Data Analyzing Malpractice Lawsuits in Shiny-R: 2004-2012
Project GitHub | LinkedIn: Niki Moritz Hao-Wei Matthew Oren
The skills we demoed here can be learned through taking Data Science with Machine Learning bootcamp with NYC Data Science Academy.
You can play around with the app here: Visualization of Malpractice Lawsuits
Despite saving lives every day, medical professionals get sued by their patients quite often. Over the nine year period spanning 2004 to 2012, over 123,000 patients successfully sued for malpractice. There is a broad spectrum of reasons as to why this happens, whether it be a mistake by a professional or an overly upset patient. While I acknowledge that doctors are going to make fatal mistakes on occasion, I wanted to test whether or not patients and malpractice lawyers have a tendency to exploit the system.
I began with a very large csv file I found on Enigma (source). Most of my initial work came in the form of data cleansing, as more than two thirds of the observations containing missing data. This made interpreting certain numerical columns a challenging exercise; for example, if the payout on a lawsuit was 'NA', did that mean that the data was missing? Or did the patient get paid nothing? Additionally, I had an overwhelming 71 columns to look at as potential variables!
Since this data-set was so rich, I made the decision to cherry pick the variables that I deemed most relevant. So initially, I asked myself the simplest filtering questions I could; What does the distribution of malpractice payments look like? Do certain medical professionals get sued more than others? What type of patients are most likely to file a malpractice lawsuit? Based on these questions, I decided my most important variables were 'Total Payment','Profession of the defendant','Allegation against the defendant', and 'Outcome of the Patient'.
So I chose my variables, but how was I supposed to visualize and interpret them? The only numerical variable was 'Total Payment', while the rest were categorical with at least 10 levels. Rather than force my own direction on the application, I decided that allowing the user to subset and compare data from all of these categories was the best way to go. This was a relatively simple task for
Shiny in R:
library(shiny)
ma = readRDS("data/final.RDS")
shinyUI(navbarPage("Interpreting Malpractice Lawsuits",
tabPanel("By Payment",
sidebarPanel(
selectInput("outcome", "Choose what happened to the patient:",
choices=c("All","Death","Unknown","Emotional Injury","Insignificant",
"Major Permanent Injury", "Major Temporary Injury",
"Minor Permanent Injury", "Minor Temporary Injury",
"Significant Permanent Injury","Quadriplegic Brain Damage Lifelong Care"
)),
checkboxGroupInput('vars', 'Compare allegations against the professional:',
choices=levels(ma$allegation)[-(1:2)], selected = "Anesthesia Related"),
hr(),
helpText("Specify a range of total payment"),
sliderInput("payrange",label = "Payment Range (millions):",min=0,max=30,value=c(0,1.5),step=0.5),
submitButton("Update Density plot")
),
mainPanel(h4("What is the distribution of malpractice payments?"),
plotOutput("hist"),
h4("Summary Statistics in millions of dollars:"),
verbatimTextOutput("summary"))
)
library(ggplot2)
library(dplyr)
library(shiny)
ma = readRDS("data/final.RDS")
ma$total_pay = ma$total_pay/(1e6)
result = ma$patient_outcome
shinyServer(
function(input, output) {
# output$imag1 = renderImage({
# src="img/patient.gif"
# })
output$hist <- renderPlot({
injury <<- switch(input$outcome,
"All" = ma,
"Death" = ma[result == 'Death',],
"Unknown" = ma[result == 'Cannot Be Determined from Available Records',],
"Emotional Injury" = ma[result =='Emotional Injury Only',],
"Insignificant" = ma[result =='Insignificant Injury',],
"Major Permanent Injury" = ma[result =='Major Permanent Injury',],
"Major Temporary Injury" = ma[result =='Major Temporary Injury',],
"Minor Permanent Injury" = ma[result =='Minor Permanent Injury',],
"Minor Temporary Injury" = ma[result =='Minor Temporary Injury',],
"Significant Permanent Injury" = ma[result =='Significant Permanent Injury',],
"Quadriplegic Brain Damage Lifelong Care" = ma[result =='Quadriplegic Brain Damage Lifelong Care',]
)
payswitch <<- filter(injury, allegation %in% input$vars, total_pay<=input$payrange[2],total_pay>=input$payrange[1])
qplot(total_pay, data = payswitch, geom = "density",color = allegation,xlab="Total Payment (millions of dollars)",main="Density of lawsuit payments by patient outcome") +
coord_cartesian(xlim = c(input$payrange[1],input$payrange[2])) +
scale_x_continuous(breaks=seq(0,input$payrange[2], input$payrange[2]/10))
})
output$summary = renderText({
payswitch <<- filter(injury, allegation %in% input$vars, total_pay<=input$payrange[2],total_pay>=input$payrange[1])
paste0(
"Minimum: ", summary(payswitch$total_pay)[1],"n",
"Maximum: ", summary(payswitch$total_pay)[6],"n",
"Average: ", summary(payswitch$total_pay)[4],"n",
"1st Quantile: ", summary(payswitch$total_pay)[2],"n",
"Median: ", summary(payswitch$total_pay)[3],"n",
"3rd Quantile: ", summary(payswitch$total_pay)[5] )
})
While this may look overwhelming to someone without R experience, the code is actually fairly straightforward. On the user-interface side, a list of choices is given for each input. On the server side, each input choice selected by the user is fed into the switch function which appropriately filters the data. Once the user submits his/her choices, the graph is created and we can finally begin answering our questions!
Q1: What is the distribution of Malpractice Payments?
With so many ways to filter a rich data-set, I felt it was extremely important to be able to graphically compare specific categories. Thanks to Shiny, this can easily be accomplished with a few lines of code! The average successful malpractice lawsuit results in a payment of around $260,000, but this is highly influenced by some of the larger payments. The 1st quantile value is much more revealing, saying that the highest proportion of malpractice lawsuits result in a payment under $50,000.
Looking more closely at the density plot, we can see that many allegation curves share a similar density spike around $500,000, and $1 million. I believe this is reflective of an underlying structure in allegations against the defendant as they relate to total payment, and look forward to analyzing this further.
Q2: Which medical professionals are found guilty of malpractice most often?
I thought that comparing lawsuit counts by professional would be more fruitful, but unfortunately there isn't much beyond the superficial. Allopathic Physicians (M.D.) get sued by far the most, with dentists coming in a very distant second. While there were nearly 100 different types of professionals that have faced a malpractice lawsuit, the number of these less frequent types was so small in comparison to the the top 10 most-sued professions that it didn't make sense to include them.
The most interesting thing I found in this comparison plot was that hospitals seem to go out of their way to shield nurses and especially interns/residents from lawsuits. Their respective bars are shockingly small in comparison to MD's. Shouldn't these employees be making more mistakes than the experienced doctors?
In truth, they are. Interns and residents are guaranteed to experience plenty of growing pains, and will almost certainly make a mistake(s) that result in a patients' death. That said, the ego's of these interns and residents can be fragile, and it is important to protect them through their growing process. As a result, the attending physician almost always takes the heat.
Q3: Are certain types of patients more likely to sue than others?
This was the question I was most interested in answering; are certain patients exploiting the system to make off with a quick score? The summary statistics encompassing all categories paint a vague picture showing that women file 33% more lawsuits than men, but are also awarded less money. This fed straight into my skeptical nature; if women are filing more successful lawsuits, shouldn't they be getting paid more?
By Patient
Filtering the data by the patient outcomes of "Insignificant Injury", "Minor Temporary Injury", and "Emotional Injury" produced some interesting results. Female patients who suffered insignificant or minor temporary injury were 100% more likely to file a lawsuit than men, and those who suffered emotional injury were beyond 200% more likely to sue than men! Given this, it seems like women are more likely to file lawsuits of a minor nature.
However, it should be noted that insignificant or emotional injuries represent a very small fraction (6.4%) of the pool of successful lawsuits by female patients.
The exploratory nature of this project was a lot of fun, but as a Data Scientist, I feel my work is incomplete; these comparative graphs were my way of performing only a superficial analysis of the problem. I left out many potentially significant variables, and I feel that a data-set this rich demands some machine learning methods to provide more valuable insight. Check back soon for a more thorough analysis!
About Author
