Aaron Broukhim | HXI - Human-centered eXtended Intelligence

UNDERSTAND: Uplifting the New generation through DBT Education and Resilience for Social Triggers, Anxiety, Negativity, and Depression

Fri, 11 Jul 2025 00:00:00 +0000

Overview

Many college students experience depression and anxiety together, which compounds both symptom severity and functional impairment. Treatments like CBT and DBT work, but they are delivered on a schedule. The skill a student needs is taught in a session on Tuesday and required at 2am on Saturday, and the gap between those two moments is where the intervention fails.

UNDERSTAND closes that gap with three components.

Sensing. Wearables and phones carry signals that precede and accompany distress: physiological arousal, sleep and activity disruption, changes in social behavior and interaction patterns. We instrument these continuously and in daily life, so that the system is reading the situation as it develops rather than reconstructing it afterwards from self-report.

AI. Machine learning models turn those raw streams into an estimate of when someone is struggling and when they are receptive, which is the harder of the two questions. Large language models then generate the support itself, and our work constrains that generation so the model stays inside the skills and principles of DBT rather than producing generic encouragement that sounds therapeutic without being it.

Just-in-time adaptive intervention. The decision of what to deliver, when, and whether to deliver anything at all is the intervention. JITAI makes that decision continuously from the sensed context, offering a DBT skill at the moment it applies, and staying silent when an interruption would do more harm than good.

By targeting transdiagnostic mechanisms such as emotional dysregulation, interpersonal difficulty, and cognitive distortion, the approach is designed to generalize across populations with comorbid mood and anxiety disorders rather than to a single diagnosis.

Funding and External Collaborations

UNDERSTAND is funded by the National Institute of Mental Health (NIMH) at NIH, with the funded project starting 1 August 2026. It is a collaboration between the HXI Lab and the UC San Diego Department of Psychiatry and the Herbert Wertheim School of Public Health, bringing together expertise in DBT, clinical psychology, digital mental health, and ubiquitous sensing.

Preference Learning: Aligning AI to Human Preferences

Tue, 15 Sep 2026 00:00:00 +0000

Overview

Preference-based reinforcement learning aligns AI systems to human judgment by showing annotators two outputs and asking which is better. The annotation protocols, agreement statistics, and training objectives supporting this approach were designed and validated on text.

Much of what this lab builds produces speech rather than text, where timing, prosody, and tone carry meaning alongside the words. While preference learning is well established for language models, its transfer to audio is largely untested: a PRISMA-guided review of roughly 500 papers found that only 6% apply it to audio tasks.

Modality also changes the judgment itself. In a controlled cross-modal study of human and synthetic annotation, identical content produced different preference ratings depending on whether raters read it or heard it. Agreement statistics imported from the text literature can therefore measure something other than what they report.

This work supports the lab’s speech and conversation projects, where feedback on how a clinician sounded, an AI persona that adapts its affect, and a simulated patient whose tone shifts all rest on defensible judgments that one generated utterance is better than another.

Collaborations

This work is carried out with collaborators across UC San Diego, including Prithviraj Ammanabrolu in Computer Science and Engineering, and with Eshin Jolly.

CSE 291A / DSC 266R

Thu, 17 Oct 2024 00:00:00 +0000

Human-Centered AI

Background

Artificial intelligence permeates the fabric of human society, and its ability to generate data-driven insights and recommendations has potential benefits at an unprecedented scale, particularly for underserved and marginalized communities. Together with that promise, AI brings new risks: it can deepen existing inequalities, reinforce injustices, and create deeper disconnects within societies. Many of those risks stem from a growing gap between a technology-centric approach to building AI and the complex socio-cultural contexts the technology ends up embedded in.

This course takes a human-first approach instead. A human-first approach means creating AI systems where human perspective and needs drive the technological innovations throughout every stage of a system’s design: data collection, learning models, inference strategies, interaction paradigms, validation, deployment, evaluation, and maintenance.

Course Description

The course teaches future data scientists how to carry that human-centered approach through the decisions they make across the Data Science Pipeline. One question runs through the whole quarter:

How can we use design thinking to ensure data science and machine learning models are more transparent, approachable, and equitable?

Each week focuses on a specific stage of the pipeline, from use case development and the design phase onward, and pairs it with the human-centered topics and tools that belong at that stage. Students work through case studies and engage directly with the socio-technical questions every data scientist should be able to reason about when designing, building, and deploying AI systems.

The course also brings in outside voices on the themes that cut across the whole pipeline, including sessions on ethics, user-centered design, and safety.

Enrollment

The course is offered jointly as CSE 291A (Topics in AI) and DSC 266R, and is open to graduate students across computer science and data science.

Designing Smart and Autonomous Vehicles

Thu, 25 May 2023 00:00:00 +0000

Overview

Adoption of autonomous vehicles is held back less by the driving than by the passenger’s judgment of it. Trust has to be calibrated: too little and the technology goes unused, too much and it is relied on where it should not be. We study that calibration at the level of the individual, using virtual reality, biometric measurement, human-centered design, and data science, rather than treating drivers as a single population.

Trust is personal. A survey of 1,457 young adults, modeled with machine learning and SHAP, showed that perceptions of AV risks and benefits, attitudes toward feasibility and usability, institutional trust, prior experience, and mental models predict trust, while psychosocial traits and driving style contribute far less.

Explanations carry that trust, and they can fail. In a simulated driving study with 232 participants, errors in an AV’s explanations reduced comfort in relying on the vehicle, confidence in its ability, and explanation satisfaction, even though the driving itself was identical. Perceived harm and driving difficulty amplified the damage, so the contexts where an explanation matters most are where getting it wrong costs most.

Studying any of this requires observing the driver. DriveSimQuest is a VR driving simulator and research platform on the Meta Quest Pro, capturing gaze, facial expression, hand activity, and full-body gesture in real time, so that studying a driver’s affective state is a matter of designing the study rather than building the rig.

The same setting looks different when the driver is a person. Fleet drivers are involved in collisions that impose severe financial costs and endanger lives, and fleet companies rely on one-to-one coaching to prepare them. We characterize that coaching from both sides, surveying coaches and interviewing drivers, and find that manager-led coaching outperforms self-coaching across experiential outcomes.

Funding and External Collaborations

The autonomous vehicle trust work is a collaboration between the UCSD departments of Cognitive Science, Computer Science and Engineering, and external industry partners. The fleet driver coaching research is conducted with Lytx.

Same Words, Different Judgments: How Preferences Vary Across Modalities

Sun, 06 Dec 2026 00:00:00 +0000

Preference-Based Learning in Audio Applications: A Systematic Analysis

Mon, 17 Nov 2025 00:00:00 +0000

Enhancing LLM-Based Just-in-Time Support with Theoretical Guidance from Dialectical Behavior Therapy

Sun, 01 Jun 2025 00:00:00 +0000

What did my car say? impact of autonomous vehicle explanation errors and driving context on comfort, reliance, satisfaction, and driving confidence

Wed, 01 Jan 2025 00:00:00 +0000