Unlock the Power of Unstructured Patient Data
The Talix platform powers intelligent workflow applications that enable risk-bearing healthcare organizations to succeed in the age of value-based care.
Our workflow solutions for payers and providers require intelligent underlying technologies to work in unison and at scale. We’ve engineered the Talix Platform to support the needs of thousands of end-users, anywhere in the world simultaneously. Moreover, our platform architecture enables multiple SaaS application solutions in order to harness the efficiencies derived from being able to process millions of patient charts and encounter data.
The Talix Platform is comprised of several technology components, intricately linked, to power software applications at scale for healthcare payers and providers. These components form the building blocks of artificial intelligence (AI). AI requires large sets of data. In healthcare, much of the data is found in the unstructured text (think, for example, physician notes on patient diagnosis and treatment). In fact, 80% of patient data today is in unstructured format, whether it’s documented in a provider office visit, telemedicine, analysis of lab or imaging result, discharge summary, or surgical notes.
Understanding and analyzing healthcare information at scale requires highly sophisticated technologies purpose-built for specific objectives. For Talix, these technologies support the enabling of value-based care, from financial (e.g., HCC coding for Medicare Advantage, ACA, Medicaid) to quality measurement (e.g., HEDIS measures) to clinical decision support (e.g., care gaps to close).
OCR performs an important role in turning images into machine-readable text. Talix has partnered with an OCR engine to develop its sensitivity to healthcare text and how patient charts are variously organized (e.g., forms, templates, headings, etc.). Years of Talix engineered refinement has yielded a market leading OCR application for healthcare
Our EZ Chart process sits between the Talix Platform and the Talix Applications it supports. Its function is to enable PDF readability and accelerate the speed for end users to navigate original documents. This invention is a Talix trade secret. Please contact us directly for more information.
4 Core Technologies
The patent-pending Talix Platform has 4 core technologies—each designed for scalability, maximum efficiency, and accuracy.
1 – Our Healthcare Taxonomy represents all concepts associated with the clinical and non-clinical nature of healthcare, and how these concepts are semantically related to each other.
2 – Natural Language Processing (or NLP) possesses the ability to put text and structured data into context so an end-user can interpret results accurately.
3 – Clinical Rules express specific criteria or guidelines adapted to the practice of medicine.
4 – Machine Learning automatically reads patterns in the prepared healthcare data set that has been organized, prepped, and contextualized after the application of OCR, Taxonomy, NLP, and Clinical Rules. The findings are fed back into the technologies to improve results.
High-Speed Data Connectors
Built into the Talix Platform are high-speed data connectors that provide our customers up to real-time access to patient encounters and chart data. Our cloud-based scalable platform can process thousands of individual charts daily. Thus, Talix can support peak load customer requirements against stringent deadlines. Pre-built adaptors enable customers to access a wide range of patient data assets, including clinical, claims, membership, and provider data.
Our customer deployment experience covers a wide range of patient data sources: PDFs, CCDs, CCDAs, Lab and Prescription data, Claims and RAPS files, and HL7 messages.
In fact, Talix workflow applications are embedded into Electronic Health Records (EHRs) and Electronic Medical Records (EMRs) at the point of care where medical professionals (e.g., nurses, doctors, other allied medical professionals) want access to suggested HCC codes, for example, that may have been overlooked.
Talix Taxonomy and NLP Overview
The Talix Taxonomy maps out over 1,000,000 health concepts and over 2,000,000 relationships and is constantly updated, expanded, and refined – making it the most comprehensive map of healthcare concepts in existence today. This twenty-year investment provides our customers with real-time access to healthcare knowledge as it evolves. Our automated adaptors fetch the latest billing codes, medical concepts, clinical rules, etc. that are produced by tens of organizations around the world. A physician-led supervised machine learning environment processes newly acquired concepts to rationalize the induction of concepts into our semantic taxonomy.
The Power of Symantics
The above example on Chronic Kidney Failure (also referred to as CKF, CKD, and CRF) helps explain the power of semantics underlying the process of what we call Natural Language Understanding (NLU). NLU is a subset of NLP, and is critical to understanding the meaning of healthcare concepts in the text. For independently sourced information on this complex topic see the Wikipedia page on NLU.
The power of semantics. Semantics is a branch of linguistics focused on assigning meaning to words or concepts. Simply put, CKD in this example is assigned a semantic type called “disease state or syndrome.” The surrounding attributes in the pale color express the various attributes of this disease state concept. This is important for accurately identifying concepts
in unstructured text. The outer border of relationships in blue defines the context this concept has in the world of healthcare. For example, we now can identify all of the symptoms associated with CKD. Conversely, we can take any single symptom and understand its relationship with disease states. This semantic two-way bridge between any two concepts helps us discover the context or meaning of that concept in a particular patient chart.
As shown in the image below, a patient chart can be processed instantly to identify the contextual relationships between any concept and it’s semantic type and the corresponding medical codes.
For more details on Talix’s core technologies, please visit the sections below:
What is the difference between the Talix Taxonomy and off the shelf terminology or data dictionary sets in use today? Some companies trying to develop NLP and AI capabilities borrow or rent off the shelf healthcare taxonomies from academic institutions. We decided against that and instead invested millions of dollars and thousands of hours by medical doctors and healthcare data scientists to build a powerful, proprietary solution that we can customize for a broad array of healthcare use cases. For example, finding applicable HCC codes in a thousand page patient chart covering three years of patient history can be accomplished in milliseconds. In addition, our Talix Taxonomy is continuously updated to address the ever changing aspects of healthcare without having to rely on third party sources. We integrate multiple terminologies representing different semantic types and interconnect concepts amongst these types. For example, we capture NDC and RXnorm for drugs, ICD10 and SnoMed for conditions, and LOINC for labs, and so on. Talix employs software, medical doctors, and scientists to build accurate relationships across concepts.
Sitting on a proprietary, configurable clinical rules database, the Talix Clinical Rules engine is tightly integrated with the Talix Taxonomy to enable condition and treatment inference based on medical evidence found in both structured and unstructured patient information. Depending on the use case, for example HCC coding, the set of clinical rules to be followed are customized to what is permissible. As an example, in HCC coding, conditions treated in past years, which are not chronic, should not be considered. Also, conditions that are documented during a non face-to-face encounter are not taken into consideration. In order to take care of the above two issues, the coding software needs to understand the associated dates for each condition identified and also understand the document type in which the condition is mentioned. Another example is understanding the medication dosage that is prescribed by the physician, or a lab value associated with the administered test. Depending on the dosage, the condition severity can be determined. Similarly, if the lab result is high, the lab value would indicate severity.
Another key component supporting Artificial Intelligence is Machine Learning (ML). The Talix Machine Learning framework employs a vast array of accumulated clinical and non-clinical input data to recognize data behavior patterns across large patient data sets. We call these “recognition models.” We refine these models into a self-learning system that continuously integrates user feedback to achieve higher and higher levels of accuracy.
Accuracy is derived from two measurements: precision and recall.
Recall represents how well the system captured all true results.
Precision measures the percentage of results that were produced that were true.
The combination of the 2 measures represents accuracy. Accuracy below 100% typically means that the system has identified false positive and false negative results along with true results. The goal of ML is to reduce the number of false results.
The Talix team gained first exposure to Artificial Intelligence techniques in the late 1980s through projects in the financial services industry. By the early 2000s, Talix engineers and healthcare data scientists came together in a Talix predecessor company to develop AI components strictly for the healthcare industry. Today, our experience in applying healthcare AI techniques covers a broad swath of healthcare, including consumer health, oncology, medication impact on chronic conditions, medical professional content publishing, and so on. Much of the work involved projects completed for the Fortune 100 of healthcare.
ML encompasses a number of analytical techniques honed by world leading computer scientists and mathematicians. Talix employs a number of these techniques in the analysis of large patient data sets.
The features that contribute to ML accuracy are aggregated over a large number of patient encounters and charts. These features expand over time to increase the accuracy of results. Features may include the exact text representing an ICD10 code, the section of the chart where the code is found, the document type (consult note), evidence of active treatment, and many more.
Our approach to ML is customer-centric. The Talix ML model takes learnings from across all of our customers and applies learnings to all users. At scale, Talix processes millions of patient encounters, which is fundamental to finding statistically relevant output to improve the precision component of accuracy. As depicted in the exhibit below, our customers are directly involved in understanding the value ML provides. For example, our Coding Insight customers provide input automatically into a live ML feedback system. Clinical findings that are not identified as accurate are separated into false positive and false negative findings. The Talix Platform ingests these inputs to be processed through “recognition models” on a test basis. Positive impacts on precision and recall are then sent back to the Talix Platform to train and improve the software.
The Talix Platform is built to serve payers and providers who want to shift to the value-based care world. This shift relies significantly more on leveraging patient information to conduct business, whether it be alternative payment models such as HCC coding, measuring quality such as HEDIS, or identifying misdiagnosed patients and organizing care management. Software technology that can understand the context of patient clinical and care management details will make healthcare delivery and financing ecosystem remarkably more efficient.
Here are 5 helpful tips in evaluating a healthcare technology platform powering workflow applications that will change the way you conduct business.
1 – Is the platform able to support workflow applications serving multiple environments? For example, can you deploy one solution to enable all employees, no matter their location, as well as third party partners?
2 – Can the platform scale? Meaning, for example, will performance be adversely affected if you increase throughput volumes by a magnitude? Similarly, how much time is required to fetch, process and deliver patient chart analysis? Is it hours, days, or weeks?
3 – Is the platform flexible and configurable and not beholden to several third party licenses? In other words, how quickly can customer feedback be incorporated into the platform?
4 – Can you run the platform through a test in hours or days, not weeks or months? Many purported AI companies use behind the scenes humans to curate output. This takes time.
5 – Look for examples on how AI training sets are formulated, how custom feedback is incorporated into the Machine Learning models, Technology platforms using AI components should not be mysterious or be viewed as obscure rocket science. Regardless of the advanced mathematics employed and the number of PhDs working on the technology, the solution needs to perform, and be understood by the business owners. Talix delivers.