Meaningful Healthcare NLP
Preparing for the next normal: how healthcare organizations can grow through AI-driven transformation
As much as 30% of the entire world's data stems from the healthcare industry. And each year, a single patient typically generates close to 80 megabytes in imaging and electronic medical record data. This trove of data holds clinical and operational value for health care providers, life sciences companies, and medical insurers. But the value of healthcare data is realized only when it can be analyzed at scale and converted into knowledge that influences everyday practice.
The failing of health information technology
Scaling healthcare data analysis is a complex, labor-intensive task when it comes to unstructured data like text, images, or videos. In today's healthcare ecosystem, teams of medical coders and auditing specialists analyze unstructured medical documents primarily by hand. This process is error-prone and does not scale. But automated document review tools either do not meet standards of excellence or cannot be easily integrated into the intricate healthcare IT environment. And healthcare professionals cannot ignore unstructured data. First, unstructured data is a common data format rich in medical insights, with as much as 80% of healthcare data being unstructured. Secondly, information about mental health, medical symptoms, or date of diagnosis can only be ascertained from medical documents. In theory, health information technology like electronic health records should expedite clinical data review and remove operational bottlenecks. In practice, many physicians report that using electronic health records adds to their daily frustration level and leads to strain and burnout.
Alternative paths forward
One can argue that the medical insights stored in the unstructured text should also be available in a structured format that computers can analyze at scale. The healthcare industry is not shy of data storage standards. Since the late 1980s, the healthcare industry has nurtured global health data interoperability standards like Health Level Seven International (HL7) and Fast Healthcare Interoperability Resources (FHIR) that facilitate storage and exchange of medical data in a structured representation. However, these data standards are not widely adopted. In a 2019 Deloitte study, healthcare executives stated that interoperability will be crucial to their organization in 3-5 years compared to today. Interoperability is a long-term journey, commonly steered by regulations mandated by the government.
There are several open questions left to be addressed on the interoperability journey. Specifically, how does an organization populate the HL7 structured fields? And how do you best harmonize information from HL7 that colludes with insights from the unstructured medical record? Without an automated solution for mapping insights from disparate data silos into the HL7 standard, only limited clinical information will be reflected in structured HL7 fields. And even in a future where interoperability standards become widely adopted, healthcare stakeholders still need automated solutions to glean and harmonize meaning from unstructured data at scale.
Teach machines to read like humans do
AI and machine learning are a promising choice to enable the automatic extraction and harmonization of knowledge stored in unstructured text. We know too well that humans communicate through language, while digital computers communicate in binary code. To bridge the divide between how humans and computers communicate, we need a human-machine interface that can disambiguate the nuances of human language into a systematic representation coded for machines. Machine learning tools like natural language processing (NLP) can translate human language into machine-readable setups and assist healthcare professionals with the review and analysis of medical documents in a repeatable, scalable way.
Intelligent healthcare applications built on NLP engines can have a substantial impact at the system- and human-level. System-wide, they enable curation, translation, and harmonization of medical insights for large scale analysis and conversion of insights into knowledge that influences everyday healthcare practice. Such healthcare NLP applications can also reduce workforce burnout and increase healthcare productivity across the healthcare ecosystem.
Drug discovery: enable faster drug development by deploying NLP to search through troves of scientific reports and published literature and characterize chemical compounds within a specific therapeutic context
Clinical trial management: streamline the operation of clinical trials by deploying NLP to translate clinical trial protocols and patient medical records into a structured representation that can be used for optimal patient matching
Real-world evidence: improve quality of care through evidence-based medicine informed by patient-reported outcomes such as adverse effects that are retrieved with NLP from observational data obtained outside the context of randomized controlled trials
Screening and diagnostics: improve the specificity of disease screening tools with information extracted by NLP from medical records. For example, the Rosner-Colditz breast cancer incidence model relies on risk factors (history of lobular carcinoma in situ, hyperplasia, or ovarian cancer; age at menopause; breast density) that are typically stored in unstructured clinical notes and pathology reports
Utilization management: understand and monitor how health plan members receive and utilize healthcare services. The holy grail in utilization management is a coordinated, member-centric care management focus. Healthcare NLP can surface insights from medical records for better and faster identification of shortfalls in adherence, compliance, and evidence-based care
Case management: document, oversee, and automate healthcare delivery through accurate information exchange between payers and providers. NLP enables the review of information in claims and medical records for automated pre-authorization and claims adjudication
Virtual care delivery: employ NLP conversational solutions to engage members in their own health. Personalized digital health solutions can drive patient understanding of chronic conditions, enhance adherence and compliance, boost self-care, and avoid more costly treatments
Administrative workflow automation and digitization: automate workflows such as pre-authorization, referrals, lab/prescription orders that could lower the overall administrative cost by $24 to $48 billion annually through productivity gains
Quality care measures reporting (HEDIS, MIPS): compile and report quality care performance data to provide healthcare purchasers and consumers with the information they need for reliable comparison of health plan performance
Risk adjustment: deploy NLP to help estimate future health care costs for patients and effectively underwrite patient populations
While healthcare NLP tools can introduce workflow efficiencies across various touchpoints in the healthcare lifecycle, this technology requires specialization to the task for high-fidelity results. In a production environment, the internal architecture of an NLP engine for clinical trial matching will look significantly different from the NLP purpose-built for virtual care delivery. Healthcare NLP cannot "read" like a primary care doctor just yet. Still, it can excel at laser-focused tasks and augment the productivity of the healthcare professional.
