Journal Club 2016 Presentation V3
Transcript of Journal Club 2016 Presentation V3
Mining Electronic Health Records Data: Network Analysis
of Adverse Health Effects among Victims of Intimate Partner Violence
By: Kathleen Whiting, Larry Y. Liu, Mehmet Koyuturk, Gunnur Karakurt
Introduction• Intimate Partner Domestic Violence (IPV) is a rising global issue• Extensive health, societal, and economic costs• In 2010, 30% of women experienced domestic abuse from an intimate
partner, 25% of which involved violent altercations with serious physical injuries
• There are notable adverse health effects associated with IPV• Ranges from minor injuries to serious disability and death• Common effects include psychological distress, STI’s, gynecological
problems, unplanned pregnancies and complications, gastrointestinal issues, mental health and substance abuse disorders, neurological symptoms, chronic physical ailments
Introduction• Current progress in rooting out key IPV-related health
effects has been lacking• Previous work has included identifying long-term
complications from IPV but no progress on mitigation or prevention• Female victims of violent crimes tend to have overall poorer
health and socio-economic status with worse adverse physical and mental conditions to be treated than female victims of non-violent crimes• They often don’t get treatment until six months after the incident,
and their treatment costs are often 15-24% higher than victims of non-violent crimes in the first 12 months
• Victims of violent crimes tend to present with symptoms that are misattributed to non-violence related causes
Introduction• Electronic Health Record (EHR) mining on adverse
health effects might help to improve screening of IPV• More accurate to place a picture of reoccurring health
effects suffered from known IPV patients in their health records then by self-reports• Easier for healthcare providers to understand and gauge• Could lead to finding associations among common health
effects that are seen in various IPV victims• Has a good sample size to work with when determining
these common health effects
Previous Work• EHR records were sourced via Explorys to identify
symptoms that were more prevalent with noted IPV victims than the general populace• The queries focused on differentiating health effects such
as:• Acute Injuries to violence-related physical injuries• Chronic or persistent conditions• Conditions that were from non-violent or violent causes
• These queries were then sorted into 28 broad categories• These categories affirmed four distinct health problems that
were complications from IPV
Previous WorkDisease Classification Category
Percent (%)
Acute Condition 34.09%Acute Injury 23.14%Chronic 16.18%Disorders 15.60%Cardiovascular 8.60%Pregnancy Related 7.53%Gynocological 7.45%Musculoskeletal 6.18%Mental Health 5.93%Gastrointestinal 5.60%Allergy 5.02%Substance Abuse 4.78%Other 4.57%Nervous system 4.53%Skin related 3.99%Respiratory 3.91%Eyes, Ears, Nose & Throat 3.75%Excretory 3.05%Personal History 2.22%Congenital/Hereditary 1.56%Endocrine 1.56%Neoplasm 1.32%Immune System 1.32%Nutrition 1.19%Procedure 0.82%Neuropathy 0.78%Family History 0.74%Diabetes 0.62%
.
Table 1. Percent of diseases considered significant in IPV falling into each category (28)
Objective /Rationale• Given the queries of adverse health effects sorted into 28
diverse categories, we wanted to see if there were any significant interactions among these health effects from IPV victims
• Using network analysis to generate a map helps to visualize these interactions• Each node on the map represents each of the 28 symptom categories• The edges connecting the nodes represents a relationship between
two categories, which signifies a shared frequency of symptoms coded between the two
• The edge thickness identifies the strength of these relationships
Methodology• A base data set from the previous study was provided
from Explorys (de-identified) • A data set containing 3458 symptom terms that are possibly
associated with IPV (domestic abuse)• The sample size (# of records) was 5870 containing
exclusively female patients between the ages of 18 to 65• Each symptom term was coded to at least one of the 28
broad categories attributed from the previous study• Among the 3458 symptom terms, 2429 terms were deemed
statistically significant @ 95% confidence (Chi-Squared Independence Test) w/ the background data set from previous study
Methodology• The key measurement to generate a network map was
to count the frequency of the occurrence of each category• Each symptom term will be accounted to at least one
category• Each symptom term with two attributed categories will be
counted as a pair• Each symptom term with three or more attributed categories
will be counted for each combinatory pairs of categories that could be associated• This results in two output files needed for mapping: nodes
and edge pairs
Methodology• Using GePhi, we inputted the two files containing the
nodes and edge pairs to generate the network map• The thickness of the edge pairs are determined by the
frequency of occurrence of that edge pair• The thickness of the nodes were determined by the number
of terms attributed to each category • The edge pairs are bidirectional as there was equal
precedence between the nodes in the pairing • The map was rescaled and transformed into a shape that
was more presentable and easier to analyze without compromising the integrity of the map
Results1a
Acute Injury
1b
Acute Condition
2 Chronic3 Substance Abuse4 Mental Health5 Other6 Disorders7 Gynecological8 Pregnancy Related9 Allergy10
Procedure
11
Congenital/Hereditary
12
Nutrition
13
Neoplasm
14
Personal History
15
Family History
16
Neuropathy
17
Diabetes
18
Gastrointestinal
19
Cardiovascular
20
Nervous System
21
Respiratory
22
Musculoskeletal
23
Eyes, Ears, Nose & Throat
24
Excretory
25
Endocrine
26
Immune System
27
Skin Related (not burns)
Figure 1. Network Map of the 28 categories made up of 2429 symptom terms found to be significantly more prevalent among victims of IPV than the general population. Larger nodes indicate higher coding frequency of the category by itself, while thicker edges reveal the strength of relationship between two nodes by signifying the frequency of any two categories being coded together on a single symptom term. Darker nodes appear more frequently among the pairs than lighter nodes.
Results1a
Acute Injury
1b
Acute Condition
2 Chronic3 Substance Abuse4 Mental Health5 Other6 Disorders7 Gynecological8 Pregnancy Related9 Allergy10
Procedure
11
Congenital/Hereditary
12
Nutrition
13
Neoplasm
14
Personal History
15
Family History
16
Neuropathy
17
Diabetes
18
Gastrointestinal
19
Cardiovascular
20
Nervous System
21
Respiratory
22
Musculoskeletal
23
Eyes, Ears, Nose & Throat
24
Excretory
25
Endocrine
26
Immune System
27
Skin Related (not burns)
• Out of 2429 symptom terms, there were at least 208 pairs of interactions present to at least one of the symptom terms
• Most significant nodes were 1b and 2, having the great frequency of occurrence
• 1b has the most significant degrees of twelve• Strong pairings with 2, and 19• Moderate pairings with 7, 18, 27,
8• Fair pairings with 3, 20, 21, 22,
23, 24• 2 has five significant degrees
• Strong pairing with 1b• Fair pairings with 4, 19, 20, 22
• It is worth noting that pairing between 1b and 2 could be due to ambiguity of a symptom term that could be chronic or acute
Results1a
Acute Injury
1b
Acute Condition
2 Chronic3 Substance Abuse4 Mental Health5 Other6 Disorders7 Gynecological8 Pregnancy Related9 Allergy10
Procedure
11
Congenital/Hereditary
12
Nutrition
13
Neoplasm
14
Personal History
15
Family History
16
Neuropathy
17
Diabetes
18
Gastrointestinal
19
Cardiovascular
20
Nervous System
21
Respiratory
22
Musculoskeletal
23
Eyes, Ears, Nose & Throat
24
Excretory
25
Endocrine
26
Immune System
27
Skin Related (not burns)
• 6 also shows significant pairings, but its target connected nodes are more significant in 1b and 2
• Node 1a does have significant symptom terms associated with it but it doesn’t have many terms that are associated with another node
Discussion• The network map showed consistent results to our
knowledge of IPV thus far• IPV victims tend to have complications of chronic and acute
conditions in addition to acute injuries• Chronic and acute conditions also have a strong link towards
conditions related to physiological systems• Chronic and acute conditions are prevalent since these
conditions often requires medical attention• Also these conditions are often related closely to each other
due to ambiguous symptom terms
Discussion• Gynecological and/or Pregnancy-related conditions also are
significant as IPV-related abuse often do severely target these aspects of the victim’s body• Mental health terms are expected to be significant but the discrete
nature of how the diagnosis terms were coded, the network did not reflect those terms• Generally mental health issues are not associated with other physiological
systems• We also hypothesized that stress may play a role in determining
category significance, because aspects of stress is involved to many of the categories that seem associated with IPV in the network map• IPV could also contribute to stress and anxiety as a side effect or
complication
Limitations• The network map does not accurately analyze subtle patterns and
cycles of these associations among the adverse health effects to IPV• The dimension of time was not added in but if added, would open up a new
layer of analysis and a more accurate network map
• The existing Explorys data set did not appear to reflect the generally accepted rate of IPV, that approximately 1 out of every 4 women would encounter an IPV event in their lifetimes• This could be due to the inadequate reporting and accuracy of self-reports of
IPV in potential victims• Healthcare professionals and EHR systems don’t consider IPV as a ‘condition’
or ‘diagnosis’ but as a ‘finding’, masking more subtle cases of IPV from being shown
Conclusion• The network map generated here by EHR mined data
of IPV victims helps to illustrate some of the relationships among adverse health effects associated with IPV• This would help to improve screening and diagnostic
measures to IPV patients as well as understanding any signs• Hope to push EHR and healthcare providers to make IPV
notation a higher priority for better accuracy in reporting and monitoring• Opens up new study angles to affirm new symptom
relationships that could aggress IPV side-effects and complications (like stress)
Personal Next Steps• Using Explorys to query new data sets to see a
relationship of adverse health effects between victims of IPV to Traumatic Brain Injuries (TBI)• Finding out significant diagnosis terms that are heavily seen
in both TBI and IPV patients• Developing a comorbidity relationship between these terms
to the general population (background)• Generating a network analysis and mappings of such
connections to hopefully find more less-likely considered diagnosis terms that can be associated with IPV-induced TBI patients
AcknowledgementsKarakurt Lab• Gunnur Karakurt• Kate Whiting
Advisor• Mehmet Koyuturk
Personal Thanks• Mario Yepez-Olvera
Funding:• Clinical and Translational Science
Collaborative of Cleveland• National Center for Advancing
Translational Sciences – National Institute of Health
Assistance:• Explorys Inc.