What is natural language processing in healthcare?
Natural language processing (NLP) is a branch of AI that uses algorithms to extract meaning from unstructured human language presented either in spoken or written format. NLP requires knowledge of computational linguistics and other machine learning skills.
Natural language processing is gaining popularity in healthcare as it can analyze large quantities of unstructured medical data, such as doctor notes, medical records, clinical trial reports, and even service reviews that patients post on social media. Research shows that about 80% of healthcare data is unstructured and not utilized to its full potential. NLP can change that by analyzing data and extracting insights to guide doctors and pharmacists to make more informed decisions.
Here are some of the main NLP techniques that are useful in healthcare settings:
Optical character recognition (OCR). OCR converts printed and handwritten text to a machine-readable format. It can extract text from images and tables and pass it to other NLP algorithms for further processing. If you want to learn more about the topic, check out our recent guide on optical character recognition.
Text classification. This approach involves assigning semantic labels based on predefined categories. For instance, it allows doctors to label a patient as “at risk of hospitalization” based on certain keywords in their medical notes.
Named entity recognition. This is an information extraction model that can pinpoint entities in text. Physicians can use this technique to extract entities, such as “treatment” and “symptoms,” from a heap of medical data.
Topic modeling. This technique can sort information based on common topics. For example, it can group together all doctors’ notes on patients suffering from a particular disorder.
Relationship extraction. This NLP method can establish semantic relationships between entities. For instance, it can point out that a particular patient was treated at that hospital under the supervision of that doctor.
NLP use cases in healthcare
After learning about what healthcare NLP is and how it works, let’s see how it can contribute to the healthcare field.
Clinical documentation management
Research shows that physicians spend 16 minutes on EHRs on average for every patient they see. This time could have been spent caring for patients. Instead, it’s wasted on the screen and contributes to doctor burnout.
Enhancing clinical documentation through speech recognition
Natural language processing in healthcare can help physicians work with electronic health records (EHRs). Doctors can use speech-to-text conversion tools with built-in NLP capabilities to transcribe their notes and enter them into the corresponding patent’s EHR fields. Also, medical staff can query the NLP tools to extract relevant data from EHRs.
For instance, Nuance’s Dragon Medical One solution is a cloud-based speech recognition tool that helps medical personnel document patients’ stories in their EHRs. The company claims that deploying their solution will cut time spent on clinical documentation by half. Concord Hospital deployed Dragon Medical One, and 75% of employees reported more accurate clinical documentation.
Supporting doctors in decision making
NLP solutions can also analyze clinical documents and support physicians in real-time decision making. For instance, after analyzing doctor notes, the system can predict hospital bed demands, which gives hospital staff time to prepare and accommodate incoming patients.
Medical coding and billing
Medical coding means deriving billable information from clinical notes and transferring it into standardized medical codes. Traditionally, a human coder would perform this task. But manual coding is slow and prone to error, meaning that the provider might not be able to claim and receive the full amount in compensations.
NLP-powered computer-assisted coding (CAC) tools can derive information from doctor notes and patient EHRs about different therapies and procedures they encountered, and provide the corresponding insurance codes to solidify claims.
One example of CAC natural language processing in healthcare is 3M 360 Encompass System. The company employs over 150 coding experts and remains up to date with the newest regulations. This solution assists coders by aggregating and analyzing patient documentation, offering auto-suggested tags and comprehensive review and approval tools.
Clinical trial management
Studies reveal that around 80% of clinical trials are delayed or terminated because they can’t recruit enough participants. Medical NLP can speed up recruitment by scanning patients’ medical data in search of eligible trial candidates. The technology can also help people who want to participate in clinical trials. It can transform eligibility criteria into questions, allowing candidates to filter through many trials fast.
Headquartered in Canada, Inspirata built a solution that employs natural language processing in healthcare to facilitate patient recruitment. Its NLP engine mines clinical documents and transferred unstructured data into structured insights on patient eligibility.
Another successful NLP platform, Linguamatics NLP, can not only help recruit patients but also assist in trial design and site selection. It can parse a variety of sources, such as news feeds, patents, medical literature, and similar trial descriptions.
Refer to our article on AI in clinical trials to discover how other subdisciplines of AI contribute to this field.
Patient sentiment analysis
Applying NLP in healthcare helps medical staff aggregate and analyze customer reviews and opinions from different social media platforms. Natural language processing tools can process thousands of reviews to gauge patients’ attitude towards the care they received. Such tools can identify aspects that frustrate patients, segment them based on frequency, and start implementing improvements for the most recurring negative feedback.
For instance, a comment such as “I waited for 30 minutes on the phone to talk to the receptionist” indicates a negative experience with the administrative staff.
Another way of benefiting from healthcare NLP in sentiment analysis is processing survey results. Instead of relying on social media, healthcare facilities can compose targeted surveys, distribute them among patients, and use NLP solutions to automatically review every response.
In a recent study, a team of researchers built an NLP solution that can scan survey results to determine how patients feel about their healthcare provider. The team designed a survey with open-ended questions, distributed it on vestibular schwannoma patients, and received 534 responses. The algorithm classified the responses into two groups — positive and negative sentiments. This classification had a 90% overlap with the manual classification produced by human experts.
Check out our blog article if you want to learn more about how sentiment analysis can boost your customer service.
To find effective candidate drugs, scientists need to identify the biological origins and understand the disease first. This includes reviewing large amounts of medical literature, patient data, and more. Natural language processing in healthcare can quickly sift through all this information, extract what is relevant, and present it to researchers in a compelling format so that they can learn about similar diseases and how they were treated.
For instance, Texas-based Lymba offers an AI-powered software that can assist in drug discovery by applying NLP to analyze unstructured data from multiple sources, such as gene cards, PubMed publications, proprietary global research data repositories, clinical trial documentation, and more. After absorbing the available disease-related knowledge, the tool builds an ontology of existing drugs that highlights promising research areas to discover new effective drugs.
Additionally, NLP can analyze social media feeds and patients’ medical records to gauge adverse effects of different drugs. Some tools can even correlate a drug’s side effects to the dosage consumed and the frequency of use.
For more exciting applications of AI in drug discovery you can refer to our blog.
Challenges of NLP implementation in healthcare
Applying natural language processing in healthcare has many benefits. But if you decide to enhance your practice with this technology, here are the challenges that you are likely to encounter.
Specific language requirements of the healthcare field
Doctors have their specific way of writing clinical notes and other medical documents. This style is typically heavy with jargon, acronyms, and abbreviations, and has limited context. To add to the confusion, a third of the clinical abbreviations present in the Unified Medical Language System Metathesaurus have several meanings each.
Even words that are spelled out in full can be confusing. For instance, “discharge” can either mean release from a hospital or bodily fluids, depending on the context.
And if you want to extend the application of NLP in healthcare beyond your clinic and use it to analyze patient’s feedback on social media, then you need to familiarize it with the specific abbreviations and emoticons that people use to express their emotions instead of words.
If you acquire ready-made general-purpose NLP algorithms, you will need to retrain them to operate in healthcare settings. And even better, familiarize them with the vocabulary used by your doctors.
Issues related to human language complexity
It’s relatively easy to deploy NLP algorithms to spot terms that are explicitly mentioned and unambiguous. But human language is not always that straightforward. Sometimes, algorithms will need to deal with the following:
Inferences. Natural language processing models in healthcare have to be able to infer information from the input text. For example, if you want to know whether a patient has social support, there might not be a clear mention of this in their medical notes. However, there might be something like “relative at bedside,” that implies the patient has a form of support.
Semantics beyond keywords. A basic keyword-driven NLP will perceive the two statements “wife helps patient with medication” and “patient helps wife with medication” as identical, while in reality the semantics in both cases are very different.
Negation. Clinicians often use negation to indicate absence of medical conditions. For example, a doctor might write “a brain malignancy was ruled out” to document the fact that a malignant brain tumor wasn’t found. Medical NLP algorithms can’t mistake this with the presence of the disease.
Decide on how advanced your algorithms need to be based on the target business cases. And if you opt for more basic solutions, be aware of their limitations.
Bias and explainability
Like any other AI technology, NLP in the medical field can acquire different types of bias during training on unbalanced datasets or when it continues to learn on the job. Gender bias is one of the most prominent types in natural language processing-based tools. For example, GPT-3 tends to associate males with jobs that require high-level education, such as a doctor, while females are connected to less knowledge-intensive occupations, like a nurse.
Unfortunately, AI bias is not uncommon in healthcare. For example, an algorithm that was supposed to determine the severity of patients’ illness significantly downplayed the level of care required for black patients just because, historically, more money was spent on white patients’ needs.
It can be hard to spot bias when the algorithms are “black box” models that don’t explain how they arrive at their conclusions. One solution to this issue is to deploy explainable AI. With this technology, healthcare natural language processing tools justify their recommendations, allowing you to verify whether they are biased or not.
Another way to spot bias is using a curated dataset that scientists designed to detect specific types of bias. This solution is not scalable to large applications but works well with limited usage.
In addition to the three NLP-specific challenges presented above, you might encounter general obstacles associated with implementing any AI-powered technology, such as:
Integration with legacy systems. Many healthcare organizations still rely on outdated legacy systems that aren’t built to work with AI-driven technology and deal with large amounts of data.
Insufficient training data. Medical facilities struggle to provide sufficient, high-quality data that faithfully represents the target population.
Ethical considerations and moral hazards. Hospitals struggle with the responsibility that comes with using AI, such as who has the final say on selecting the right treatment, and who is responsible if the treatment didn’t work.
For a more detailed explanation, check our article on top AI implementation challenges.
Step-by-step implementation of natural language processing in healthcare
If you want to use NLP in healthcare, you can follow these steps:
Identify potential use cases. Determine which problems you want to solve with NLP and whether you need advanced language processing capabilities, such as negation and inference.
Build/buy an NLP solution. Decide if you want to acquire a ready-made NLP tool or build your own customized product. Off-the-shelf solutions are suited for companies with limited funds and flexible internal processes that can be adapted to accommodate an external product. Consider a custom-built solution if you have rigid workflows, are looking for something scalable, and need to integrate the new tool with legacy systems. In the latter case, you will need to look for a trusted AI healthcare solutions company to collaborate with.
Prepare the training dataset. Even if you opted for a ready-made solution in the previous step, you will still need to retrain it for optimal performance, as mainstream NLP models are not designed to handle the specifics of the healthcare sector. You will need to prepare a training dataset that is bias-free and representative of your target population.
Train and validate the model. Test the tool for bias, usability, and how it integrates with the rest of the workflow.
Make sure your algorithms are compliant. If you want to build and/or adopt AI models in the medical sector, compliance is of utmost importance. Your tools need to comply with the US Food and Drugs Administration (FDA), the General Data Protection Regulations (GDPR), and any other regulatory bodies in the countries of your operations. You can find more about the topic in our recent article on healthcare IT standards.
Integrate with your existing systems. If you opted for a custom healthcare solution and hired a tech vendor, they will design and build the NLP tool with the specifics of your system in mind. They will also help you integrate the solution with your legacy systems.
Monitor and adjust the algorithms constantly. The effort doesn’t end at deploying natural language processing in healthcare. You need to audit the NLP models to make sure they are still producing the desired results and update them to reflect any changes in the target use cases.
NLP has many applications in healthcare. It can help save time on medical coding, facilitate recruiting participants for clinical trials, keep patients happy through sentiment analysis, and much more.
To successfully deploy natural language processing in healthcare, dedicate time to finding a trusted machine learning vendor who has experience in the medical field. A tech partner will help you with:
Creating and training custom NLP models that:
Are free from bias
Fit seamlessly with the rest of your system
Tailor to your practice’s vocabulary and have the desired degree of complexity
Complying with the privacy requirements of the healthcare sector
Supporting you in auditing and updating the algorithms when needed