algorithms are developed and evaluated.

Use of Machine Learning to Identify Follow-Up Recommendations in Radiology Reports.

Evaluation of an Automated Information Extraction Tool for Imaging Data Elements to Populate a Breast Cancer Screening Registry.

Traditional AI models are developed and utilized in several studies for risk stratification and prediction.

Risk Stratification Model: Lower-Extremity Ultrasonography for Hospitalized Patients with Suspected Deep Vein Thrombosis.

A Clinical Model for the Early Diagnosis of Acute Pancreatitis in the Emergency Department.

created using python and dash plotly

Meet CEBI Miner

CEBI-Miner is a rule based natural language processing application geared for querying radiology report texts. The application can help synthesize a large cohort of report texts to the ones of a particular interest. Examples of questions it has helped answered here at the Center for Evidence-Based Imaging include which radiology reports contain miscellaneous liver lesions, whether prostate MRI reports were completed with an endorectal coil or without, and which reports contain adrenal lesions. The three main features of the application are search keywords, associated words, and negation words. 

LEARN MOREGET STARTED

Download CEBI Miner

Center-for-Evidence-Based-Imaging/CEBI-Miner: Created using Python and Dash Plotly – GitHub

CEBI Miner

The three main features of the application are

search keywords, associated words, and negation words.

Search Keywords

Are the words/phrases that must be present within the report text to indicate that a report may be of interest.

After running the NLP search, your results will contain a new column, TERM_COUNT, which is a count of the number of appearances of all your search keywords for each report text.

Think of the search keywords as the anchor to the user’s search.

It is through instances of these keywords that associated, and negation words will be searched.

Associated Words

Are the words/phrases most likely to be surrounding the user’s defined search keywords that would indicate it is a report of interest.

Associated words taken alone in the report texts may not indicate that a report is of interest, but when within a close proximity to search keywords they strengthen the interest in that specific report.

Negation Words

Are words/phrases that contradicts (or negates) all or part of the meaning of the sentence. This feature is important if the user’s particular interest is often negated in reports.

An example could be a sentence such as “There is no evidence of gallstones or IHBR dilatation”, within an abdominal MRI report text.

In this example you could be looking for reports where there is evidence of gallstones, and thus adding negation words could ensure you don’t pick reports such as the one above. 

Additional Features

Scope of Search – Allows the user to determine what part of the report the search should be executed on. The default option is the whole report, meaning it searches the entire text for each row of the column selected. The other options are patient information, exam information, report (with impression), impression only, and technique section. These options are parsed from the format of Brigham and Women’s radiology reports. Thus, if the format of your reports varies from their structure you should choose the option to run on the whole report text.

Associated/Negation Words Search Distance and Direction – Associated/Negation words are searched around the user’s defined search keywords. The user can choose whether these words are searched before or after the search keywords. They can also define the search distance from the user’s keywords. Distance can be in words or sentences. For example, an associated distance of 4 words before will look for associated words within four words before the user’s keyword. Note that if the sentence started only 3 words before the search keyword the program will not search the last word of the previous sentence. If the user sets search distance to 0 sentences, the program searches for associated words within the same sentence as the user’s search keywords.

Index Column – Index column lets the users define a column which to filter on. For example, if you had a modality column it could be filtered to only run CEBI-Miner on modalities equal to CT.

Back To Top