Blog

Why negation detection?

The recognition of negations is an important task in medical text mining: during the treatment of patients many examinations are carried out and both the presence and absence of symptoms and conspicuous findings are documented. The following are examples of negated findings.

• Patient shows no meningism, normal light reaction, no indication of scleral icterus, oral mucosa and pharyngeal ring are irritation-free and inconspicuous, thyroid gland not enlarged, nausea did not occur.

Due to the complexity of human language there is the particular challenge of e.g. double negations, pseudo-negations etc. In the following examples, findings are not negated, although there are clear indicator words (“inconspicuous”, “not”, “excluded”) in the same sentence.

• Apart from a slight skin rash, the physical examination not inconspicuous.
• A tumor cannot be excluded with certainty.
• Spleen and liver not palpable due to ascites.

Machine learning forms the basis for improved negation detection

In rule-based approaches, these complex patterns are recognized via a wealth of rules. Here Averbis uses the text mining rule engine UIMA RUTA, which is developed by Averbis and made available to the community open-source. Many negation patterns can already be reliably recognized using rules.

Internal analyses have shown that the reliable detection of complex negation patterns leads to an enormous complexity in the rules and the maintainability suffers greatly from this. For this reason, we have added a machine learning approach to this procedure in recent weeks. For this purpose, we have compiled and annotated a large internal training data set consisting of English and German training data. Additionally, we used data from the public i2b2-2010 dataset.

For each diagnosis, the ML model gets information about the words before and after each diagnosis, whether the diagnosis is in an enumeration and whether a negation indicator such as “none” or “not” is present. The rest of the decision making process is fully automated. We have improved individual errors of the ML model by applying specific post-processing rules.

Negation status of diagnoses can be determined with high accuracy

Fortunately, we were able to increase the performance of the negation detection to 96% F1 score (English) and 95% (German) by combining the rule-based and ML method. An in-depth analysis of the errors showed that the ML approach was also able to identify faulty annotations in the gold standard.

EconSight is a new generation, independent and neutral consulting company in the best Swiss sense. Dr. Jochen Spuck, Chief Technology Officer (CTO) at EconSight in Basel, has prepared a scientific analysis for the Bertelsmann Foundation on the topic of “WORLD CLASS PATENTS IN FUTURE TECHNOLOGIES”, in which the Patent Monitor from Averbis is used.

(Read the complete article here, for the methodology see page 66)

The Bertelsmann Stiftung is one of the most influential think tanks in Germany. “The Bertelsmann Stiftung is a place where we look to the future without party-political boundaries and develop impulses for change.” Reinhard Mohn, Founder

Dr. Spuck, most patent analyses use either very broad technology fields or very specific patent classes to categorize the contents. EconSight therefore strikes a balance between these two approaches by developing specific technology definitions in order to capture the technological activities of companies, research institutions, regions and countries in the best possible way. For this purpose you use an AI-based application, our Patent Monitor. Can you explain why you have chosen Averbis?

Firstly, because we have already had very positive experiences with the results for industrial customers for several years, which go far beyond what can be achieved with cluster engines or other automatic categorization tools on the market. Secondly, because Averbis is a competent and “close” supplier with whom you can easily strive for solutions instead of discussing problems.

Where do you see the biggest challenge in the area of worldwide patent administration?

On the one hand, the mass, since the number of patent applications is increasing more and more and at the same time more and more patents are available in languages other than English, especially Chinese and Korean. On the other hand, the dynamics of technologies, triggered by digitalization. The classical classification is often too complex and too slow, which makes it difficult to bring meaningful thematic structures into the patent landscape. And without these, the patent world remains closed to non-experts.

Which potential do you see in the future in the field of AI to support the various questions of a patent analyst with software solutions?

We call it adaptive classification or categorization, i.e. categorization that is quickly adapted to requirements, we believe is the ideal solution for patent analysts. This has to be done on different levels, i.e. large topics, for example “robotics”, have to work as well as on an FTO level, for example “left-hand threaded screw for knee joints”. In addition, AI is increasingly being used to overcome the language barrier once and for all. AI will offer real added value if it can analyse technical similarities in patents on a large scale in order to make disruptive technological developments visible.

In your opinion, how will the world look in 10 years in the field of patents? Where will the development leading to?

Automated prior art searches, AI novelty searches and semi-automated FTOs will completely change the landscape of patent analysts and examiners at the offices. AI will not only do the research, but also the inventions themselves. AI will be used to make trends visible earlier, and new methods for automatic predictions will be investigated. Patent quality indicators will be available that are based on similarities in content, in addition to or instead of citations, and patents will become recognised indicators of the development and commitment of companies, e.g. in the climate and environmental sector (based on AI-supported categorisations). In addition, the use of AI will develop from patents to scientific literature and will significantly expand the search possibilities.

Thank you very much for this informative conversation and your valuable time, Dr. Spuck

• Averbis Health Discovery enables reliable de-identification of medical documents in accordance with the HIPAA Safe Harbor method.
• Sensitive patient information remain protected.
• Patient information can be used for medical research, quality assurance and clinical studies in a privacy-compliant manner.

Secondary use of routine medical data in research
The provision of clinical raw data is an indispensable basis for numerous applications in medical research: e.g. aggregated patient data can help to identify disease mechanisms, reduce recruitment times of patients in clinical studies or improve medication safety monitoring. However, such projects often fail because the raw data contain personal information and there is no way to quickly and  reliably deidentify large volumes of data. This is exactly where Averbis Health Discovery can help.

High need for protection of clinical data
Personal medical data is highly sensitive and subject to strict data protection regulations. All personal information must be removed before the data can be released for medical research. In addition to personal names and dates of birth, telephone numbers, names of medical staff and relatives, many other text passages are also part of the information that needs to be protected. Depending on the application, information such as dates should not be completely removed, but should be coarsened to one or more years for specific studies.

Best possible protection through a combination of AI and pattern recognition
Averbis Health Discovery supports you in the de-identification of personal data in medical free texts in compliance with HIPAA. In order to ensure the best possible protection of data, current technologies from the field of deep learning are combined with pattern-based procedures: for example, names, location information, occupational information and other features are recognized by means of artificial intelligence, the identification of e-mail address and date information is additionally secured by means of pattern recognition.

Individually adaptable to your needs
The marking of personal data and their further processing are logically separated in Averbis Health Discovery. Thus, you have the possibility to treat the identified characteristics differently according to requirements. You would like to obtain the patient’s year of birth in one study, and completely remove the year in another study? The de-identification is flexibly adaptable and provides reliable protection in every case of application.

De-identification fast and safe
The de-identification of Averbis Health Discovery is an indispensable tool for distributed medical research and finds its application wherever personal information is to be protected. It supports the data protection-compatible handling of medical documents in clinical studies, quality assurance and medical research. The de-identification as well as all other modules of Averbis Health Discovery are available for different languages.

For diagnosis confirmation and reimbursement optimization

For the clinics and hospitals under cost pressure, a complete coding of diagnoses and therapies is essential for the billing of services.  The completeness of the coding becomes a particular challenge when medical documentation from other areas such as the laboratory has to be taken into account.  The new version of Health Discovery addresses precisely this challenge.

In addition to clinical notes, surgery and pathology reports, microbiological reports  can now also be analyzed. Numerous DRG-relevant secondary diagnoses that were previously often overlooked can be recognized and coded.

View of a microbiology finding in the annotation editor of Health Discovery

Case study microbiology: 10% uncoded ICD secondary diagnoses

The new module has already been successfully tested in a large German hospital.  For this purpose, a total of more than 70,000 microbiological reports were analysed using Health Discovery. The result: more than 12,500 ICD codes were derived from positive microbiological findings. About 10% of these codes had not been coded by the hospital’s medical controllers so far.

Top 10 ICD-10-Codes (German version)

Assessment of CCL relevance

A comparison with the clinical complexity level CCL (Complication or Comorbidity level) showed that of the approximately 12,500 predicted codes, more than 7,500 represent CCL-relevant secondary diagnoses. Of the diagnoses not yet coded, a total of 825 were found to be CCL-relevant.

The new microbiology module can now be used by all Health Discovery customers and, through the partnerships with Cerner, is also available to all hospitals that use Meta-KIS as DRG-optimization software.