Personalized Depression Treatment
For a lot of people suffering from depression, traditional therapy and medication are ineffective. A customized treatment may be the solution.
Cue is an intervention platform that transforms sensors that are passively gathered from smartphones into customized micro-interventions that improve mental health. We parsed the best-fit personalized ML models for each subject using Shapley values to identify their feature predictors and reveal distinct features that are able to change mood over time.
Predictors of Mood
Depression is a major cause of mental illness around the world.1 Yet, only half of those affected receive treatment. To improve the outcomes, healthcare professionals must be able to recognize and treat patients with the highest probability of responding to particular treatments.
A customized depression treatment is one method of doing this. By using mobile phone sensors as well as an artificial intelligence voice assistant, and other digital tools, researchers at the University of Illinois Chicago (UIC) are developing new methods to predict which patients will benefit from which treatments. Two grants were awarded that total over $10 million, they will make use of these tools to identify the biological and behavioral factors that determine the response to antidepressant medication and psychotherapy.
The majority of research on predictors for depression treatment effectiveness has been focused on clinical and sociodemographic characteristics. These include demographics such as age, gender and education, and clinical characteristics like symptom severity and comorbidities, as well as biological markers.
Very few studies have used longitudinal data to determine mood among individuals. Many studies do not consider the fact that moods can differ significantly between individuals. Therefore, it is important to develop methods which permit the analysis and measurement of individual differences in mood predictors and treatment effects, for instance.
The team's new approach uses daily, in-person evaluations of mood and lifestyle variables using a smartphone app called AWARE, a cognitive evaluation with the BiAffect app and electroencephalography -- an imaging technique that monitors brain activity. This enables the team to create algorithms that can systematically identify different patterns of behavior and emotions that are different between people.
In addition to these modalities, the team also developed a machine-learning algorithm to model the changing variables that influence each person's mood. The algorithm combines the individual differences to produce a unique "digital genotype" for each participant.
This digital phenotype was correlated with CAT DI scores that are a psychometrically validated symptoms severity scale. However the correlation was tinny (Pearson's r = 0.08, adjusted BH-adjusted P-value of 3.55 x 10-03) and varied widely among individuals.
Iampsychiatry of symptoms
Depression is a leading cause of disability in the world, but it is often misdiagnosed and untreated2. In addition, a lack of effective treatments and stigmatization associated with depression disorders hinder many individuals from seeking help.
To aid in the development of a personalized treatment plan in order to provide a more personalized treatment, identifying predictors of symptoms is important. However, the current methods for predicting symptoms are based on the clinical interview, which is not reliable and only detects a limited number of features associated with depression.2
Using machine learning to integrate continuous digital behavioral phenotypes of a person captured through smartphone sensors and a validated online mental health tracker (the Computerized Adaptive Testing Depression Inventory, CAT-DI) along with other indicators of severity of symptoms could increase the accuracy of diagnostics and treatment efficacy for depression. These digital phenotypes provide a wide range of unique behaviors and activities, which are difficult to document through interviews and permit continuous and high-resolution measurements.
The study included University of California Los Angeles (UCLA) students with moderate to severe depressive symptoms. participating in the Screening and Treatment for Anxiety and Depression (STAND) program29 developed under the UCLA Depression Grand Challenge. Participants were directed to online support or in-person clinical care according to the severity of their depression. Those with a score on the CAT-DI scale of 35 or 65 were assigned to online support with a peer coach, while those with a score of 75 patients were referred to psychotherapy in person.
At the beginning, participants answered an array of questions regarding their personal demographics and psychosocial features. The questions asked included age, sex and education as well as marital status, financial status and whether they were divorced or not, current suicidal ideas, intent or attempts, as well as how often they drank. Participants also scored their level of depression symptom severity on a 0-100 scale using the CAT-DI. CAT-DI assessments were conducted each other week for participants that received online support, and every week for those who received in-person support.
Predictors of Treatment Reaction
A customized treatment for depression is currently a research priority and many studies aim to identify predictors that help clinicians determine the most effective medications for each patient. Pharmacogenetics, for instance, identifies genetic variations that determine how the human body metabolizes drugs. This lets doctors choose the medications that are most likely to work for each patient, reducing the time and effort needed for trial-and error treatments and avoid any negative side consequences.
Another option is to create prediction models combining the clinical data with neural imaging data. These models can be used to determine which variables are most predictive of a particular outcome, like whether a medication can help with symptoms or mood. These models can be used to predict the patient's response to treatment, allowing doctors maximize the effectiveness.
A new generation of studies uses machine learning methods, such as supervised learning and classification algorithms (like regularized logistic regression or tree-based methods) to combine the effects of multiple variables and improve the accuracy of predictive. These models have shown to be effective in the prediction of treatment outcomes like the response to antidepressants. These models are getting more popular in psychiatry, and it is expected that they will become the standard for future clinical practice.
Research into the underlying causes of depression continues, as well as ML-based predictive models. Recent research suggests that the disorder is associated with dysfunctions in specific neural circuits. This theory suggests that a individualized treatment for depression will be based upon targeted therapies that restore normal functioning to these circuits.
One method of doing this is by using internet-based programs which can offer an personalized and customized experience for patients. One study found that a web-based program was more effective than standard treatment in improving symptoms and providing an improved quality of life for patients with MDD. A randomized controlled study of a customized treatment for depression revealed that a significant number of participants experienced sustained improvement and had fewer adverse consequences.
Predictors of Side Effects
A major obstacle in individualized depression treatment is predicting which antidepressant medications will have minimal or no side effects. Many patients are prescribed various medications before settling on a treatment that is effective and tolerated. Pharmacogenetics offers a new and exciting method of selecting antidepressant medicines that are more effective and precise.
Many predictors can be used to determine the best antidepressant to prescribe, such as gene variants, patient phenotypes (e.g., sex or ethnicity) and the presence of comorbidities. However finding the most reliable and valid predictors for a particular treatment is likely to require controlled, randomized trials with significantly larger numbers of participants than those normally enrolled in clinical trials. This is due to the fact that the identification of interaction effects or moderators may be much more difficult in trials that only focus on a single instance of treatment per patient, rather than multiple episodes of treatment over time.
Additionally, the prediction of a patient's reaction to a particular medication will likely also require information on the symptom profile and comorbidities, in addition to the patient's prior subjective experience with tolerability and efficacy. Currently, only some easily identifiable sociodemographic and clinical variables are believed to be correlated with the severity of MDD like gender, age race/ethnicity, BMI and the presence of alexithymia and the severity of depressive symptoms.
The application of pharmacogenetics to depression treatment is still in its beginning stages and there are many obstacles to overcome. First it is necessary to have a clear understanding of the underlying genetic mechanisms is required, as is a clear definition of what is a reliable indicator of treatment response. Ethics, such as privacy, and the responsible use genetic information should also be considered. Pharmacogenetics could be able to, over the long term help reduce stigma around treatments for mental illness and improve treatment outcomes. However, as with any approach to psychiatry careful consideration and application is required. For now, the best course of action is to provide patients with a variety of effective depression medications and encourage them to speak openly with their doctors about their concerns and experiences.