Download Subtitles for Health Care Data Analytics Lecture B

welcome to healthcare data analytics machine learning and natural language processing this is lecture b the component healthcare data analytics covers the topic of healthcare data analytics which applies the use of data statistical and quantitative analysis and explanatory and predictive models to drive decisions and actions in health care the learning objectives for this unit machine learning and natural language processing are to describe the major tasks for which machine learning is used compare and contrast the major approaches for machine learning describe the major tasks for which natural language processing is used and discuss the major approaches and challenges for processing clinical narratives in this lecture we begin our discussion of natural language processing or nlp of clinical text first we'll look at basic definitions and approaches to nlp this will be followed by challenges in processing the clinical narrative in the next lecture we'll discuss various clinical nlp approaches and projects and finally we'll describe alternatives and future directions let's begin with basic definitions and approaches successful nlp of the clinical narrative could help better enable the use of data in electronic health records or ehrs we know for example that current coded data such as icd-10 does not cover the complexity of what's described in the clinical narrative we also know that a good deal of clinical information is locked in that text meaning we cannot easily extract and process the information to use for various purposes some have noted that the term nlp could actually be better described as natural language understanding because the goal of nlp is the understanding of natural language in computerized text for those desiring more detail on the various approaches to nlp and its uses the references given in the last few slides of this presentation can be consulted what are some of the use cases for clinical nlp the three major ones are listed on this slide the first use case is classification where we're trying to classify what we find in the text into some sort of category for example we may want to classify a patient finding into a category such as when determining if they might be eligible for a clinical study probably the major use case for clinical nlp is extraction where we want to extract information from a clinical narrative for example we might want to extract the findings that occur in a radiology report and even the measurements that are reported within that text a third use case is summarization where we may want to summarize or abstract the information that's in the narrative we may do this for medical literature to summarize scientific information or the clinical narrative where we're trying to summarize the major findings that the patient has we can delve further into use cases for nlp by considering cancer care this set of use cases comes from some promotional literature from a company that sells clinical nlp products but actually gives a good set of use cases for which nlp might help us for example we might identify potential clinicals trials matches something akin to what we mentioned on the last slide we might be able to do advanced information extraction from complex patient documents we may be able to carry out more precise information retrieval for clinical case histories and outcome studies we also may be able to streamline the process of entering patients into cancer registries in addition we may be able to use the data that we extract using nlp to apply predictive models and care coordination rules to clinical narratives in the patient record we may also be able to perform semantic enrichment of patient documentation to improve the ability to search their nodes we can analyze patient narratives for insights into treatment outcomes and also to assess the effect of genetic aberrations on disease finally we may be able to support tumor boards where the care of patients who developed cancer is discussed by those providing care for them let's take a more detailed look at human language so we can understand the applications and limitations of nlp tools and clinical documents linguists talk about the levels of human language we begin with phonology the sound units that make up a language's discrete sounds are called phonemes next level up is morphology which is the analysis of parts of words which are called morphemes sometimes a whole word is the morpheme but other times they may be bound morphemes that are part of the word for example many anatomic locations such as the appendix are bound to another word such as itis indicating inflammation thus appendicitis indicates there's inflammation of the appendix there are other morphemes such as fairing and itis there are also bound morphemes that indicate procedures such as an appendectomy syntax refers to the rules that govern the construction of language sometimes called the grammar semantics describes the meanings of the words phrases and the sentences that make up language linguists also talk about pragmatics which is how the context of language affects its meaning and then there's the larger world knowledge that's not explicitly part of language but is the general knowledge that's necessary to understand it the classic approach to nlp goes through three phases the first phase is syntax where we attempt to recognize the grammatical constituents of language sentences phrases within them and down to nouns verbs adjectives etc the next phase is semantics where we attempt to recognize the meaning of those words phrases and sentences finally is the context in which the sentence occurs each of these levels is successively harder and requires more knowledge engineering but would add more value if we could solve those problems one of the ways we address the inability to completely perform classic nlp is through the use of rules and matching where we don't aim for complete understanding of everything in the document but instead try to recognize the terms that occur and perhaps normalize them this may allow us to understand what was said or instead of using detailed grammar rules we may use machine learning techniques where we learn the rules of parsing rather than developing human enumerations of all the possible grammar rules that might exist let's explore syntax and semantics in more detail processing of syntax is usually done via a technique called parsing this requires a grammar which is the rules that govern the syntax of language the most common way that we express a grammar is as a set of rewrite rules where a more complex grammatical construct is rewritten from constituent parts for example a relatively simple grammatical rule is that a sentence consists of a noun phrase a verb phrase and a noun phrase an example is the patient has severe hypertension the first noun phrase is the patient the verb phrase is the verb has the second noun phrase is severe hypertension of course noun phrases themselves can be rewritten into more basic constituents there are determiners such as a grammatical article an example of which is the word the there are also adjectives such as severe and a noun phrase can also just consist of a single noun the symbols that cannot be further decomposed such as an adjective and a noun are called terminal symbols likewise those that can be further decomposed such as sentence and noun phrase are called non-terminal symbols as you can imagine the grammar supporting the english language can get highly complex with many many rewrite rules this is why the machine learning approach has superseded the approach of trying to enumerate every last grammar rule in semantics we aim to map the parts of speech these nouns adjectives verbs etc into standardized terminology for medicine probably the most descriptive terminology is snomed ct processing language has been one of the most challenging computer tasks and is difficult not only in the clinical narrative but almost in all forms of natural language clinical narratives such as progress notes and discharge summaries can be even more difficult to process than other types of text for many reasons one is that clinical narratives are written in a telegraphic elliptical style oftentimes the narratives are not complete sentences we'll see examples of that in a moment clinical text also may have spelling errors or grammatical errors we also know that physicians and others may take license with language and oftentimes there may be important information that's buried within normal language that's implicit but not actually in the words and phrases we'll look at some of the challenges at the syntactic semantic and contextual levels here's a look at some of the syntactic challenges that were first enumerated by sager in the 1980s others have since validated these challenges as mentioned in the previous slide a great deal of clinical narrative text is syntactically incomplete that is at least according to sager's analysis half of all sentences in the clinical narrative were found to be grammatically incomplete if we think of the minimal english sentence as subject-verb object we see different types of incomplete sentences for example the medical record may delete the verb and object when the text says stiff neck and fever there has been a deletion of the verb and object from the sentence in brain scan negative there's deletion of the verb is for the statement positive for heart disease there is deletion of the subject and verb such as the patient has and finally was seen by local doctor has deletion of the subject as humans we can read these and still for the most part understand what's happening but computer algorithms especially those that are solely based on rules have difficulty with these sorts of violations of rules of basic english grammar there are also semantic challenges which again as humans especially those who have some clinical knowledge we readily understand but to a computer that's just functioning based on rules there's a lot more difficulty we know that words have different senses and meanings for example when we read in a medical chart murmur is appreciated we know that likely there's a clinician who's listening probably with a stethoscope to the heart and there's a murmur it's not so much that the murmur is appreciated in the sense of it being liked by the same token when we read about eye drops we're thinking about drops of liquid containing medication put into the eye and not the eye physically dropping likewise when we read mass at three o'clock we know that we're likely reading about something that's felt on the left-hand side of the abdomen and not that there's a religious service in the afternoon another semantic challenge is synonymy where different words and phrases have the same meaning but they're expressed differently for example consider the phrase epigastric pain after eating versus another phrase postprandial stomach discomforts these two phrases have no words in common but essentially mean the same thing there is also polysemi where the same words and phrases have different meanings for example someone might say the pcp of the patient with pcp advised him to stop using pcp pcp is an acronym that stands for several things such as primary care physician pneumocystis carini pneumonia or an abbreviated name for the drug fincyclidine there are a number of additional semantic challenges one is negation the clinical narrative is often full of negation clinicians may say the patient does not have this finding or that finding or that this disease is not present or saying we're choosing not to use this treatment and instead are using another one negation is common in medical text for example patient does not have any chest pain there is also uncertainty in natural language text clinicians may say things like patient treated for possible pneumonia there is also temporality just because something is mentioned doesn't mean that it's present now for example patient has history of pneumonia or there might be something that's been resolved such as chest pain resolved after administration of nitroglycerin there are also contextual challenges in the clinical narrative the term that describes a broad category of these is co-reference which is the relation between linguistic expressions that refer to the same real world entity consider the sentence chest x-ray shows nodule in left upper lobe followed by another sentence the tumor has increased in size to two centimeters the phrase the tumor from the second sentence is actually referring to that same nodule from the first sentence there's a particular type of co-reference that can be challenging which is anaphora or the use of pronouns consider these two sentences he complains of chest pain it awakens him at night it in the second sentence refers to chest pain in the first sentence there's another type of contextual challenge where there's the deletion of subjects this is quite common in clinical narratives so we may see strings of sentences such as complains of chest pain increasing frequency worse in the morning again as human readers we usually understand that quite easily but when we have a natural language processing system the computer may not make the connections across the sentences are there any silver linings that may enable us to have hope that we can carry out clinical nlp it turns out that there are first is the notion of sub-grammars after work by sager in medicine and other disciplines she determined that there were sub-grammars that were grammars that were specific to disciplines and that there was a sub-grammar of clinical narratives that were actually fairly regular and predictable another finding is that medical charts tend to have a predictable discourse especially documents like the history and physical where the document begins with the history of the patient goes into the past medical history and then into the physical exam physicians for the most part follow a well-prescribed pathway through the exam more recently another silver lining has been that we should abandon the notion of processing the entire clinical narrative and instead focus on specific elements that we can identify to indicate whether or not a specific disease or specific clinical finding is present thus giving up on the approach of processing everything and instead focusing on specific elements present before we look at usage of clinical nlp and systems for it in the next lecture let's talk briefly about how we evaluate how well nlp systems work there are a variety of ways that systems can be measured but basically we want to determine how well they identify correct concepts and how well they don't identify incorrect concepts the measures that we typically use are recall and precision recall is the proportion of correct concepts found for example if there are 100 concepts that should be found by an nlp system and 75 actually are found then the recall is 75 percent precision is the proportion of found concepts that are correct so if we identify 150 concepts and 75 of them are correct then our precision is 50 percent many evaluations of nlp are carried out in so-called challenge evaluations where there is a common data set that different researchers use these different research groups will compare the results on the same task for the clinical nlp community the largest and most participatory challenge evaluation has been the i2b2 nlp shared task there has also been a systematic review of all studies through 2010 that was published and will be described more in the next lecture this concludes lecture b of machine learning and natural language processing in summarizing this lecture we learned the major use cases for nlp are classification extraction and summarization the major phases of nlp are syntax semantics and context each of which has challenges and is successively harder to do with computers and there may be some silver linings to help with nlp such as sub grammars predictable discourse and focus on processing less than the entire meaning of everything in the document you