Progress: A Machine Learning Approach to Anomaly Detection in Mental Health Data

Abstract

The mental healthcare industry is in need of technical solutions to improve patient outcomes. More specifically, we see a major need for mental health patients to have the ability to track and monitor their mental health progress in a quantitative manner. To tackle this issue, we designed a mobile application that is used for journaling. Each journal entry provides a qualitative insight into a patient's mental health status. Using sentiment analysis, we convert users' qualitative journal entries into quantitative metrics of emotion and tone. These quantitative metrics allow users to track their mental health progress over time. Most importantly, we also dynamically perform anomaly detection as users create journal entries. These anomalies could correspond to high levels of rage, or other mental health episodes. This allows the user to further track and understand their own mental health. To perform sentiment analysis, we use IBM Watson which, in our testing, on average, correctly identified the primary emotion for all categories of journal entries. These categories included angry, sad, happy, and fearful. To perform anomaly detection, we use a modern technique known as isolation forests, which performs well on small amounts of testing data and is relatively computationally efficient. Given a single anomalous emotion in large number of primary emotions, our techniques can correctly label the anomalous journal entry 71\% of the time. We view this as major success and see this as a strong proof of concept for data and machine learning driven mental healthcare. Beyond the core implementation of each component, we also architected a system that can effectively scale to many thousands of users.