63/100

Moderate evidence

Cite
More
EmailLinkedInX

AI reading of 8,348 antidepressant reviews matched trial data on tolerability

Study title: Can AI-assisted analysis of antidepressant drug online reviewers (ADORE) inform real-world use of antidepressants?

Evidence type
Observational study
Sample
n = 8,348
Sources
1 sourcePeer-reviewed
Published
2 min read

Study title: Can AI-assisted analysis of antidepressant drug online reviewers (ADORE) inform real-world use of antidepressants?

Researchers analysed 8,348 patient-written reviews of 16 antidepressants taken for major depressive disorder, posted to Drugs.com between February 2008 and March 2025. Two AI models, Claude Sonnet 4.5 and Gemini 2.5 Pro, sorted each review across 17 clinical dimensions. Each drug's share of top ratings correlated strongly (ρ = +0.79) with retention odds ratios from the 2018 Cipriani network meta-analysis.

Read the study abstract

Antidepressant selection is constrained by limited comparative-effectiveness data, with randomized trials measuring symptom reduction rather than patient-reported tolerability. We tested whether AI-assisted analysis of patient-written online reviews recovers drug-level tolerability signals that converge with the strongest randomized evidence. We did a cross-sectional analysis of 8348 reviews of 16 antidepressants for major depressive disorder posted to Drugs.com (February 2008-March 2025; ≥1 month reported use).…

Can AI-assisted analysis of antidepressant drug online reviewers (ADORE) inform real-world use of antidepressants?
"Prélèvement sérologie au Laboratoire Hopital Douala Cameroun 10" by Happi Raphael, CC BY-SA 4.0.

The gain

In a cross-sectional analysis of 8,348 Drugs.com reviews of 16 antidepressants, two AI models agreed closely (mean κ 0.81), and review ratings correlated strongly with randomized-trial retention odds ratios (ρ = +0.79).

In brief

The result matters because randomized trials usually measure symptom reduction, not how well patients tolerate a drug. Top ratings did not track efficacy odds ratios (ρ = -0.14, p = 0.64), and emotional blunting was the strongest negative predictor of satisfaction (odds ratio 0.20). The authors also found substantial differences within the SNRI class.

The findings are presented as a hypothesis-generating signal. Reviews come from one platform and from self-selected users, the drug-level comparison rests on few agents, and the design is cross-sectional and ecological.

Main points

  1. The study covered 8,348 Drugs.com reviews of 16 antidepressants for major depressive disorder, posted February 2008 to March 2025, from users reporting at least one month of use.
  2. Claude Sonnet 4.5 and Gemini 2.5 Pro classified each review across 17 clinical dimensions using identical structured prompts.
  3. Mean agreement between the two AI models was κ = 0.81; two experts rated a 100-review subsample and agreed with the AI consensus at mean κ = 0.75 and 0.80.
  4. Top-rating prevalence correlated with retention odds ratios from the 2018 Cipriani network meta-analysis (ρ = +0.79, p < 0.001).
  5. Top rating did not correlate with efficacy odds ratios (ρ = -0.14, p = 0.64).
  6. Emotional blunting was the strongest negative predictor of satisfaction, with an odds ratio of 0.20.
  7. Heterogeneity within the SNRI class was substantial.

The full story

The team used a cross-sectional design. Only reviews describing at least one month of use were included. Two large language models classified every review with the same structured prompts, and a stratified 100-review subsample (1,700 classifications per reviewer) was rated independently by two experts. All agreement statistics were highly significant.

To test validity, the researchers compared each drug's share of top ratings with the retention and efficacy odds ratios from Cipriani and colleagues' 2018 network meta-analysis, using Spearman correlation. Ratings followed retention closely but not efficacy.

The authors conclude that, among contemporary antidepressants with comparable efficacy, tolerability, especially emotional blunting, correlates most with patient satisfaction. They describe the SNRI differences as a hypothesis-generating signal.

Sources

1 on file
  1. Journal of Affective Disordersdoi:10.1016/j.jad.2026.122743
    Peer-reviewed2026-11-15

Talk it over

Three questions to start on, two more if you have time
Top ratings tracked retention odds ratios from the 2018 Cipriani network meta-analysis (ρ = +0.79) but not efficacy odds ratios (ρ = -0.14). What does that suggest patients are rating when they review an antidepressant?

Start with a question

AI reading of 8,348 antidepressant reviews matched trial data on tolerabilityhttps://truvace.com/item/can-ai-assisted-analysis-of-antidepressant-drug-online-reviewers-adore-inform-re

  1. Top ratings tracked retention odds ratios from the 2018 Cipriani network meta-analysis (ρ = +0.79) but not efficacy odds ratios (ρ = -0.14). What does that suggest patients are rating when they review an antidepressant?

    Answer in the thread
  2. The 8,348 reviews came from self-selected Drugs.com users who reported at least one month of use. Who might be missing from that group, and how could that change what reviews say about emotional blunting?

    Answer in the thread
  3. Two AI models, Claude Sonnet 4.5 and Gemini 2.5 Pro, agreed closely (mean κ 0.81), and two experts rated a 100-review subsample. What further check would make you more confident the AI sorted the reviews correctly?

    Answer in the thread
If you have more time
  • The authors describe substantial differences within the SNRI class as a hypothesis-generating signal. What would you want to look up or ask next before treating that signal as meaningful?

    Answer in the thread
  • What would a study look like that follows the same patients over time to test whether emotional blunting, with its odds ratio of 0.20, predicts people stopping a drug?

    Answer in the thread

What readers say