Leily M Behbehani
@leilyb.bsky.social
Clinical Psychology PhD student at Yale | Prior Research Asst at Zucker Hillside/Feinstein Institutes | Haverford College ‘21 | Computational Clinical Science
And yes! I totally agree that examining latent embedding dimensions or supervised projections to characterize the endpoints vs. the gray zone would be a great next step.
December 10, 2025 at 2:20 PM
And yes! I totally agree that examining latent embedding dimensions or supervised projections to characterize the endpoints vs. the gray zone would be a great next step.
Thank you so much! One thing we noticed is the “I don’t know” topic also peaked in the mid range, suggesting people experience the middle of the scale as harder to pin down. So part of the grey zone may be heterogeneity and genuine uncertainty about what the mid range represents.
December 10, 2025 at 2:18 PM
Thank you so much! One thing we noticed is the “I don’t know” topic also peaked in the mid range, suggesting people experience the middle of the scale as harder to pin down. So part of the grey zone may be heterogeneity and genuine uncertainty about what the mid range represents.
A HUGE thank you to all of my collaborators Dr. Ruben Van Genugten, @kathrynfox.bsky.social @sharinahamm.bsky.social @francesghart.bsky.social and my brilliant advisor @shirleybwang.bsky.social for supporting and contributing to this project!
December 9, 2025 at 4:40 PM
A HUGE thank you to all of my collaborators Dr. Ruben Van Genugten, @kathrynfox.bsky.social @sharinahamm.bsky.social @francesghart.bsky.social and my brilliant advisor @shirleybwang.bsky.social for supporting and contributing to this project!
6/6
These findings challenge assumptions in our field: that numeric ratings function as uniformly interpretable indicators of suicidal thinking.
Mid scale ratings often reflect qualitatively different experiences, not merely different intensities.
These findings challenge assumptions in our field: that numeric ratings function as uniformly interpretable indicators of suicidal thinking.
Mid scale ratings often reflect qualitatively different experiences, not merely different intensities.
December 9, 2025 at 4:36 PM
6/6
These findings challenge assumptions in our field: that numeric ratings function as uniformly interpretable indicators of suicidal thinking.
Mid scale ratings often reflect qualitatively different experiences, not merely different intensities.
These findings challenge assumptions in our field: that numeric ratings function as uniformly interpretable indicators of suicidal thinking.
Mid scale ratings often reflect qualitatively different experiences, not merely different intensities.
5/6
What we found:
1️⃣ Shared meaning emerged only at the ends of the scale.
2️⃣ The middle of the scale was a conceptual gray zone.
3️⃣ Between person consistency never exceeded ~ 20%.
4️⃣ Within person consistency was strongest at scale endpoints.
5️⃣LLM refinement outperformed BERTopic alone.
What we found:
1️⃣ Shared meaning emerged only at the ends of the scale.
2️⃣ The middle of the scale was a conceptual gray zone.
3️⃣ Between person consistency never exceeded ~ 20%.
4️⃣ Within person consistency was strongest at scale endpoints.
5️⃣LLM refinement outperformed BERTopic alone.
December 9, 2025 at 4:34 PM
5/6
What we found:
1️⃣ Shared meaning emerged only at the ends of the scale.
2️⃣ The middle of the scale was a conceptual gray zone.
3️⃣ Between person consistency never exceeded ~ 20%.
4️⃣ Within person consistency was strongest at scale endpoints.
5️⃣LLM refinement outperformed BERTopic alone.
What we found:
1️⃣ Shared meaning emerged only at the ends of the scale.
2️⃣ The middle of the scale was a conceptual gray zone.
3️⃣ Between person consistency never exceeded ~ 20%.
4️⃣ Within person consistency was strongest at scale endpoints.
5️⃣LLM refinement outperformed BERTopic alone.
4/6
Using a 2 stage NLP pipeline (BERTopic -> LLM refinement), we extracted coherent themes from thousands of participant responses and mapped them onto the numeric rating scale.
We also looked at within person consistency (across time) and between person consistency (across people).
Using a 2 stage NLP pipeline (BERTopic -> LLM refinement), we extracted coherent themes from thousands of participant responses and mapped them onto the numeric rating scale.
We also looked at within person consistency (across time) and between person consistency (across people).
December 9, 2025 at 4:30 PM
4/6
Using a 2 stage NLP pipeline (BERTopic -> LLM refinement), we extracted coherent themes from thousands of participant responses and mapped them onto the numeric rating scale.
We also looked at within person consistency (across time) and between person consistency (across people).
Using a 2 stage NLP pipeline (BERTopic -> LLM refinement), we extracted coherent themes from thousands of participant responses and mapped them onto the numeric rating scale.
We also looked at within person consistency (across time) and between person consistency (across people).
3/6
Across 2 independent cohorts of adolescents and young adults, participants were randomly assigned number ratings (0-10) of suicide urge and provided open ended descriptions of what thoughts they would be having at those ratings.
Across 2 independent cohorts of adolescents and young adults, participants were randomly assigned number ratings (0-10) of suicide urge and provided open ended descriptions of what thoughts they would be having at those ratings.
December 9, 2025 at 4:28 PM
3/6
Across 2 independent cohorts of adolescents and young adults, participants were randomly assigned number ratings (0-10) of suicide urge and provided open ended descriptions of what thoughts they would be having at those ratings.
Across 2 independent cohorts of adolescents and young adults, participants were randomly assigned number ratings (0-10) of suicide urge and provided open ended descriptions of what thoughts they would be having at those ratings.
2/6
We usually treat numeric ratings as if they directly reflect suicide risk. But psychological measurement tells us something trickier:
People map internal states onto numbers in highly personal ways.
We usually treat numeric ratings as if they directly reflect suicide risk. But psychological measurement tells us something trickier:
People map internal states onto numbers in highly personal ways.
December 9, 2025 at 4:25 PM
2/6
We usually treat numeric ratings as if they directly reflect suicide risk. But psychological measurement tells us something trickier:
People map internal states onto numbers in highly personal ways.
We usually treat numeric ratings as if they directly reflect suicide risk. But psychological measurement tells us something trickier:
People map internal states onto numbers in highly personal ways.
Reposted by Leily M Behbehani
(6/6) Big thank you to co-authors @sharinahamm.bsky.social , @leilyb.bsky.social , @kathrynfox.bsky.social , @shirleybwang.bsky.social- and bluesky-less Kathryn Coniglio, Erin Reilly, and Leah Somerville- for their support in this project!
July 25, 2025 at 6:26 PM
(6/6) Big thank you to co-authors @sharinahamm.bsky.social , @leilyb.bsky.social , @kathrynfox.bsky.social , @shirleybwang.bsky.social- and bluesky-less Kathryn Coniglio, Erin Reilly, and Leah Somerville- for their support in this project!