The collaboration between a machine-learning specialist and a psychology expert might appear unexpected at first glance. However, MIT's Rosalind Picard and Massachusetts General Hospital's Paola Pedrelli share a common vision: leveraging artificial intelligence to enhance accessibility in mental healthcare for patients worldwide.
With over 15 years of experience as both clinician and researcher in psychology, Pedrelli emphasizes that "significant obstacles prevent individuals with mental health conditions from accessing and receiving appropriate care." These challenges include identifying when and where to seek assistance, locating nearby providers accepting new patients, and securing financial resources and transportation for attending appointments.
Pedrelli serves as an assistant professor of psychology at Harvard Medical School and holds the position of associate director at the Depression Clinical and Research Program within Massachusetts General Hospital (MGH). For more than five years, she has partnered with Picard, who is a professor of media arts and sciences at MIT and a principal investigator at the Abdul Latif Jameel Clinic for Machine Learning in Health (Jameel Clinic). Together, they work to develop sophisticated machine-learning algorithms designed to assist in diagnosing and monitoring symptom progression in patients suffering from major depressive disorder.
Machine learning represents a subset of artificial intelligence that, when provided with extensive data and examples of desired outputs, becomes capable of autonomously performing tasks with increasing accuracy. This technology excels at identifying meaningful patterns that might remain hidden to human observers without computational assistance. Through wearable devices and smartphones belonging to study participants, Picard and Pedrelli collect comprehensive data on skin conductance, body temperature, heart rate, activity levels, social interactions, self-reported depression assessments, and sleep quality. Their objective is to create machine learning algorithms capable of processing this vast amount of information and extracting meaningful insights—identifying when an individual might be experiencing difficulties and determining potentially beneficial interventions. They anticipate that these algorithms will eventually provide both physicians and patients with valuable information regarding individual disease progression and effective treatment approaches.
"We're developing advanced models that not only identify common patterns across populations but also learn to recognize categories of changes within an individual's life," Picard explains. "Our goal is to offer those who desire it access to evidence-based, personalized information that can positively impact their health outcomes."
The Intersection of Machine Learning and Mental Wellness
Picard's journey at the MIT Media Lab began in 1991. Three years later, her book "Affective Computing" catalyzed the emergence of an entire research field bearing the same name. Today, affective computing stands as a vibrant discipline focused on creating technologies capable of measuring, sensing, and modeling data related to human emotions.
While initial research concentrated on determining whether machine learning could utilize data to identify a participant's current emotional state, Picard and Pedrelli's current work at MIT's Jameel Clinic has evolved significantly. They now investigate whether machine learning can predict disorder trajectories, detect behavioral changes in individuals, and provide data that informs personalized medical interventions.
In 2016, Picard and Szymon Fedor, a research scientist in Picard's affective computing laboratory, initiated collaboration with Pedrelli. Following a successful small-scale pilot study, they are now in the fourth year of their five-year study funded by the National Institutes of Health.
To conduct their research, the team recruited MGH participants diagnosed with major depression who had recently modified their treatment approach. To date, 48 individuals have enrolled in the study. For 22 hours daily over a 12-week period, participants wear Empatica E4 wristbands. These wearable devices, created by one of Picard's companies, capture biometric information including electrodermal (skin) activity. Participants also install applications on their smartphones that collect data on text messages, phone calls, location information, and app usage patterns. Additionally, these applications prompt participants to complete biweekly depression surveys.
Each week, patients meet with a clinician who evaluates their depressive symptoms.
"We input all the data collected from wearables and smartphones into our machine-learning algorithm and evaluate how effectively the machine learning predicts the labels assigned by clinicians," Picard states. "Currently, our prediction accuracy is quite high."
Empowering Individuals Through Technology
While developing effective machine-learning algorithms presents one set of challenges, designing a tool that genuinely empowers and supports users represents another significant hurdle. Picard notes, "The critical question we're now addressing is: once we have these machine-learning algorithms, how can they truly benefit people?"
Picard and her team are carefully considering how their machine-learning algorithms might present findings to users—potentially through a dedicated device, smartphone application, or even a system that notifies predetermined doctors or family members about optimal ways to support the user.
For instance, consider a technology that detects that an individual has recently experienced reduced sleep, spent more time indoors, and exhibited an elevated heart rate. These subtle changes might escape notice by both the individual and their loved ones. Machine-learning algorithms could analyze these data points, mapping them against the individual's historical patterns and the experiences of other users. The technology might then encourage the individual to engage in behaviors that previously improved their well-being or prompt them to contact their healthcare provider.
If implemented inappropriately, this type of technology could potentially produce adverse effects. For example, if an application alerts someone that they're approaching a severe depressive episode, this information might trigger discouragement or intensify negative emotions. To address this concern, Pedrelli and Picard are actively involving actual users in the design process to create a tool that provides assistance without causing harm.
"An effective tool might inform an individual: 'The reason you're feeling down may relate to changes in your sleep patterns, reduced social activity, and decreased time spent with friends, along with diminished physical activity. We recommend finding ways to increase these activities,'" Picard explains. The team also prioritizes data privacy and ensures informed consent from all participants.
Artificial intelligence and machine-learning algorithms can identify connections and patterns in large datasets that humans might easily overlook, Picard concludes. "I believe there's a compelling argument for technology helping humans better understand and support other humans."
