Syndrome Anomaly Detection Chapter 3 - Are These Symptoms Normal?
In our previous post we discussed how to utilize NLP to anonymously extract patient presenting symptoms, we had to overcome the next hurdle “are these symptoms normal?” in the eventual quest to detect syndrome anomalies. To answer this question, we identified three main steps to this portion of the system. First, we needed to analyze each patient based upon the type of illness the patient most likely had. Second, based upon that illness we had to evaluate whether or not that symptom presentation was atypical or not. And lastly, we needed to produce an alert system that could handle noisy data.
To first determine what type of disease a patient most likely had our goal was to train an “out-of-the-box" machine learning classifier such as a support vector machine or tree-based classifier. However, before we could train the classifier, we need to be able to represent the symptoms as a uniform feature input. The challenge here was that there are many ways to say the same symptom (runny nose, sniffles, or medically known as rhinitis). To have a reduced and uniform input feature space to train our disease classifier we developed a “symptom converter” using an ontology approach to handle the different variations of verbally described symptoms. This approached worked well and we were able to accurately train our classifiers to correctly identify the type of disease a patient likely had.
Now that our system was able to provide the type of illness the patient most likely had, the goal became to evaluate how typical or more importantly how atypical the presenting symptoms were on a per-patient basis. Due to the fact that our goal was to detect a novel disease that had yet been contracted, we need to solve this challenge through unsupervised AI. We turned to an autoencoder. Utilizing a repository of over 300,000 patient symptoms and associated diagnoses from the CDC we train autoencoders to output atypical/typical symptom scores for the given disease categories. This way we are able to provide a metric of how atypical or typical a patient's symptoms are the given disease category the machine learning classifier determined the patient to be associated with.