Search the KHIT Blog

Wednesday, May 31, 2017

Continuing with NLP, a $4,200 "study."

OK, I've been keyword-searching around on Natural Language Processing (NLP) topics in the wake of my prior post, while I finish Daniel Dennett's glorious new witty and detailed 496 pg book on the evolution of the human mind (btw, $9.18 Kindle price, and of direct relevance to AI/NLP/CL). #NLP

Ran into this beaut.

Physician Computer Assisted Coding for Professionals: Market Shares, Strategies, and Forecasts, Worldwide, 2017 to 2023

LEXINGTON, Massachusetts (March 13, 2017) – WinterGreen Research announces that it has published a new study Professional Physician Practice and Ambulatory Clinical Facility Computer Assisted Coding: Market Shares, Strategy, and Forecasts, Worldwide, 2017 to 2023. Next generation Computer Assisted Coding of medical information is able to leverage natural language software technology to support some automation of the billing process and use of analytics to achieve higher quality patient outcomes. The study has 299 pages and 110 tables and figures.

Computer assisted coding of medical information uses natural language solutions to link the physician notes in an electronic patient record to the codes used for billing Medicare, Medicaid, and private insurance companies. 

Natural language processing is used determine the links to codes. 88% of the coding can occur automatically without human review. Computer assisted coding is used in all parts of the healthcare delivery system. The coding systems work well to implement automated coding process.

Physicians think about patient conditions in terms of words. Software is configured to achieve working with physicians who are more comfortable describing a patient treatment in words than codes. The electronic patient record, created using physician dictation, is used to form the base for the coding.  Natural language solutions implement computer coding to identify key words and patterns of language. The physician dictation can be done using regular language that the software recognizes and translates into billing codes.  

Properly designed natural language processing (NLP) solutions do not require physicians to change the way they work. They can dictate in a free-flowing fashion, consistent with the way they think, and are not limited to structured inputs that may or may not fully capture the unique circumstances of each patient encounter.

Matching codes generated from physician notes to standard treatment protocols promises to improve health care delivery. Accompanying that type of physician patient management against best practice promises to revolutionize health care delivery. The ability to further check as to whether the recommendations for follow up made by radiologists and matching the commendations with the actual follow up heralds’ significant promise of vastly improved health care delivery. 

Computer assisted coding applications depend on the development of production quality natural language processing (NLP)-based computer assisted coding applications. This requires a process-driven approach to software development and quality assurance. 

A well-defined software engineering process consists of requirements analysis, preliminary design, detailed design, implementation, unit testing, system testing and deployment. NLP complex technology defines the key features of a computer assisted coding (CAC) application.

Automation of process will revolutionize health care delivery. In addition to automating the insurance, billing, and transaction systems, streamlined care delivery is an added benefit. The ability to look at workflow and compare actual care to best practice is fundamental to automated business process. 

The ability to link diagnostic patient information to treatment regimes and drug prescriptions is central to improving medical care delivery. Once a physician can see what conditions need to be followed, and see that appropriate care has been prescribed 100% of the time, care delivery improves dramatically. Diagnosis of conditions using radiology frequently results in detection of events that need follow-up.

According to Susan Eustis, lead author of the team that prepared the study, “Growing acceptance of computer assisted coding for physician offices represents a shift to cloud computing and billing by the procedure coded. Because SaaS based CAC provides an improvement over current coding techniques the value is being recognized. Administrators are realizing the benefits to quality of care. Patients feel better after robotic surgery and the surgeries are more likely to be successful.” 

The worldwide market for Computer Assisted Coding is $898 million in 2016, anticipated to reach $2.5 billion by 2023. The complete report provides a comprehensive analysis of Computer Assisted Coding in different categories, illustrating the diversity of software market segments. A complete procedure analysis is done, looking at numbers of procedures and doing penetration analysis. 

Major health plans report a smooth transition to ICD-10. This is due to rigorous testing for six years. ICD-10 has had a positive impact on reimbursement. ICD-10 coding system requires use of 72,000 procedure codes and 68,000 CM codes, as opposed to the 4,000 and 14,000 in the ICD-9 system. Managing high volume of codes requires automation. Healthcare providers and payers use complex coding systems, which drives demand for technologically advanced CAC systems. 

The market for computer-assisted coding grows because it provides management of workflow process value by encouraging increasing efficiency in care delivery for large Professional Physician Practice and Ambulatory Clinical Facility. By making more granular demarcation of diagnoses and care provided for each diagnosis, greater visibility into the care delivery system is provided. Greater visibility brings more ability to adapt the system to successful treatments...
Need I elaborate? Seriously? The writing is painful, as are the topic-skipping lack of focus and blinding-glimpses-of-the-obvious/not-exactly-news "research analysis" observations.
"Patients feel better after robotic surgery and the surgeries are more likely to be successful."
And, this is a result of back-office NLP/CAC precisely how?

Okeee-dokeee, then. A mere 14 bucks a page for a PDF file?

First of all, EHR narrative fields "text mining" functionality has been a commonplace for years now across a number of mainstream platforms (as is the converse; turning codes and data into faux-narrative text dx "impressions'). Re-labeling such now with the trendy "NLP" moniker doesn't change that (none of which is to imply that the infotech is not improving). Second, I'm not gonna pay $4,200 to maybe verify whether "[exactly?] 88% of the coding can occur automatically without human review" (in payor audit-defensible fashion). Finally, we all "think" about things in terms of "words," but dx narrative impressions are a result of the SOAP process, not the cause. They're the "A" and the "P." The "S" and "O" comprise a mix of numeric data, informatics codes, and open-ended deliberative textual information.

Beyond those, the rest of the foregoing is a poorly-written rambling melange of platitudes and unhelpfully vague filler assertions.

CEO and "Senior Analyst" Susan Eustis:

Lordy, mercy. A proxy spokesmodel? How about the CEO herself?

Stay tuned, just getting underway. Behind the curve this week, tending to my ailing daughter.


From TechCrunch:
Mary Meeker’s latest trends report highlights Silicon Valley’s role in the future of health care

Mary Meeker’s latest Internet Trends Report, out today, was full of insights on how tech is shaping our future — including now in health care. This was the first year Meeker included healthcare in her report and it shows just how much of a role tech is going to play in improving our lives going forward...
Free download, 355 pages.

Mary Meeker, 2016:



Finished Daniel Dennett's book.

Consider medical education. Watson is just one of many computer-based systems that are beginning to outperform the best diagnosticians and specialists on their own turf. Would you be willing to indulge your favorite doctor in her desire to be an old-fashioned “intuitive” reader of symptoms instead of relying on a computer-based system that had been proven to be a hundred times more reliable at finding rare, low-visibility diagnoses than any specialist? Your health insurance advisor will oblige you to submit to the tests, and conscientious doctors will see that they must squelch their yearnings to be diagnostic heroes and submit to the greater authority of the machines whose buttons they push. What does this imply about how to train doctors? Will we be encouraged to jettison huge chunks of traditional medical education— anatomy, physiology, biochemistry— along with the ability to do long division and read a map? Use it or lose it is the rule of thumb cited at this point, and it has many positive instances. Can your children read road maps as easily as you do or have they become dependent on GPS to guide them? How concerned should we be that we are dumbing ourselves down by our growing reliance on intelligent machines?

So far, there is a fairly sharp boundary between machines that enhance our “peripheral” intellectual powers (of perception, algorithmic calculation, and memory) and machines that at least purport to replace our “central” intellectual powers of comprehension (including imagination), planning, and decision-making ...We can expect that boundary to shrink, routinizing more and more cognitive tasks, which will be fine so long as we know where the boundary currently is. The real danger, I think, is not that machines more intelligent than we are will usurp our role as captains of our destinies, but that we will over-estimate the comprehension of our latest thinking tools, prematurely ceding authority to them far beyond their competence.

Dennett, Daniel C. (2017-02-07). From Bacteria to Bach and Back: The Evolution of Minds (Kindle Locations 6649-6666). W. W. Norton & Company. Kindle Edition.
"How concerned should we be that we are dumbing ourselves down by our growing reliance on intelligent machines?"

Well, I recall a couple of books I've heretofore cited on that issue.

Back to Daniel Dennett, some concluding 'Bacteria to Bach" broad stroke thoughts:.
We have now looked at a few of the innovations that have led us to relinquish the mastery of creation that has long been a hallmark of understanding in our species. More are waiting in the wings. We have been motivated for several millennia by the idea expressed in Feynman’s dictum, “What I cannot create, I do not understand.” But recently our ingenuity has created a slippery slope: we find ourselves indirectly making things that we only partially understand, and they in turn may create things we don’t understand at all. Since some of these things have wonderful powers, we may begin to doubt the value— or at least the preeminent value— of understanding. “Comprehension is so passé, so vieux jeux, so old-fashioned! Who needs understanding when we can all be the beneficiaries of artifacts that save us that arduous effort?”

Is there a good reply to this? We need something more than tradition if we want to defend the idea that comprehension is either intrinsically good— a good in itself, independently of all the benefits it indirectly provides— or practically necessary if we are to continue living the kinds of lives that matter to us. Philosophers, like me, can be expected to recoil in dismay from such a future. As Socrates famously said, “the unexamined life is not worth living,” and ever since Socrates we have taken it as self-evident that achieving an ever greater understanding of everything is our highest professional goal, if not our highest goal absolutely. But as another philosopher, the late Kurt Baier, once added, “the over-examined life is nothing to write home about either.” Most people are content to be the beneficiaries of technology and medicine, scientific fact-finding and artistic creation without much of a clue about how all this “magic” has been created. Would it be so terrible to embrace the over-civilized life and trust our artifacts to be good stewards of our well-being?

I myself have been unable to concoct a persuasive argument for the alluring conclusion that comprehension is “intrinsically” valuable— though I find comprehension to be one of life’s greatest thrills— but I think a good case can be made for preserving and enhancing human comprehension and for protecting it from the artifactual varieties of comprehension now under development in deep learning, for deeply practical reasons. Artifacts can break, and if few people understand them well enough either to repair them or substitute other ways of accomplishing their tasks, we could find ourselves and all we hold dear in dire straits. Many have noted that for some of our high-tech artifacts, the supply of repair persons is dwindling or nonexistent. A new combination color printer and scanner costs less than repairing your broken one. Discard it and start fresh. Operating systems for personal computers follow a similar version of the same policy: when your software breaks or gets corrupted, don’t bother trying to diagnose and fix the error, unmutating the mutation that has crept in somehow; reboot, and fresh new versions of your favorite programs will be pulled up from safe storage in memory to replace the copies that have become defective. But how far can this process go?

Consider a typical case of uncomprehending reliance on technology. A smoothly running automobile is one of life’s delights; it enables you to get where you need to get, on time, with great reliability, and for the most part, you get there in style, with music playing, air conditioning keeping you comfortable, and GPS guiding your path. We tend to take cars for granted in the developed world, treating them as one of life’s constants, a resource that is always available. We plan our life’s projects with the assumption that of course a car will be part of our environment. But when your car breaks down, your life is seriously disrupted. Unless you are a serious car buff with technical training you must acknowledge your dependence on a web of tow-truck operators, mechanics, car dealers, and more. At some point, you decide to trade in your increasingly unreliable car and start afresh with a brand new model. Life goes on, with hardly a ripple.

But what about the huge system that makes this all possible: the highways, the oil refineries, the automakers, the insurance companies, the banks, the stock market, the government? Our civilization has been running smoothly— with some serious disruptions— for thousands of years, growing in complexity and power, Could it break down? Yes, it could, and to whom could we then turn to help us get back on the road? You can’t buy a new civilization if yours collapses, so we had better keep the civilization we have running in good repair. Who, though, are the reliable mechanics? The politicians, the judges, the bankers, the industrialists, the journalists, the professors— the leaders of our society, in short— are much more like the average motorist than you might like to think: doing their local bit to steer their part of the whole contraption, while blissfully ignorant of the complexities on which the whole system depends. According to the economist and evolutionary thinker Paul Seabright (2010), the optimistic tunnel vision with which they operate is not a deplorable and correctable flaw in the system but an enabling condition. This distribution of partial comprehension is not optional. The edifices of social construction that shape our lives in so many regards depend on our myopic confidence that their structure is sound and needs no attention from us.

At one point Seabright compares our civilization with a termite castle. Both are artifacts, marvels of ingenious design piled on ingenious design, towering over the supporting terrain, the work of vastly many individuals acting in concert. Both are thus by-products of the evolutionary processes that created and shaped those individuals, and in both cases, the design innovations that account for the remarkable resilience and efficiency observable were not the brain-children of individuals, but happy outcomes of the largely unwitting, myopic endeavors of those individuals, over many generations. But there are profound differences as well. Human cooperation is a delicate and remarkable phenomenon, quite unlike the almost mindless cooperation of termites, and indeed quite unprecedented in the natural world, a unique feature with a unique ancestry in evolution. It depends, as we have seen, on our ability to engage each other within the “space of reasons,” as Wilfrid Sellars put it. Cooperation depends, Seabright argues, on trust, a sort of almost invisible social glue that makes possible both great and terrible projects, and this trust is not, in fact, a “natural instinct” hard-wired by evolution into our brains. It is much too recent for that. 104 Trust is a by-product of social conditions that are at once its enabling condition and its most important product. We have bootstrapped ourselves into the heady altitudes of modern civilization, and our natural emotions and other instinctual responses do not always serve our new circumstances.

Civilization is a work in progress, and we abandon our attempt to understand it at our peril. Think of the termite castle. We human observers can appreciate its excellence and its complexity in ways that are quite beyond the nervous systems of its inhabitants. We can also aspire to achieving a similarly Olympian perspective on our own artifactual world, a feat only human beings could imagine. If we don’t succeed, we risk dismantling our precious creations in spite of our best intentions. Evolution in two realms, genetic and cultural, has created in us the capacity to know ourselves. But in spite of several millennia of ever-expanding intelligent design, we still are just staying afloat in a flood of puzzles and problems, many of them created by our own efforts of comprehension, and there are dangers that could cut short our quest before we— or our descendants— can satisfy our ravenous curiosity. [Dennett, op cit, Kindle Locations 6729-6787]
He closes,
[H]uman minds, however intelligent and comprehending, are not the most powerful imaginable cognitive systems, and our intelligent designers have now made dramatic progress in creating machine learning systems that use bottom-up processes to demonstrate once again the truth of Orgel’s Second Rule: Evolution is cleverer than you are. Once we appreciate the universality of the Darwinian perspective, we realize that our current state, both individually and as societies, is both imperfect and impermanent. We may well someday return the planet to our bacterial cousins and their modest, bottom-up styles of design improvement. Or we may continue to thrive, in an environment we have created with the help of artifacts that do most of the heavy cognitive lifting their own way, in an age of post-intelligent design. There is not just coevolution between memes and genes; there is codependence between our minds’ top-down reasoning abilities and the bottom-up uncomprehending talents of our animal brains. And if our future follows the trajectory of our past— something that is partly in our control— our artificial intelligences will continue to be dependent on us even as we become more warily dependent upon them. [ibid, Kindle Locations 6832-6840]
'eh? A lot to think about, in the context of "AI/IA" broadly (and "NLP/NLU" specifically).

Back to my original "NLU" question: will we be able to write code that can accurately parse, analyze, and "understand" arguments composed in ordinary language?

A lot more study awaits me. Suffice it to say I'm a bit skeptical at this point.

Maybe we could put the hackers at Hooli on it! ;) (NSFW)

The evolution of computational linguistics and where it's headed next
May 31, 2017 by Andrew Myers

Earlier this year, Christopher Manning, a Stanford professor of computer science and of linguistics, was named the Thomas M. Siebel Professor in Machine Learning, thanks to a gift from the Thomas and Stacey Siebel Foundation.

Manning specializes in natural language processing – designing computer algorithms that can understand meaning and sentiment in written and spoken language and respond intelligently. His work is closely tied to the sort of voice-activated systems found in smartphones and in online applications that translate text between human languages. He relies on an offshoot of artificial intelligence known as deep learning to design algorithms that can teach themselves to understand meaning and adapt to new or evolving uses of language...
Worth re-citing/linking these prior posts at this point:
The Great A.I. Awakening? Health care implications?
Are structured data the enemy of health care quality?
"I'm a bit skeptical at this point."

15 Computational Semantics

1 Introduction
In this chapter we will generally use ‘semantics’ to refer to a formal analysis of meaning, and ‘computational’ to refer to approaches that in principle support effective implementation, following Blackburn and Bos (2005). There are many difficulties in interpreting natural language. These difficulties can be classified into specific phenomena – such as scope ambiguity, anaphora, ellipsis and presuppositions. Historically, different phenomena have been explored within different frameworks, based upon different philosophical and methodological foundations. The nature of these frameworks, and how they are formulated, has an impact on whether a given analysis is computationally feasible. Thus the topic of computational semantics can be seen to be concerned with the analysis of semantic phenomena within computationally feasible frameworks...

2.1 A standard approach
In general it is difficult to reason directly in terms of sentences of natural language. There have been attempts to produce proof-theoretic accounts of sentential reasoning (for example, Zamansky et al., 2006; Francez & Dyckhoff 2007), but it is more usual to adopt a formal language, either a logic or some form of set theory, and then translate natural language expressions into that formal language. In the context of computational semantics, that means a precise description of an algorithmic translation rather than some intuitive reformulation of natural language. Such translations usually appeal to a local principle of compositionality. This can be characterized by saying that the meaning of an expression is a function of the meaning of its parts...

2.2 Basic types
When considering the representations of nouns, verbs, and sentences as properties, relations, and propositions respectively, we may have to pay attention to the nature of the permitted arguments. For example, we may have: properties of individuals; relationships between individuals; relationships between individuals and propositions (such as statements of belief and knowledge); and, in the case of certain modifiers, relations that take properties as arguments to give a new property of individuals. Depending upon the choice of permitted arguments, and how they are characterized, there can be an impact on the formal power of the underlying theory. This is of particular concern for a computational theory of meaning: if the theory is more powerful than first-order logic, then some valid conclusions will not be derivable by computational means; such a logic is said to be incomplete, which corresponds with the notion of decidability (Section 1, and Section 1.2 of Chapter 2, COMPUTATIONAL COMPLEXITY IN NATURAL LANGUAGE)...

2.3 Model theory and proof theory
There are two ways in which traditional formal semantic accounts of indicatives have been characterized. First, we may be interested in evaluating the truth of indicatives (or at least their semantic representation) by evaluating their truth conditions with respect to the world (or, more precisely, some formal representation or model of a world). This can be described as model-theoretic semantics. Model-theoretic accounts are typically formulated in set theory. Set theory is a very powerful formalism that does not lend itself to computational implementation. In practice, the full power of set theory may not be exploited. Indeed, if the problem domain itself is finite in character, then an effective implementation should be possible regardless of the general computational properties of the formal framework (see Klein 2006 for example).

On the second characterization of formal semantic accounts, the goal is to formalize some notion of inference or entailment for natural language. If one expression in natural language entails another, then we would like that relation to be captured by any formalization that purports to capture the meaning of natural language. This can be described as proof-theoretic semantics. Such rules may lend themselves to fairly direct implementation (see for example van Eijck and Unger (2004); Ramsay (1995); Bos and Oka (2002), the last of which supplements theorem proving with model building).

Although a proof-theoretic approach may seem more appropriate for computational semantics, the practical feasibility of general theorem proving is open to question. Depending on the nature of the theory, the formalization may be unde-cidable. Even with a decidable or semi-decidable theory, there may be problems of computational complexity, especially given the levels of ambiguity that may be present (Monz and de Rijke 2001)...

(2013-04-24). The Handbook of Computational Linguistics and Natural Language Processing (Blackwell Handbooks in Linguistics) (pp. 394 - 402). Wiley. Kindle Edition.
Most of what I'm finding thus far is a lot of jargon-laden academic abstraction, none of it going to answering my core NLP question: can we develop code capable of accurately analyzing the logic in prose arguments -- the aggregate "semantic" "meanings" comprising a proffer? This book, notwithstanding its 801 pg. heft, frequently begs off with "beyond the scope" apologies.
Perhaps the difficulties are simply too numerous and imposing to surmount (as of today, anyway) in light of the innumerable ways to state any given prose argument -- ranging from the utterly inelegant (e.g., Eustis) to the eloquently evocative (e.g., Dennett), from the methodically Socratic/trial-lawyer-like to the rambling, unfocused, and redundant, from the fastidiously lexically and grammatically precise to the sloppily mistake-ridden, from the explicit and accessible to the obtusely inferential...
But, then,
"Computers will understand sarcasm before Americans do."  - Geoffrey Hinton
There's certainly a thriving international academic community energetically whacking away at this stuff.

From one of the "invited papers" (pdf) in this edition:


Good article.
How Language Led To The Artificial Intelligence Revolution
Dan Rowinski

In 2013 I had a long interview with Peter Lee, corporate vice president of Microsoft Research, about advances in machine learning and neural networks and how language would be the focal point of artificial intelligence in the coming years.

At the time the notion of artificial intelligence and machine learning seemed like a “blue sky” researcher’s fantasy. Artificial intelligence was something coming down the road … but not soon.
I wish I had taken the talk more seriously.

Language is, and will continue to be, the most important tool for the advancement of artificial intelligence. In 2017, natural language understanding engines are what drive the advancement of bots and voice-activated personal assistants like Microsoft’s Cortana, Google Assistant, Amazon’s Alexa and Apple’s Siri. Language was the starting point and the locus of all new machine learning capabilities that have come out in recent years...

More to come...

No comments:

Post a Comment