| Student's name | SCIPER |
|---|---|
| Jeremy Dezalos | 283885 |
| Hojjat Karami | 332629 |
| Léonard Vaney | 284023 |
10% of the final grade
MIMIC-IV is a publicly available dataset containing de-identified electronic health record data from over 2 million ICU patients at Beth Israel Deaconess Medical Center from 2008 to 2019. It includes vital signs, lab results, medications, and clinical notes, and can be used for clinical research and machine learning.
Sepsis is a life-threatening condition caused by the body's response to an infection. It can cause widespread inflammation, organ failure, and septic shock. Early recognition and treatment are crucial for a positive outcome.
We filtered the MIMIC-IV dataset, which has data from over 60,000 ICU patients, to select only those who have at least one ICD code that contains the word "sepsis." This resulted in approximately 8,000 patients being included in our analysis.
Electronic health records (EHRs) pose several challenges related to data management and analysis. One significant challenge is managing multi-modality data, which may include diverse types of data such as imaging, audio, and text data. Time-series and event sequence data is also common in EHRs, but presents challenges in terms of analysis and modeling due to the complexity of temporal relationships. EHRs often contain heterogeneous data, meaning that different data types are collected and stored in different formats, and may be of varying quality. Additionally, EHRs contain data on different disease subtypes, and understanding the unique characteristics and patterns of these subtypes is a complex task. Addressing these challenges requires sophisticated data management techniques, as well as advanced data and visual analytics that can accommodate these diverse data types and structures and provide insights into specific disease subtypes.
We are interested in acquainting students with patient data in electronic health records (EHRs). To achieve this, we will develop several visual analytics tools that can facilitate exploring the dataset by adopting recent advancements in visual analytics technology. These visual tools will provide students with insights into the data and help them understand the underlying patterns and trends.
Based on our literature review, we have identified several components that we plan to develop in our project. These components include:
- A global overview of the entire dataset.
- Interactive subset selection, allowing the user to select a subset of the data by interacting with the overview panel.
- The ability for the user to specify desired outcomes, such as length of stay, specific diseases (using ICD codes), or in-hospital mortality.
- A simple clustering algorithm (e.g. K-means) to show potential patient subgroups. The user should be able to select two subgroups for further analysis.
- A panel utilizing recent "event sequence simplification" algorithms to show differences in medication order for each subgroup.
- A panel to show aggregated trends and temporal patterns in clinical variables, such as laboratory and vital values, for each subgroup.
The dataset contains 7889 admissions that contains an ICD code that is related to sepsis. This exploratory data analysis will focus on the patients themselves and will try to explain who they are.
Top 15 ICD codes are illustrated in the pie chart below. we can see that severe sepsis with septic shock is the most frequent icd code in our dataset followed by other sepsis-relevant icd codes.
The age repartition of this cohort is this:
We can see that no one has a age between 0 and 20. This is because the dataset has been anonymized to attempt at the privacy of the subjects. Thus, every person with an age less than 20 is represented as being 0 years old. So without even if we do not know what a spesis is we can already deduce that it affects patients of all ages.
Let's look at the outcome of the patients at the end of their stay at the hospital. We defined an outcome as positive if the patient survived their stay.
Because sepsis is a life-threatening condition, around two thirds of the patients do not survive their stay (In the graph, label 0 means being dying and label 1 means surviving).
The time of the stay is also useful to understand the patient's financial status. One could argue that wealthy patients stay longer at the hosptial because they can afford to pay the expanses longer compared to poorer patients.
Let's look at the distribution of time stayed given the patient's insurance
We notice that the patients using medicare have a shorter stay in the hospital than patients who have other insurances. Medicaid is an health insurance for people with limited income or resources. This tends to verify our hypothesis that poorer people leave the hospital faster to reduce the cost of their bills.
Ethnicity is also an attribute available in the database. Let's look at how they are represented.
White people are the most represented in the cohort. This representation is coherent because the hospital is situated in Boston.
Ethnicity can potentially correlate with other values such as time of stay or outcome.
It is difficult to determine a correlation between the time and the ethnicity of the patient. So let's look at the mortality ratio given the ethnicity.
We can see that people with unspecified ethnicity have higher chances to survive. The unable to obtains column only 74 people so we do not consider this to be significant. However the people who have an unknown ethnicity are almost eight hundred and we still do not have a good explanation for this.
This database has been used extensively for analysis or for tool testing. Here is a link to a paper that provides a detailed analysis of MIMIC-III.
We used as a base point for our visualizations an interesting paper on the state-of-the-art EHR visualization (Link here). Section 4 contains some links to other papers with useful visualizations such as events sequence simplification, clustering or cohort comparison.
Some examples below:
EventFlow Transform temporal events into an aggregated display to identify hidden trends in the data.
OutFlow Visualization of temporal event sequence data.
CoCo Tool for comparing groups (cohorts) of temporal event sequence data.
DICON Visualization tool to explore similarity in cohorts of patients.
The originality of our project lies in the fact that we want to make the exploration of this dataset understandable and usable by students or inexperienced users with EHR data and to give them useful insights on this dataset.
10% of the final grade
We will be using a pre-processed subset of the MIMIC dataset, which includes clinical data for 12,000 ICU patients. The dataset provides 36 irregularly-sampled sensor observations and 6 static demographic features for each patient, along with an in-hospital mortality label. However, the subset is highly imbalanced, with only 7% negative samples, and has a missing ratio of 88.4%.
To analyze this dataset, we will use a pre-trained deep neural network model and assess its predictive output. The model is attention-based, and its weights can be interpreted as feature importance.
The skeleton of our visualization can be found here
The website will be divided in panels. Each panel will cover the entirety of the screen and the user will be able to change panels by scrolling. The scrolling will always go to the next panel and not stop between panels. When using a trackpad, the system is a bit sensitive for now.
Our final visualization core will consist of two panels:
- Panel I: a t-SNE plot that visualizes the learned representation of all patients.
- Panel II: a single-patient view that visualizes all measurements taken throughout their entire ICU stay.
Next, We explain each panel in detail regarding its expected features and behavior.
The t-SNE plot is a scatter plot that visualizes high-dimensional patient representations in 2D coordinates. The following features are expected:
- It displays positive and negative labels, as well as model predictions with varying degrees of uncertainty.
- The scatter plot is zoomable and should respond to the
mouseoverevent, which will displayPanel IIfor the selected patient.
In the figure below, a single patient's data is illustrated using a heatmap. We aim to enhance its informativeness for the user by introducing additional features.
Here is a list of our expected features:
- Brush and zoom: This feature allows the user to zoom in on the time axis by brushing.
- Variable re-ordering: When the user hovers over a specific timestamp, the variables (on the y-axis) are realigned based on their deviation from the normal range. For instance, at the top, we should see very high values, and at the bottom, we should see very low values. It is crucial to add some degree of tolerance to make the feature less sensitive to small changes when hovering over the entire timeline.
- Attention weights: When the user requests to see the attention weights, appropriate channels are used to display the data.
- In this panel, our aim is to create an interactive visualization for electronic health records. The user can select variables of interest from a dropdown menu based on their knowledge and also select a subgroup based on static features.
The next step will display selected variables for the patients. The selected variables fall into two categories:
-
Event sequence data where each event is represented as
$(t_i,e_i)$ for timestamp and the event time, respectively. -
Time-series variables where each data point is represented as
$(t_i,m_i,v_i)$ for time, modality, and the corresponding value, respectively. Here, we convert the time series to events based on normal ranges. For example, HR>120 is regarded asHR-High.
We show a raw visualization for the selected variables, where for each data point
We will use a triangle to represent each event, with the x and y axis representing channels. Examples are the sequence of medications, hospital wards, and ICD codes.
Finally, when the user clicks on the Simplify button, we want to show a simplified version of the modalities illustrated in the previous step using the following techniques:
-
Time alignment: We will explore various methods for time alignment such as Dynamic Time Warping (DTW), Edit Distance, LSTMs, and more.
-
Modality alignment: Based on the alignment of time stamps and the similarity of sequences, we will reorder the modalities on the y-axis.
-
Sankey diagram: We will use a visualization tool similar to a Sankey diagram to simplify similar sequences.
We also considered the following additional ideas:
-
Brush and Zoom: The user can zoom over both the x and y-axis to see more details for the selected area.
-
t-SNE plot: We will provide a simple t-SNE diagram for the selected variables (in the first step) based on sequence so that the user can easily select a subgroup by brushing and selecting over it.
As recommended in the course, the project will be written in javascript using the D3 library to manage data and show visualizations. We host our webpage on the free service offered by github.
To complete the project, we will mostly need to use those lectures from the course:
- Lecture 4 & 5: We will be using D3, display interactive data and do some data filtering to choose what we will be presenting according to user's inputs.
- Lecture 6: Our dataset contains lots of various data and we will need to display them with carefully choosen colors and patterns. The goal would be to make them as understandable as possible for a non-expert user.
- Lecture 11: The mimic dataset is only comprised of tabular data. We will display them using stacked bar, Sankey diagram, time alignment.
80% of the final grade
Most of our work throughout the semester can be found on this personal repository: https://github.com/JeremyDezalos/data-viz-project
Link to the website.
Link to the screencast.
Link to the process book.
- < 24h: 80% of the grade for the milestone
- < 48h: 70% of the grade for the milestone