Language model ratings of depression often reflect rater not patient
Language-model ratings of depression reflect the rater more than the patient
Computation and LanguageArtificial Intelligence
Summary
Depression is hard to measure objectively because there are no simple tests. This paper looks at how computer language models rate how depressed someone is based on interviews. The authors found that differences between models affected ratings more than differences between patients. Even models with similar accuracy disagreed a lot on whether someone should be flagged for depression. Adjusting the models improved accuracy and agreement but didn’t eliminate disagreements about individuals.
What this means in practice
- •For mental health software developers: Improve depression screening tools by calibrating language models to reduce disagreement between automated raters on patient symptom scores.
- •For clinical data analysts: Use model comparison and calibration methods to understand variability in automated depression ratings and improve decision consistency.
Authors
Baihan Lin
Abstract
Depression has no diagnostic blood test. Language models promise tireless, consistent assessment, but can accurate raters disagree about individuals? We pre-registered 880 language-model raters, crossing 11 open models with prompting and scoring choices, and applied them to 189 interviews against the eight-item Patient Health Questionnaire. Model choice explained 30.0% of summed-symptom score variance, stable participant differences 10.5%. Two randomly drawn raters with area under the receiver operating characteristic curve (AUC) >= 0.70 disagreed on screening decisions for 40% of participants, on average. Average over-rating governed how many were flagged, yet equal-capacity raters chose differently for about one participant in five. A locked analysis of 86 new interviews reproduced the main pre-registered findings. Exploratory recalibration with 40 labelled participants raised accuracy from about 60% to 75% and halved disagreement, leaving one participant in five decided differently. Calibration repaired much of the rater dependence without securing agreement about individuals.