“We Talk, They Listen” - Paving the Way for AI Fluency in Healthcare
by Ioanna Deni
For more insights on the intersection of AI and healthcare, visit our website and subscribe for regular updates.
The importance of human language
Unlike other forms of natural animal communication, human language is compositional. It owes its uniqueness to the construction of sentences with subjects, verbs, and objects, as well as past, present, and future tenses. It is also referential, allowing speakers to exchange specific information about people or objects. Animal "languages," on the other hand, lack compositionality and are often limited to simple associative learning. The precise beginning of human language remains a challenging question but limited evidence of symbolic thinking in Neanderthals suggests that language likely emerged only 200,000 years ago [1].
In healthcare, language plays a crucial role. It is the vehicle through which knowledge and information are shared, and it is a means by which patients can express and communicate their needs. Medical terminology is important because it is the language that healthcare providers use to discuss diagnosis, treatment, and prognosis. It has become a standardized language to reduce the risk of errors, particularly when documenting medical conditions [2].
There are two significant problems associated with medical language. Firstly, it often proves challenging for patients to comprehend, which can be both frustrating and humiliating [3]. Secondly, the extensive need for meticulous record-keeping results in a vast quantity of patient forms and clinician-written notes usually stored as Electronic Health Records (EHRs). These data are immensely valuable for improving health outcomes but are often challenging to access and analyze.
What is a natural language processing algorithm?
It's hardly surprising that the next significant advancement in human language lies with the development of natural language processing algorithms (NLPs). NLPs can efficiently navigate the wealth of medical language and extract meaningful insights for patients and healthcare teams.
NLPs are specialized algorithms that can understand and interpret human speech. In simple terms an algorithm cleans and organizes the human speech, then it breaks it into smaller units or "tokens" through tokenization and then an NLP system interprets the text. There are 5 main types of NLPs [4]:
-?????? Optical Character Recognition: convert handwritten or printed text into digital formats, making it valuable for digitizing clinical notes, medical records, and other healthcare documents.
-?????? Named Entity Recognition: segment subjects like people, places, organizations, or products into predefined categories.
-?????? Sentiment Analysis: assess the underlying sentiment in text, often applied in healthcare for patient feedback.
-?????? Text Classification: assigns tags or labels to text based on predefined categories, useful for identifying at-risk patients based on specific keywords.
-?????? Topic Modeling: groups documents based on common words or phrases to identify semantic structures or "topics."
There are also the large language models or LLMs. These algorithms belong to a class of even more complex machine learning models that are designed for a wide range of NLP tasks such as classification, or even text generation. The LLMs train using vast datasets and can learn to predict the next token and eventually generate the tokens themselves resulting in producing language.
If this sounds familiar, it is because you have used GPT-3 from Open AI. Other famous LLMs are BERT from Google and Megatron-Turing from NVIDIA.
The reason LLMs are great at generating language is because they use a transformer architecture to understand the relationship between tokens (= words in a sentence). This means that instead of evaluating the relationship of tokens based on recurrence they give an attention score based on the importance of the word in a sentence [5].
LLMs can also be great medical assistance, helping both patients and healthcare providers. Join us on an exploration of the booming world of NLP in healthcare and life sciences, where the global market is expected to reach a value of $7.2 billion by 2027, driven by a remarkable CAGR of 27.1% [7]. Let's unravel the mysteries behind this captivating AI-driven sector and discover what's fueling the excitement!
Healthcare NLPs in action
Cutting-edge LLM and NLP models have evolved into the central components of contemporary health search engines, voice assistants, and chatbots. Their intelligence is continually enhanced through training on real-world data, encompassing both scientific and everyday language. Their use in healthcare can be divided into the following main categories:?
LLMs & NLP for clinicians and hospitals
In the clinical sector LLMs and NLP technology can revolutionize clinical documentation, medical decision making and enhance the Electronic Health Records (EHR). By integrating NLP into an automatic speech recognition system, clinicians’ notes can be taken in real-time during conversations and mapped to a standardized medical ontology. This real-time medical transcription saves clinicians time and ensures comprehensive documentation. Additionally, NLPs rapidly and efficiently access health-related information, by identifying, categorizing and data mining the appropriate data [8]. LLMs can interpret the EHR and predict a diagnosis with high accuracy while they are also able to mark an incorrect diagnosis 16 out 0f 17 times [9].?
Multiple key players in the clinical documentation and diagnosis are employing NLPs:
-?????? The GE Healthcare EDISON NLP is designed to extract meaningful insights from unstructured clinical data that can help advance healthcare solutions that comply with HIPAA regulations.
-?????? Roche harnessed the power of Spark NLP to decipher clinical terminology, extract specialized facts, and leverage transfer learning to reduce the need for task-specific annotations.
领英推荐
-?????? Mount Sinai deployed real-time NLP pipelines to input different clinical documents, extract information from them, and generate different risk stratification scores for oncology and psychiatry patients in outpatient and inpatient settings [10].
LLMs are growing in popularity with leading player Med-PaLM 2 that harnesses the power of Google's LLMs. Mayo Clinic is currently trying to use Med-PaLM 2 in multiple hospitals as a medical assistant for its staff [11].
LLMs & NLP for pharmaceutical and biotech
In the pharmaceutical sector LLMs and NLPs can extract valuable insights from scientific articles and propose knowledge gaps. Pharmaceutical companies are already utilizing LLMs and NLPs for various applications, including drug discovery, or drug repurposing, and clinical trials [12].
-?????? Researchers at Merck used Linguamatics NLP and other tools to discover potential novel biomarkers for diabetes and obesity, from PubMed, clinical trial data, and internal Merck research documents [13].
-?????? Novartis used NLP models to combine all its historical clinical trial protocol records and data sets to enable scientists to ask questions and perform specific inquiries into disease areas that they previously lacked the capability to execute [14].
-?????? NVIDIA deployed large biomolecular language models (they can have their own acronym LBioLMs) such as MegaMolBART that can demonstrate cheminformatics applications in drug discovery and find novel therapies for existing diseases [15].
These advances are particularly interesting in rare diseases with no current treatments plans.
LLMs & NLPs for patients
NLPs and LLMs are often used in the form of chatbots such as GPT-3 and they can answer a wide range of medical queries and provide personalized information. More specialized chatbots can help with scheduling appointments, and offer simple at-home recommendations, especially in regards to mental health (spoiler alert for our next week’s article!).
Breaking down the current NLP/LLMs barriers to adoption
While LLMs and NLPs offer great promise in healthcare, they face various limitations that can hinder their wide deployment and adoption. These challenges include:
Data Access and Quality: training these tools requires access to appropriate and high-quality data, which can be limited. Availability and curation of datasets that are comprehensive are essential for successful language model deployment.
Bias and Health Disparities: Failing to account for biases during the development and deployment of NLP/LLM models can lead to unintended biases in model outputs. These biases can perpetuate health disparities and negatively impact patient care. Addressing bias is a crucial consideration when using language models in healthcare.
Privacy Concerns: Regulations governing data use and privacy protections for language models are yet to be fully established. This raises concerns about the privacy and security of healthcare patient data.
Framework and Guidelines: LLMs and NLPs require robust frameworks and guidelines to ensure they function as intended. The absence of well-defined frameworks and guidelines for their deployment in healthcare presents a challenge in achieving consistent and reliable results.
Evaluation Methods: Language modeling tools are typically evaluated on word-, sentence-, or document-level annotations, while clinical research studies operate on a patient or population level. Bridging this gap and defining appropriate evaluation methods for NLP-driven medical insights is a significant challenge.
Performance Standards: NLP tools relying on speech recognition software can have up to 7% errors thus human review and manual editing may be necessary to ensure data accuracy and patient safety.
These limitations, however, are driving research and innovation to overcome them and maximize the potential of LLMs and NLPs in healthcare [16]. At Intelligence Ventures, we actively seek out companies that tackle these challenges head-on, demonstrating excellence in addressing them and effectively deploying LLMs and NLPs in healthcare.
The 4 takeaways for NLP investing
Language modeling technologies have already found extensive applications across various industries, and their integration across the healthcare sector is inevitable. At Intelligence Ventures, we are dedicated to supporting companies in disrupting the healthcare industry and crafting unique innovations that will shape the future. Our key takeaways include:
Innovation: We seek advancements that propel LLM/NLP to new heights, making it faster, smarter, and responsible. Does it require extensive datasets for training, and can it address inherent biases? We are interested in nurturing innovations that have transformative potential for their capacity to revolutionize healthcare.
Scalability: LLM/NLP technology should not only handle the exponential growth of electronic medical records but also facilitate the swift adoption and seamless integration into clinical workflows. It should efficiently process data from diverse sources in a reproducible manner, offering clear advantages and demonstrable improvements in its insights.
Interoperability: We are keen to explore companies that excel in the aspect of connectivity amongst the different systems of data input while they ensure data privacy. LLM/NLP can play a pivotal role in enhancing communication between systems but how does the technology prevent information loss and preservation of clinical intent while upholding data privacy?
User Base: LLMs/NLPs should add value to both the healthcare professionals work and in the patient experience. We assess factors such as user-friendliness, cost-effectiveness, and demonstrable improvements in outcomes for both user audiences.
At Intelligence Ventures, we strategically invest in the next wave of AI-driven healthcare and we support innovations that drive LLMs and NLPs adoption forward, fostering significant transformations in the healthcare landscape. If your LLM/NLP-focused venture aligns with our values and objectives, we eagerly invite you to join us in the new era in healthcare.
References
PhD, Senior ML Engineer @ Sopra Steria
1 年David Pasdeloup
Venture Capital | Investing in the intersection of healthcare & AI
1 年More on NVIDIA's LLM Developer Day here: https://www.dhirubhai.net/posts/chelsea-sumner-pharmd-rph-34199757_nvidia-llm-developer-day-activity-7128080243185045504-Nq9J?utm_source=share&utm_medium=member_desktop