section

developmental history

probe population-level factor responses to categories of discrete events also labeled via LM-QA.

a stability-based dimensionality criterion

NMF

structured question answering (LM-QA) paradigm

o label a set of 52 consensus features of depression

First, we demonstrate that the chosen LM is able to produce automated, dense, high-dimensional time series annotations of mental health–related symptom scores from a cohort of hundreds of individuals, followed over a full year, by answering natural language questions taken from a broad battery of depression-related symptoms collated post hoc, without directly interacting with subjects.

aken together, this work demonstrates a practical and generalizable blueprint for using LMs and large, naturalistic, long-timescale behavior observations to identify candidates for granular and robust mental health phenotypes, and to facilitate automated quantification of those phenotypes with data-derived and interpretable behavioral biomarkers. We propose, in line with recent perspectives (Fried, Flake, and Robinaugh 2022), that this blueprint serves as a key building block of a new iterative program for the development and interrogation of better-circumscribed categories in the spectrum of mental life for guiding mechanistic inquiry, clinical decision-making, and public mental health policy.

Here, we apply an off-the-shelf, moderate-sized LM to a recently-proposed framework for data-driven discovery of granular mental health phenotypes (Fried, Flake, and Robinaugh 2022). First, we demonstrate that the chosen LM is able to produce automated, dense, high-dimensional time series annotations of mental health–related symptom scores from a cohort of hundreds of individuals, followed over a full year, by answering natural language questions taken from a broad battery of depression-related symptoms collated post hoc, without directly interacting with subjects. These high-dimensional annotations have a highly interpretable latent factor structure which is more robust than that of some commonly-used clinical depression inventories (Shafer 2006; Furukawa et al. 2005).1 After characterizing the collective time-evolution in these mental factors within individuals in this cohort, we hypothesize that the time-evolution of each individual’s mental state falls within one of a finite collection of common mental dynamical regimes. To tractably address this hypothesis, we discretize the instantaneous mental factor vectors across the cohort into a finite collection of common discrete states, then search for groups of individuals with similar state-state transition probabilities. Finally, we suggest etiological differences in the development of these dynamical regimes in individuals through LM-assisted qualitative analysis of subjects’ self-reported histories.

Here, we apply an off-the-shelf, moderate-sized LM to a recently-proposed framework for data-driven discovery of granular mental health phenotypes

These advancements raise the promise of being able to leverage the incredible richness of large volumes of extant naturalistic text data to drive foundational research advances in psychiatry.

Meanwhile, as these concerns have been brewing, advancements in the development of large, domain-general language models (LMs) have paved the way for radical shifts in empirical approaches to natural language processing tasks like annotation and information extraction (CITES). Most intriguingly, recent work has begun to suggest the tremendous promise that these models might hold for quantitative phenotyping in mental health in their ability to analyze naturalistic text data at massive scale, with LM-based methods demonstrating good performance in tasks like zero-shot psychiatric diagnosis (CITE).

TE)

as these concerns have been brewing

research optimized for the chosen metric

metrics

by lumping populations for whom an intervention is effective with those for whom it is not

improperly delineating indications without a truly shared etiology and mechanism can lead efforts to discover and validate therapeutic efficacy far astray

indications

concerns of core definitions and measurements have large downstream

Both the clinical delineation of which ways of mental being which constitute “major depression” as a unified entity

two notions interact in a closed loop

However, a spotlight of recent interest has been focused instead on more foundational questions:

converging work in the last few decades has paved the way toward interventions that are hoped to supply more efficacious levers than current agents (Carhart-Harris et al. 2021).

storically centered on variants in neuromodulatory signaling (Owens et al. 1997; Hirschfeld 2000; Walker 2013; Dale, Bang-Andersen, and Sánchez 2015

clinical need

,

major depression

(Jorm et al. 2017

to similarly uncovering dynamical regimes in model systems to serve as granular and robustly-quantified phenotypes in investigating, and reverse-generalizing to humans, neurobiological mechanisms impacting long-timescale structure of behavior and mental life.

using similar inputs from artificial intelligence–driven behavioral segmentation tools applied to observation modalities adapted to other organisms

to other domains of psychiatry

data-driven hypotheses

cluster the structure of how individuals’ mental states unfold in time into a small set of “dynamical regimes”, phenotypes capturing subgroups with similar state-transition propensities; and

in self-reported developmental contributions behind these inferred dynamical phenotypes

(1) estimate high-dimensional time series of symptoms related to anxiety and depression post-hoc with no direct subject input;

misalignment with underlying biology, and incomplete generalization to model organisms

clinical reliability

shortcomings