A feasibility pilot study of an AI-based emotional support application (TerapIA) in a Spanish-speaking community population: recruitment, retention and parameter estimation for a controlled trial

Before running a controlled trial of an emotional support application, it is worth checking that the trial can actually be run. We present a feasibility pilot study of TerapIA, the AI-based emotional support application of LenguajeNatural.AI, with a pre/post design and a control arm, in a Spanish-speaking, non-clinical community population recruited through Fundación Social Padre Ángel. The primary outcome is feasibility: recruitment, baseline completion, 21-day retention and data quality. Anxiety, measured with the GAD-7, is a secondary outcome for which the study was not powered. We recruited 67 participants in 6 days, 66 of whom (98.5%) completed the baseline measure; 32 (47.8% of those enrolled) completed the 21-day measure. In the intervention arm the mean GAD-7 score went from 9.44 to 8.00 (-1.44 points; 95% CI -2.79 to -0.09), while in the control arm it did not change (7.00 to 7.00). The baseline-adjusted difference between arms was -0.99 points (95% CI -3.67 to 1.70; p=0.459), with g=-0.45 (95% CI -1.29 to 0.40) in the direction favoring the intervention. All secondary indicators point the same way, but with 7 controls and 25 intervention participants with follow-up the minimum detectable effect at 80% power is g=1.24, and the estimate changes sign when a single control participant is removed. The direction observed cannot therefore be separated from chance. The clearest finding concerns the process: recruitment and baseline measurement work well, and the bottleneck is the follow-up, where 90% of responses arrive within about six days and later reminders recover very few people. Assuming the effect that recent meta-analyses attribute to conversational agents on anxiety (SMD_post=0.28), a baseline-adjusted analysis and 1:1 allocation, a controlled trial would need 222 complete follow-ups, that is, about 370 participants to recruit with 60% retention, or 465 with the retention observed here.