To provide effective therapy for people suffering from a depression or any mental health problem, it is important to understand the impact therapies and experiences can have on the future state of a person (for instance upon the mood). If we understand this influence, it allows us to select specific treatments that are more effective and speed up recovery. Currently, limited evidence is present within the domain of psychology that allows us to obtain a deeper understanding and therefore make well-founded decisions.
This is where the area of Artificial Intelligence (AI) can contribute: trying to create predictive models for the development of mental states, in our case mood, and thus estimate how fast a person would recover given certain conditions. Within E-COMPARED, this contributes to one of the intended outcomes of the project: can we identify specific groups of patients for which a certain therapy (blended therapy or treatment as usual) is preferred over the alternative? In this article, we aim to explain what approaches can be used to create predictive models. Before going into details about the AI approach, we will elaborate more on the data that is available within therapies that we can use to learn from.
Data in Mobile Interventions
One of the main advantages of modern mobile-based interventions for depression is that they generate a lot of data about the behavior of patients. This data can be collected in different ways: (1) by prompting the user to answer questions via the smart phone, and (2) by using the sensors of the smart phone. Psychologists refer to these approaches as Ecological Momentary Assessments (EMA): the user behavior is measured during regular daily routines. Of course we should refrain from asking too many questions and maximize the information we use from the sensors as they are measured in an unobtrusive way. Altogether, this data provides us with a good impression of the patients’ behavior: we can ask regularly how the patient is doing; measure how active and engaged patients are. The resulting data can be used to feed algorithms to distill accurate predictive models for the patient.
So how do we generate accurate predictive models using AI techniques and the data we have just described? Essentially, there are two options: top-down and bottom-up.
In the approach we call top-down, we start from the existing body of knowledge within the domain of psychology and try to create computer models from the existing theories. These computer models together form a virtual patient model that allows us to simulate the behavior of the patient given certain therapeutic influences. Of course one might ask: where does the data come into play in this option? This is certainly used, but mainly to tailor the settings of the model. It is imaginable that some people have better coping skills than others. Therefore, we use a part of the data collected for the patient(s) and try to see what settings fit the data of the patient best. Then, we use these settings to make predictions on the behavior of the patient in the future. The figure below shows an example of such a model. Here, the circles represent the relevant states of a patient while the arrows represent the dependencies between the states. It goes beyond the scope of this article to dive into the details of the model, but it suffices to say that these models are described in a very precise way such that they can be executed on a computer.
Figure 1. Top-down mood and coping model
A second approach is the bottom-up, or machine learning approach. In this case, we start from the data and try to automatically identify what measurements are good predictors (in our case for the mood). There are a lot of different approaches to generate such predictive models. All approaches have a notion of the importance of a measurement and they create a model using the most important measures. Models come in different forms ranging from mathematical equations and decision trees to sets of rules. In the end, we seek to find predictive models with measurements that were known in the literature to confirm that we are indeed on the right track. In addition, we would be interested in finding a few new ones that clearly contribute to more accurate predictions.
So far, we are still collecting data within the E-COMPARED project. However, preliminary studies with datasets taken from other studies show that making predictions is certainly not a trivial task. Figure 2 shows a visualization of the measured mood of a patient versus the mood described by a bottom-up approach. We need to tailor and improve the AI approaches to the case at hand in order to make sufficient accurate predictions. But we are confident that we will be able to create insightful models that can contribute to more insights into predicting therapeutic outcome and developments of mood.
Figure 2. Example of reported values for mood (gray) versus the machine learning prediction (red).
Mark Hoogendoorn, Altaf Abro, Ward van Breda & Michel Klein
VU University Amsterdam
Dennis Becker, Vincent Bremer & Burkhardt Funk
Leuphana Universität Lüneburg