Research
Depression is, at its core, a disorder of emotion: of feelings that become rigid, self-reinforcing, and hard to escape. Yet emotions are usually measured coarsely: a questionnaire every few weeks, a single score per session. The CARED Lab, directed by Hadar Fisher, PhD, uses computational methods, including large language models (LLMs), to measure emotional processes as they actually unfold, moment to moment in daily life and session by session in treatment, and to turn those measurements into a better understanding of how depression develops and how treatment works.
Click any area to read more.
How emotions move, get stuck, and recover in daily life, and what that reveals about who will become depressed.
Read moreA single sad moment is not depression. Depression emerges when emotional states start feeding on themselves and each other: sadness invites rumination, rumination invites withdrawal, withdrawal deepens sadness. We treat emotional life as a dynamic system with feedback loops, inertia, and equilibria that a person keeps returning to. What matters is not only how someone feels, but how their feelings move: whether emotions recover flexibly after a bad moment or get stuck in self-reinforcing loops.
Studying these dynamics requires dense measurement over time. Using ecological momentary assessment (EMA) and dynamic modeling, we study how the temporal structure of emotion (rigidity, feedback loops, and the equilibria people return to) prospectively predicts depressive symptoms, with a focus on adolescence, when depression risk rises sharply.
EMAemotion rigiditydynamic systemsadolescence
What people say, and how they say it, carries rich signals about their emotional state.
Read moreLanguage is the most natural record of emotional life we have, and modern NLP and large language models (LLMs) make it measurable at scale. We develop and evaluate language-based approaches for detecting depression and tracking negative emotions, from daily-life speech and text messages to full therapy transcripts.
Our recent systematic review and meta-analysis in npj Digital Medicine takes stock of the whole field of language-based depression detection with machine learning: what works, what is overclaimed, and what evidence is still needed before these tools can be trusted in the clinic.
NLPLLMsmachine learningmeta-analysis
Smartphones passively capture movement, sleep, sociability, and routine. What can that tell us about depression?
Read moreThe phone in a person's pocket quietly records behavior that questionnaires can only ask about: how much they move, sleep, leave the house, and reach out to others. We test how these passive signals, combined with LLM-derived ratings, can measure clinically meaningful constructs, like behavioral activation during treatment, and support personalized early detection of depression onset.
We are equally interested in the limits of this technology. Passive sensing comes with real pitfalls of validity, privacy, and interpretation, and we study the pathway to clinical integration as carefully as the sensors themselves.
passive sensingsmartphonesearly detection
How emotional experience, expression, and the patient-therapist relationship drive therapeutic change.
Read moreThe therapy room is where emotional change is supposed to happen, and we treat it as a natural laboratory. We study how emotional experience, emotional processing, and even facial expression shift between and within sessions, and for whom these shifts drive recovery.
We also study the relationship in which this change unfolds: the patient-therapist alliance, its biological correlates such as oxytocin and cortisol, and its costs as well as its benefits. Being deeply attuned to another person's emotions helps treatment, but our work shows it can also carry a hidden price for both patient and therapist.
psychotherapy processemotional experiencealliancefacial expression
People differ in why they are depressed, so they should differ in what treatment they need.
Read moreTwo people with the same diagnosis can be depressed for very different reasons: one trapped in rumination, another cut off from reward, a third stuck in an interpersonal pattern that plays out in therapy itself. We study depression from a personalized, within-person perspective, asking what maintains this person's depression and what needs to change for this person to get better.
Methodologically, this means combining two toolkits that are usually kept apart. Data-driven prediction tells us what will happen; theory-driven explanation tells us why. We think clinical science needs both, integrated rather than siloed, to move toward measurement-based, personalized care.
precision psychiatrymechanisms of changeprediction
Computational tools that read emotion raise real questions of validity, privacy, and ethics.
Read moreAlgorithms that infer how people feel, from emotion classifiers to large language models (LLMs), are no longer science fiction: they are being built into products, workplaces, and health care. Because these tools increasingly touch real patients, we study their validity, limits, and ethics with the same rigor we apply to the models themselves.
Our position, argued in Nature, is that emotional AI is here, and the task for clinical science is to shape it rather than shun it: to set rigorous evaluation standards so that what reaches patients actually helps them.
ethicsvalidityclinical integration
Ecological momentary assessment · dynamic systems and time-series modeling · natural language processing and large language models · supervised machine learning and prediction · facial expression analysis · passive smartphone sensing · randomized controlled trials and session-by-session psychotherapy process data · systematic review and meta-analysis.
