Developers
August 14, 2020

Using ML to Subtype Patients Receiving Digital Mental Health Interventions

Microsoft Research team is using machine learning to detect depression and anxiety early on.

Today we will talk about how Microsoft is using machine learning for mental health. Microsoft sees a clear path of helping people with diagnosed mental illness through a process called subtyping.

What is subtyping? It already happens at the moment we receive a suggestion for purchase for example. It works by using past behavior to do recommendations. It is based on the premise that states that people will respond favorably to one thing similar to what they previously responded positively.

It is not yet known how the different forms of engagement affect mental health. Studies have shown that engagement with treatments leads to better outcomes. For example, check-ins with clinician’s vs. self-directed work are not the same and have been proven over time.

Project Talia

The Microsoft Research department has been working on a project together with the Trinity College of Dublin and SilverCloud Health. The project is called Project Talia. SilverCloud Health is the world's largest provider in digital mental health.

SilverCloud Health is currently providing internet-delivered cognitive behavioral therapy (IBCT) to treat depression, anxiety, and other mental conditions.

Based on data from 54,605 SilverCloud Health patients, Microsoft has built a machine-learning framework that works by identifying patients´ subtypes that are based on the ICBT.

Probabilistic graphical models provide a rich framework to represent complex data structures. When trying to understand disease progression based on symptoms and factors, there are several examples of how probabilistic models can be used.

They can be used to understand disease heterogeneity in a longitudinal, observational clinical context, including sepsis, asthma, and kidney failure.

What kind of probabilistic graphical model was used?

The probabilistic graphical modeling they used is the hidden Markov model. It assumes that the behavior of today is conditioned by the behavior of yesterday. Similar to the deterministic philosophy which states that one event is influenced by past events.

By using the model, we want to know what the probability is of moving from one state to another. Traduced to a medical situation, what is the probability a patient will engage with a specific tool?

The probabilistic patterns of behavior are governed by something that is not directly observer but rather latent, like changing behavior from one day to the other.

The model captures what can't be measured directly with standalone metrics. A common metric might be, how engaged a person is. The model specifically shows how observed management on the SilverCloud platform transitions over time and whether the engagements have different sections representing the subtype.

A hidden state is defined. This hidden state represents true engagement. It is based on the initialization probability, using the uniform Dirichlet prior probability. With this, they can know the probability of patients transitioning across states over time.

They part from the base of the unknown. They don't know the number of subtypes, but they can learn the optimal number by using penalized log-likelihood based on the Bayesian information criteria.

SilverCloud Health has done many other works prior to the one with Microsoft. When they give permission to access data, they make sure they anonymize the data given. Age, gender, and private demographics are always kept private. They only provide relevant and needed data such as sections clicked, amount of time spent on sections, support conversations, among others.

Using the probabilistic model, five subtypes of engagement had been identified. Class 1 aka as “low engagers”, class 2 aka as “late engagers”, class 3 aka “high engagers with moderate decrease”, class 4 aka “high engagers with moderate decrease” and class 5 “highest engagers”.

The goal here is to deliver more effective treatments, nothing else than that. They have investigated if the prototypical patient behaviors were associated with depression or anxiety. Patients had to complete questionnaires to inform the symptoms.

This approach for understanding and predicting future interactions is not only being used in ICBT and is showing promise for early detection and intervention strategies. It helps gain time and work more efficiently from doubt to diagnosis. By identifying subtypes of patients early, doctors and medical workers can recommend resources that can improve symptoms in a direct way.

In conclusion, The Microsoft Research department has been working on a project together with the Trinity College of Dublin and SilverCloud Health. The project is called Project Talia. Project Talia uses subtypes of information to help diagnose depression and anxiety at an early stage. They do this by using the hidden Markov model. This model allows the medical worker to determine whether a present condition may come from past behavior. Similar to the deterministic philosophy which states that one event is influenced by past events. By working with probabilities, Microsoft, and the entire Health community make sure to have effective results.

TagsMachine LearningMental HealthMicrosoftMicrosoft Research
Lucas Bonder
Technical Writer
Lucas is an Entrepreneur, Web Developer, and Article Writer about Technology.

Related Articles

Back
DevelopersAugust 14, 2020
Using ML to Subtype Patients Receiving Digital Mental Health Interventions
Microsoft Research team is using machine learning to detect depression and anxiety early on.

Today we will talk about how Microsoft is using machine learning for mental health. Microsoft sees a clear path of helping people with diagnosed mental illness through a process called subtyping.

What is subtyping? It already happens at the moment we receive a suggestion for purchase for example. It works by using past behavior to do recommendations. It is based on the premise that states that people will respond favorably to one thing similar to what they previously responded positively.

It is not yet known how the different forms of engagement affect mental health. Studies have shown that engagement with treatments leads to better outcomes. For example, check-ins with clinician’s vs. self-directed work are not the same and have been proven over time.

Project Talia

The Microsoft Research department has been working on a project together with the Trinity College of Dublin and SilverCloud Health. The project is called Project Talia. SilverCloud Health is the world's largest provider in digital mental health.

SilverCloud Health is currently providing internet-delivered cognitive behavioral therapy (IBCT) to treat depression, anxiety, and other mental conditions.

Based on data from 54,605 SilverCloud Health patients, Microsoft has built a machine-learning framework that works by identifying patients´ subtypes that are based on the ICBT.

Probabilistic graphical models provide a rich framework to represent complex data structures. When trying to understand disease progression based on symptoms and factors, there are several examples of how probabilistic models can be used.

They can be used to understand disease heterogeneity in a longitudinal, observational clinical context, including sepsis, asthma, and kidney failure.

What kind of probabilistic graphical model was used?

The probabilistic graphical modeling they used is the hidden Markov model. It assumes that the behavior of today is conditioned by the behavior of yesterday. Similar to the deterministic philosophy which states that one event is influenced by past events.

By using the model, we want to know what the probability is of moving from one state to another. Traduced to a medical situation, what is the probability a patient will engage with a specific tool?

The probabilistic patterns of behavior are governed by something that is not directly observer but rather latent, like changing behavior from one day to the other.

The model captures what can't be measured directly with standalone metrics. A common metric might be, how engaged a person is. The model specifically shows how observed management on the SilverCloud platform transitions over time and whether the engagements have different sections representing the subtype.

A hidden state is defined. This hidden state represents true engagement. It is based on the initialization probability, using the uniform Dirichlet prior probability. With this, they can know the probability of patients transitioning across states over time.

They part from the base of the unknown. They don't know the number of subtypes, but they can learn the optimal number by using penalized log-likelihood based on the Bayesian information criteria.

SilverCloud Health has done many other works prior to the one with Microsoft. When they give permission to access data, they make sure they anonymize the data given. Age, gender, and private demographics are always kept private. They only provide relevant and needed data such as sections clicked, amount of time spent on sections, support conversations, among others.

Using the probabilistic model, five subtypes of engagement had been identified. Class 1 aka as “low engagers”, class 2 aka as “late engagers”, class 3 aka “high engagers with moderate decrease”, class 4 aka “high engagers with moderate decrease” and class 5 “highest engagers”.

The goal here is to deliver more effective treatments, nothing else than that. They have investigated if the prototypical patient behaviors were associated with depression or anxiety. Patients had to complete questionnaires to inform the symptoms.

This approach for understanding and predicting future interactions is not only being used in ICBT and is showing promise for early detection and intervention strategies. It helps gain time and work more efficiently from doubt to diagnosis. By identifying subtypes of patients early, doctors and medical workers can recommend resources that can improve symptoms in a direct way.

In conclusion, The Microsoft Research department has been working on a project together with the Trinity College of Dublin and SilverCloud Health. The project is called Project Talia. Project Talia uses subtypes of information to help diagnose depression and anxiety at an early stage. They do this by using the hidden Markov model. This model allows the medical worker to determine whether a present condition may come from past behavior. Similar to the deterministic philosophy which states that one event is influenced by past events. By working with probabilities, Microsoft, and the entire Health community make sure to have effective results.

Machine Learning
Mental Health
Microsoft
Microsoft Research
About the author
Lucas Bonder -Technical Writer
Lucas is an Entrepreneur, Web Developer, and Article Writer about Technology.

Related Articles