The post New review: Machine Learning for Clinical Text Data appeared first on gnTEAM.

As part of the NIHR-funded project Developing and Enhancing the Usefulness of Patient Experience and Narrative Data (DEPEaND), we have developed a text mining software to analyse themes and sentiment expressed in free-text patient service feedback comments (e.g. Family and Friends Test).

The topic-specific opinion mining techniques have been applied to extract commonly mentioned themes from patient comments and to detect the polarity related to each theme. Following an initial manual inspection of a small sample, we focused on four main themes (staff attitude, quality of care, waiting time and environment) and associated sentiment (positive and negative/neutral). Two machine-learning methods have been developed (using Python and R), focusing on the segmentation of patient comments, and then prediction of the themes (and sentiment) using various machine learning algorithms look answer to the question. The system can also combine outputs of the two systems using a probability-threshold technique, and show top comments for each of the themes, for which the system is most confident (i.e. have the most confident prediction). The approach has been tested in two clinical settings – a general hospital and a mental health trust.

The methodology is explained in the paper “Segmentation-based mining of free-text patient feedback comments” (in preparation).

The code to train the software is available at here. The user manual is available here. Note that an annotated corpus is needed to generate the models.

The output of the software can be further analysed (using Stata) and visualised (using LaTeX) to automatically generate reports containing main themes, sentiment and most representative comments over a period of time. The software for this is available here (Stata code) and here (LaTeX). We note that Stata does require a licence.

Funded by:  

The post DEPEND appeared first on gnTEAM.

The paper presents our experience in learning to identify personal information as part of the 2016 CEGS N-GRID Shared Tasks Track 1, which evaluated de-identification methods on a set of psychiatric evaluation notes for up to 25 different types of Protected Health Information (PHI). The methods we used rely on machine learning on either a large or small feature space, with additional strategies, including two-pass tagging and multi-class models, which both proved to be beneficial. The results show that the integration of the proposed methods can identify Health Information Portability and Accountability Act (HIPAA) defined PHIs with overall F1-scores of ∼90% and above. Yet, some classes (Profession, Organization) proved again to be challenging given the variability of expressions used to reference given information.

The post Learning to identify Protected Health Information in free text appeared first on gnTEAM.

CliNER is a command line tool for identification of mentions of four categories of clinically relevant events: Problems, Tests, Treatments and Clinical Departments. It also recognises and normalises clinical temporal expressions. It was developed as part of the i2b2 2012 text mining challenge and therefore has been trained and optimised on the i2b2 data.

For example, for input:

He had surgery about 3 weeks ago and had the lining cleaned and a biopsy was performed.

CliNER will produce the following output (stand-off):

<?xml version="1.0" encoding="UTF-8" ?>
<ClinicalNarrativeTemporalAnnotation>
<TEXT><![CDATA[
He had surgery about 3 weeks ago and had the lining cleaned and a biopsy was performed.
]]></TEXT>
<TAGS>
<EVENT id="E1" start="65" end="73" text="a biopsy" modality="FACTUAL" polarity="POS" type="TEST" />
<EVENT id="E1" start="8" end="15" text="surgery" modality="FACTUAL" polarity="POS" type="TREATMENT" />
<TIMEX3 id="T1" start="22" end="33" text="3 weeks ago" type="DATE" val="2014-07-29" mod="NA" />
</TAGS>

System Architecture

• Implemented in Java, using cTAKES, CRF++ and Clinical NorMA
• Support for multiple formats, currently supporting:
  • standoff XML
  • character offset-based format

Algorithm details
Conditional random fields with IO scheme and five groups of features:

• Lexical features included the token itself, its lemma, and POS tag, as well as lemmas and POS tags of the surrounding tokens. Each token was also assigned features from its associated chunk (phrase): the type of phrase (nominal, verbal, etc), tense and aspect (if the phrase was verbal), the location of the token within the chunk (beginning or inside), and the presence of negation.
• Domain features capture mentions of specific clinical/healthcare concepts. Mentions of Problem, Test, and Treatment (as generated by cTAKES) were assigned to the token.
• Semantic role features model dependencies between the token and associated verb. Each token is assigned the role, the verb, and their combination (eg, ‘object+perform’) in order to capture particular verb–role preferences.
• Section type feature represents the section type in which the token appeared.
• Temporal expression (TE) features represent five features that indicated the presence of the five common types of constituents of TEs in a given token.

More details on the architecture and the performance of the tool can be found in the paper below. Please cite this publication if you use CliNER:

Kovačević, A., Dehghan, A., Filannino, M., Keane, J. A., & Nenadic, G. (2013). Combining rules and machine learning for extraction of temporal expressions and events from clinical narratives. Journal of the American Medical Informatics Association, 20(5), 859-866.

The post CliNER appeared first on gnTEAM.

The post Mining term associations and events from bio-literature appeared first on gnTEAM.

The post Blog sentiment analysis appeared first on gnTEAM.

The post Dr Michele Filannino appeared first on gnTEAM.