What if computers would understand human language in the same way people do?
A few systems, including Siri and Alexa, are already providing us with valuable feedback and information via voice. Yet, these voice agents’ applications are very broad, making them not suitable for medical-grade purposes.
However, the same logic applied to build these voice-activated solutions can be used to train specific systems for healthcare. A particular area where such a tool would be a significant step forward is radi-ology.
This is due to the vast quantity of digital data which makes its way into patients’ electronic health rec-ords and which could potentially help other people with the same condition.
Making Sense of Radiological Reports
Although there are some guidelines for how to write a radiological report, these are not set in stone. Usually, such a document consists of sections of free text, which is an unstructured format that is hard to analyze if not processed by machine learning through NLP.
Text mining applications transform electronic health records into a structured database which can be further used for a wide range of purposes. The good news is that once trained, an NLP algorithm can go through massive amounts of patient data at once. As most of this data contain patterns, the per-formance of the algorithm is usually very high.
How NLP Works
Starting with a free text report, the first stage is feature extraction. Similar to a human reading a book, the system first segments the text then performs boundary detection and word normalization. Once the text has been split into words, syntax and semantic parts, the last step is to identify negation which can change the meaning of the entire sentence.
The tokens resulting from this process are then processed either by following set rules or through ma-chine learning. The result is then compared with estimated outcomes and experts' opinions.
In the case of radiology, for excellent NLP results, a specialist needs to annotate data sources that will be used as input.
Advancements So Far
NLP for radiology has already found numerous applications which have been included in medical stud-ies and academic research. We will highlight further the primary use cases.
Diagnosis observation
There are some medical conditions including appendicitis, pneumonia, acute lung injury, and other ma-lignant lesions which can be easily overlooked by doctors. These are detectable on X-rays but can be easily missed if the focus of the investigation is completely different.
The advantage of using machine learning is that it can be fine-tuned to detect patterns associated with each of these diseases thus reducing the risk of overlooking a life-threatening condition.
Automatic patient selection
The ability to process electronic health records without human intervention can be used when creating cohorts for different medical studies.
Depending on the criteria required by the study, the NLP-equipped system can identify candidates with the studied conditions as well as suitable control groups.
Manual validation is not necessary but sometimes employed, although the sensitivity of the algorithm is usually about 95%.
This method has already been applied successfully to select patients with renal cysts, arterial disease, liver conditions, metastasis, and pulmonary conditions.
Such a solution yields far better results compared to traditional medical coding and offers the possibil-ity of selecting patients with comorbidities.
Another difference from traditional medical coding systems is in the fact that NLP can identify those patients who have specific symptoms without being classified with unknown illnesses.
This approach can help medical professionals create particular groups for study.
Database query
Since NLP classifies the data obtained as a result of converting free text reports, the structured data can be regarded as an indexed library.
As long as the system is trained with specific terms related to radiology, it can be used as a search en-gine for patient records.
Such an index is essential both for academic studies and data training purposes, as medical students can just ask the system to provide them with relevant images related to a particular disease.
Records quality evaluation
NLP can be used to evaluate the quality and consistency of medical records by comparing paper-based notes with those in EHRs. The topics of interest include the completeness of the report, the quality of therapeutic communication, case management, recommendations, and ongoing support.
It is essential that medical recommendations be consistent across the same age groups and associated conditions. This can potentially uncover significant inconsistencies in recommendations. For example, some radiologists performing abdominal imaging might recommended further imaging examination, while others might stop after the first test. Such a variation can compromise the quality of care deliv-ery altogether.
Support services
An NLP-based system with a library of patient records can act as a real-time assistant for the doctor who provides real-time feedback. Such a system can interact with the doctor through voice and can be incorporated in the workflow more easily.
Another support service may be in automatically transforming free text reports into highly structured data by using text analysis software. This can then be indexed, searched or retrieved as necessary. Furthermore, structured data can then be mapped to healthcare codes for automatic invoicing and insurance purposes.
NLP for Radiology- Does It Work?
Radiology is a healthcare area dominated by images and free text observations. This makes it difficult to get searchable, indexed information at hand.
NLP can solve this problem by automatically converting unstructured data into structured reports.
Applications for this are numerous and include identifying hidden conditions by scanning medical ob-servations, selecting the right patients for clinical studies based on a set of characteristics ,as well as doctor support services and evaluation of medical records consistency.