Machine Learning models for Mental Health use cases
Image Source: Elisa Riva
As more time passes and innovation progresses, it is easy for the term ‘Artificial Intelligence’ to become another hackneyed buzzword thrown around in the media. However, the importance of those innovations must not be downplayed by our overfamiliarity with the words. Psychiatry is one field experiencing and engaging in these innovations with promising results. At BrainSightAI, we have front row seats to the tableau and I am excited to share an introduction to my work.
Be it in virtual agency, intelligent wearables or human behaviour simulation and modelling, Artificial Intelligence is making inroads in the mental health care system and its effects are noticeable. With high computational speed, smart device capability, artificial intelligence and other enabling technologies, the ambulatory monitoring of a person’s mental state is now a field of digital medicine that is rapidly progressing. These monitoring systems have moved past the tracking of observable physical quantities into the realm of intelligent detection of unobservable mental states such as stress, mind wandering, etc. This field of study known as automated mental state detection is interdisciplinary in nature with roots in psychology and the technical domains of sensors, wearable devices and machine learning. This hypothesis proposes the detection of mental states using bodily responses to certain internal and external stimuli. The extraction of this clinical information using computational tools is where interesting things begin to happen. The variety of sensors and the amount of information that can be extracted from those sensors are factors taken into consideration while designing the experiment or research study. The design architect must consider which predictive model archetype suits their purpose and proceed accordingly. The most popular machine learning techniques today are the supervised and unsupervised methods.
Supervised machine learning models (predictive algorithms built using labelled data) have long been used to extract a variety of standard features, such as specific facial landmarks, head position, fundamental frequency of speech, and overall level and variance in each physiological signal to detect human emotions. It has even been observed that the machines outperform humans at detecting emotions. Taking the clinical example of identifying patients suffering from anxiety disorders for differentiating them from healthy controls, a set of features might be chosen to assess the patients. Supervised learning methods build a model between the outcome (in this case: anxiety or not) and a series of features, such as age, gender, education background, work type and so on, which are collected from different data sources.
Unsupervised machine learning models (predictive algorithms with no prior hypothesis about the output and based solely on the input information) are popular because they provide a level of honesty to the outputs inferred that results extrapolated from a laboratory produced, labelled data set might be biased against. To offset this, some studies have compared the results of their work to existing psychiatric measures and diagnostic scales. Taking the example of a study that used an unsupervised k-means clustering algorithm to determine levels of depression and assess the validity of the clusters against existing depression scales, it was able to adjust for real world context such as perceived stress, sleep quality and anxiety. This was done by taking into account the inputs obtained from the individuals involved in the study.
Our app, Snowdrop, uses an unsupervised learning model to track the user’s sleeping patterns. The usage of automatically generated data reduces personal biases that may come into play while answering a questionnaire on their sleep habits. The passive nature of the data collected also limits a feeling of intrusion a self-questionnaire may cause.
The advances in AI will transform behavioural and mental health care in the years to come. Given this transition, it is important for healthcare professionals and AI systems designers and developers to be aware of the current and emerging strengths, intricacies, and opportunities that AI brings to the behavioural, emotional and cognitive disorder domains.
4. Yang, Z., Chen, C., Li, H., Yao, L., & Zhao, X. (2020). Unsupervised Classifications of Depression Levels Based on Machine Learning Algorithms Perform Well as Compared to Traditional Norm-Based Classifications. Frontiers in Psychiatry, 11.