carywoods / arthritisdm243

Classification of Autoimmune Diseases

Complete Report

CSE_243_Project_Report

Motivation

We would like to increase the accuracy and shorten the time to diagnosis of sufferers with the following Autoimmune diseases: Ankylosing Spondylitis (AS), Psoriatic Arthritis (PsA), Systemic Lupus Erythematosus (SLE), Sjögren’s Syndrome (SS), Adult-Onset Still’s Disease (AOSD), and Rheumatoid Arthritis (RA). We hope to accomplish this by identifying the common early symptoms for each disease as reported by the patient.

Autoimmune Arthritis diseases are difficult to diagnose. This difficulty is exacerbated by discrepancies between the symptoms given for each disease in medical journals and the symptoms that patients report experiencing prior to diagnosis. These discrepancies prolong the time to disease diagnosis. It is common for sufferers of the six diseases in the study to go years without an accurate diagnoses. Since these diseases are autoimmune diseases, during this time, the body is literally attaching itself. The more time that lapses between disease onset and disease diagnosis, the more damage occurs. Damage which is irreversible and which reduces the patient’s mobility, longevity, and quality of life. Any work in this area that either confirms early symptoms commonly associated with one of these diseases or identifies new ones, aides in the diagnosis process and possibly shortens time to diagnosis resulting in better long term prognosis.

Challenges with Dataset

The dataset is built against a set of questions organized as a survey. The accompanying questionnaire has 59 pages. While the meaning of the data may seem obvious to appropriate medical practitioners, the data was far from ready to be analyzed systematically. Here we list some of the major challenges which we tackled with a great effort:

1. Not all the questions are about earlier symptoms. Some of the questions are related to posterior symptoms or diseases. So, they surface after a person already has developed an autoimmune disease. And as the data-set holds data for those questions, too, it led us to filter those posterior attributes from the dataset. So, first challenge was to confirm which were in fact early symptoms.

2. Mapping from questions to data object attributes were not obvious. For this reason, we couldn’t readily tell which column refers to which question. This made our first challenge even more difficult. In the early experiments, we inadvertently included some posterior attributes for prediction models and ended up very high accuracy rates but unfortunately wrong. To make the mapping clear, we renamed all the columns in such a way that we could find the associated question fast. But that was not enough as many of the questions didn’t have question numbers. So, for some questions, we created unique prefixes which can group related questions. We also made the prefixes to hold some semantics. For example, we use the prefix Post_AD_, for all the attributes which relates to diseases happening after the first autoimmune disease. Smart prefixing also helped us to filter out attributes automatically. So, for first disease prediction, we matched the column names with Post_AD_ and AAD_AD_ prefixes and dropped them. Without renaming columns it would be a tiresome work for us to manipulate columns for transformations and clean-up.

3. Inconsistent attribute values. When it comes to attribute values for a real-life survey, all kinds of surprises are common. Here are some that we experienced and how we handled it:

Final pipeline:

This is a final pipeline of our whole process. It was developed through process iterations in our experiements. Final process pipeline. The bluebordered methods requires manual interventions and optionally some automated processes.

Data mapping & Attribute Renaming

The pivotal element of a smart mapping between the questions and dataset attribute is the Prefixing mechanism we employed. The prefixed we created served several purposes:

• To denote which section, group, or question the attribute belongs to. For example, AAD_AD prefix denotes the group of the chosen 6 autoimmune diseases which can develop after the first autoimmune disease.
• Some questions has multiple choice options. Each of the option becomes a attribute in the dataset. So, they were prefixed with the same string. This helps to filter out the whole question from dataset.
• Prefixes also help us to segment data. Though, we did not conduct any experiment with segmentation in this project, it will be helpful for future work. In addition to help us to build effective mapping, renaming also helped us to create transformers easily. For example: “[M_Thr_Nk_Ns_24] 34. Did you have any of the following issues with your mouth/ throat/ neck/ nose region during the first 24 months after initial onset? [Painful swollen and tender lymph nodes in areas of the body not including the face and/or neck?]” is an actual attribute name in the raw dataset. If we want to refer to the column in our automated cleaning and transformation scripts, that would be cumbersome to work with. But our new name for the column, Ex5_34_body, was easy to refer. Here Ex5 denotes section of "Exhibit 5", 34 denotes "Question no 34", and body denotes the option with body excluding face and neck.

Data Cleaning & Transformations

We call automated transformation algorithms by transformers. Before transforming values of an attribute, we need to make sure they are clean and consistent enough for the transformer to work. Each transformer can apply a set of ordered basic transformers. For continuous values like age, we automated detection of non-numeric values and manually fixed them in Microsoft Excel. For categorical valueswe followed the following steps:

1. Figure out definitions of missing values for the attribute
2. Replace missing values with string "NA2". "NA2" was chosen instead of "NA" as some analytic and machine learning libraries automatically processed "NA" as missing values. We wanted full control on our processes.
3. Figure out unique values. This itself is two step process. First step is automatically extract unique string values. Then manually detect "duplicates" and map duplicates to a single value.
4. Create a transformer to transform all the values for the attribute. For mixed attributes having data from both pre and post symptoms, we needed to extract new attributes before applying the transformations.

Experimental pipeline

Our final process pipeline 6 emerged from experimental pipeline in 7. We revised the cleaning and transformation processes several times which resulted in a sequence of revisions in and execution of other processes.

Results

Feature Importance

Future Work

Possible future work includes:

1. Create survey that lends itself to analysis and that also includes questions that will help us answer questions raised during analysis. Identifying patterns of symptoms over time for different diseases/ segments.
2. Segmentation Analysis
3. More experiments with models
4. Predicting a sequence of Autoimmune diseases
5. Identifying relevant symptoms