The incidence of suicide in American service members represents a major health crisis. Precision medicine and machine learning offer recent breakthroughs and advances that could be used to better understand and mitigate this crisis. However, given the sensitive nature of medical records, and particularly that of physicatric patient information, there is a major bottleneck in the available data that be utilized to make progress in the task of suicide detection and prevention. Here we introduce a system, SynthNotes, which aims to provide sufficiently realistic synthetic psychiatric notes that can be freely shared with the research community without compromising patient confidentiality.
Synthea is a tool developed by the Mitre Corporation for developing full synthetic populations of patients. Their software relies on census data, medical incidence rates, domain experted crafted care plans (Need to figure out how this is created). Census data at the (county or city?) level provides probability distributed for demographic information ranging from gender, socioeconomic status, education levels, alcoholism, sexual orientation etc. We use this rich source of quantitative information to increase the realistic patient models behind our note generation.
General module framework: A collection of independent(? do we have co-occurrence rates for disease?) modules that function as a state machine. Probabilities for transitions between states come from (? what? check references in synthea). Also, I don't think the patient modeling itself is the same kind of state graph. No need for transitions. However, some demographic information is injected into other modules, like homelessness, alcoholism, etc.
Paragraph on disease profiling.... I think this mostly relies on medical incidence rates in a given demographic region, county or city. Care plans are created. Need to learn where the probabilities for transitions in the state graphs come from. This is likely in the module files (src/main/resources/modules)
Rely on Synthea for initial demographic information about patients that we can use to enhance our Phase I notes. Some existing modules and demographic information is particularly useful. Modules that we can use now:
1. Medication:
* Pain relievers: Includes moderate and strong opiods and otc
2. Injuries
3. Cancer modules:
* colorectal cancer
* lung cancer
4. Alcohol and substance abuse:
* Discrete True False for alcoholism with transitions to recovery possible
* Opiod addictions
5. Self harm:
* Distributions by gender, and refined by age group
* Fatal and non-fatal attempts with simple care plans
* Non-fatal methods:
* poisoning (overdose), cutting, suffocation
* Fatal methods:
* Firearm, suffocation, poisoning
* Other - vehicular, drowning, falls, etc
Note: While fatal and non-fatal attempts have different distributions based on gender, the methods do not. A simple addition for us is to split these further to accomodate for the differences between men & women. (Needs citation)
6. Injuries:
We know that patients at high risk for suicide are also at risk for non-suicidal injuries. Synthea supported injuries are
* Gunshot
* spinal
* concussion
* whiplash
* bone fractures
* burns
* laceration (with body location information)
* sprains
* knee injuries (torn ligaments)
* shoulder injuries (torn rotator cuffs)
Injuries include pain medication (opiates and otc) as part of care plans). We can integrate medical injury history into military service narratives.
7. Homelessness:
National numbers from HUD. Synthea captures temporary and chronic states of homelessness. Also captures visiting homeless shelter, and some social history during those events. Includes HIV status and current opiod abuse. I need to figure out where synthea writes this information. Looks like information gathering event goes into "procedure.csv", while information gained is in "observations.csv". Visiting a homeless shelter is in "encounter.csv"
8. Fibromyalgia
9. Arthritis / Joint replacement
We don't have mobility information related to these types of diseases and treatments. This is another area for enhancement as I think the literature points to mobility as a factor in suicide risk?
10. Pregnancy: This is a large module but I'm not sure if it is relevant at the moment. Does include information regarding miscarriage and abortion rates by age group.
Our patient modeling: Extend with fields: * Firearm ownership (We can get this into the Synthea demographic model easily I think. Maybe not at a finegrained city level, but certainly nationally. Can also split firearm ownership by gender, and veteran or non-veteran).
The first phase of our synthetic note generation takes the most simplistic approach of using a simple template based approach. We use canned text containing placeholders for variables commonly found in sample SOAP notes and from note guidelines [reference to guidelines]. We also include keywords and life stressor indicators found to be effective as suicide risk indicators. Most fields are selected from a configuration file by random selection although some are held constant to preserve basic logical coherence.
Phase II of the synthetic note generation process will incorporate classical NLG techniques. Due to limitations in data access this is a necessary intermediary step. There are two areas for improvement which will be discussed below.
Our first step in enriching the synthetic notes model will be patient modeling. The primary goal of this is to create patients with richer informational properties that can then be reported in the synthetic notes. This is important as a key goal of our work is to enable search and patient modeling that is useful to the research community at large. As such, we will be continually enhancing the scope and copmlexity of our synthetic patient models so that they maintain an increasing fidelity to the underlying sensitive medical records of the VA.
The early stages of patient modeling will entail creating patient profiles that are logically coherent. Initial properties will cover areas such as demographic information, living conditions, occupation, socioeconomic factors, military service history, marital status, weapon ownership, and some personal narrative information. The narrative information, as will all other properties will be informed by past research into suicide risk factors. We envision this to initially provide information on areas such as past criminal history, physical and emotional trauma, history of abuse, etc.
The set of options for the above mentioned properties will continue to be either hardcoded discrete values or some range or set of values. It will be the responsibility of some sort of logic engine to search the possible space of options and create logically sounds patient profiles. Logic rules will also be hardcoded.
Disease modeling is itself a rather large undertaking. Initial progress can be made by referring to the literature on conditions most commonly associated with suicide risk and focusing on the properties and patient profiles which are present with each disease. In addition to the corpus of suicide risk research we may also lean upon existing clinical tools such as the UMLS (Unified Medical Language System)1 which provides large dictionaries of medical terms and clusters them into conceptual categories. Within UMLS is also a semantic net which holds relational information. We need to adequately explore the capabilities of UMLS. However, from an initial investigation it appears that a great deal of information regarding relationships of treatments to diseases are encoded in the semantic net. We can also use the metathesaurus for syntactic variability. We may also be able use this to add other features to increase the realistic nature of our synthetic notes. For instance, common abbreviations, misspellings, etc., can be added to notes based on lookups from UMLS.
A later step, following the initial construction of key disease profiles, would be the logical integration of patient and diseases. That is, our patients, based on their profile information, should match the probabilistic likelihood of suffering from some disease.
Physician modeling refers creating synthetic clinical notes that would be particular to the style of a physician. As mentioned above, this could contain common misspellings or usual shorthand or abbreviations that a physician may use. Given time and a determination of sufficient research value, it could be beneficial to pair semantic meaning with syntactic structure of sentences and explore whether we can find distinct physician groups. Other language modeling techniques are certainly available such as n-gram modeling.
The previously discussed profiles do not discount the value of our current template design structure but rather serve to augment it's informational load and logical coherence. This in turn helps with search methods, and is a useful tool for benchmarking and testing our text processing infrastructure. However, it is a brittle system that involves a great deal of time in crafting sentences. The next step, given our limited data access is to create a sentence generation system. There is a great deal of literature on this process and we have many options. However, there is no opensource implementation that is also domain independent.
Classic sentence generation has typically been described broadly as a three stage pipeline: document planning, microplanning, and surface realization (the actual writing of sentence structures that are present in the final document). For the beginning of our second phase of SynthNotes, we will be focusing on surface realization from a grammar perspective. That is, using formal grammars to construct a possible set of legal sentences. There are of course many different grammars to choose from, but this decision has not yet been set.
Provided there is a grammar in place, the framework for constructing sentences will involve sets of "messages" that can be constructed with appropriate arguments. Each message will be given a set of arguments and contain the grammatical structure to write a set of different possible sentences. For example, we could have a message like PreviousSubstanceAbuse which when given a patient profile will construct a grammatically correct sentence that relates to a patients history of substance abuse. The most notable opensource tool for surface realization is SimpleNLG2, a Java based tool. Our current software is in Python, but there are options for integrating the two. Or, a determination could be made that Java is more suitable because it seems the surface generation of text will consume a large portion of the library's code content.
Another tool that has been discovered in the last week is the Grammatical Framework (GF)3. GF is a functional type based language, based on Haskell, that appears to provide tools for sentence construction from a higher level of abstraction, notably the semantic level. It seems that in some way it is paired with FrameNet4, a semantically annotated dataset, and could possibly lead to a leaner and more concise codebase. However, the learning curve for this tool seems steep and is an important consideration.
Given that we are able to successfully create sufficiently realistic notes through the modeling and NLG methods outlined above, we will need to turn to document planning. This is an area that needs to be explored more by our group. However, the most common and aggreed upon language structuring theoretical foundation appears to be found in Rhetorical Structure Theory (RST)5, 6. This is a natural starting point. There are other issue related to the second stage of the NLG pipeline, microplanning, that contribute to textual fluency, but at this point are of a smaller concern. This is especially true if we are assuming search methods that are restricted to the word or phrase level.
Footnotes
-
Unified Medical Language System, https://www.nlm.nih.gov/research/umls/ ↩
-
SimpleNLG, https://github.com/simplenlg/simplenlg ↩
-
Grammatical Framework, http://www.grammaticalframework.org/ ↩
-
Rhetorical Structure Theory, https://en.wikipedia.org/wiki/Rhetorical_structure_theory ↩
-
Rhetorical Structure Theory: Toward a functional theory of text organization, https://doi.org/10.1515/text.1.1988.8.3.243 ↩