Evidence-Grounded Mapping of Multimodal Human Sensing Psychological Transdiagnostic Dimensions
This paper introduces a clinician-in-the-loop benchmark using the GLOBEM dataset to evaluate how well large language models can generate evidence-grounded B-HiTOP profiles from multimodal sensing data, finding that while semantic abstraction effectively organizes self-report evidence, it acts as an information bottleneck for indirect passive behavioral signals.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine a world where your phone and watch could quietly track your daily life—your sleep patterns, how far you walk, who you talk to, and how you feel in the moment. For years, researchers have hoped to turn this stream of data into a clear picture of a person's mental health, moving beyond simple checklists to understand the complex, overlapping ways people experience distress. Instead of asking if someone has a specific disease like depression or anxiety, a newer approach called the Hierarchical Taxonomy of Psychopathology suggests looking at broad dimensions of human experience, such as emotional instability or social withdrawal, which often appear together across different conditions. The challenge has been figuring out how to translate the raw, silent signals from a device into these meaningful psychological concepts without a doctor sitting right there to interpret them.
In this study, researchers set out to test whether modern artificial intelligence could act as that interpreter. They built a system to see if large language models—advanced computer programs trained on vast amounts of text—could look at a person's daily digital footprint and generate a profile of their mental state that a human clinician would find reasonable. The team used a massive dataset called GLOBEM, which contains years of data from thousands of people, including passive sensor readings from their devices, short daily surveys about their mood, and longer questionnaires about their overall well-being. Because the original dataset did not include the specific mental health scores the researchers wanted to predict, they could not simply check if the computer was "right" in a traditional sense. Instead, they asked a different, more practical question: given the evidence available, is the computer's guess about a person's mental state logical and defensible?
To answer this, the researchers created a rigorous testing ground involving nearly 15,000 daily snapshots of participants' lives. They asked the artificial intelligence to predict scores for 29 different mental health indicators, ranging from feelings of sadness to problems with impulse control. They tested two different ways for the computer to do this. In the first method, the computer looked directly at all the raw data—sensor numbers, survey answers, and questionnaire results—and tried to assign a score to each mental health indicator in one go. In the second method, the computer first summarized the data into a broad psychological description, like a short paragraph describing the person's general state, and then used that summary to assign the specific scores, without looking at the original numbers again.
The results revealed a surprising truth about how these machines process information. When the computer relied on self-reported data, such as the daily surveys and questionnaires where people explicitly described their feelings, the two-step method worked significantly better. By first organizing the scattered survey answers into a coherent story, the computer was able to make much more sensible predictions about specific mental health symptoms. However, the story changed completely when the computer relied on passive sensing data, such as how much someone moved or how often they used their phone. In this case, the two-step method made the computer worse. When the AI tried to summarize the raw sensor data into a general description first, it seemed to lose the subtle, specific details that were actually needed to make an accurate judgment. The computer became overly cautious, often assigning the lowest possible score to almost everything, effectively saying "nothing is wrong" when the data might have suggested otherwise.
This finding suggests that there is no single best way to use artificial intelligence for mental health monitoring. The approach that works best depends entirely on the type of information being used. If the data comes from people telling the computer how they feel, summarizing that information first helps the computer understand the bigger picture. But if the data comes from silent sensors tracking behavior, summarizing it first can strip away the very details that matter, acting as a bottleneck that blocks the computer from seeing the full reality. The study concludes that for these systems to be truly useful, they must be designed to match the specific nature of the data they are analyzing, rather than applying a one-size-fits-all strategy. This careful alignment is essential if we are to turn the quiet signals of our daily lives into a tool that genuinely helps us understand our mental well-being.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.