23. COULD A PHONECALL TO AN AI SIMPLIFY MEASUREMENT-BASED CARE FOR OLDER ADULTS: PROOF OF CONCEPT

Abstract

Introduction

The value of measurement-based care (MBC) in psychiatry and primary care is well-established. Implementation of MBC frequently leverages digital approaches including apps or ecological momentary assessment. However, for older adults only have low digital literacy simplifying the process further may lead to more efficient collection of meaningful clinical data. Voice as a modality can offer several advantages over web or text interfaces for older adults: significantly improve accessibility, reduce cognitive load, and lower physical barriers. A completely hands-free, voice only interaction provides a high level of accessibility and independence for users. Thus, we tested the feasibility of connecting with an AI via a simple phone call, for mental health screening

Methods

This work represents the first application of the 3rd Ear voice interviewing platform for building conversational agents that dynamically adapt their questioning to each patient’s responses. By focusing on interviewing rather than offering clinical judgment or recommendations, the platform minimizes the risks of AI hallucinations and bias. The platform orchestrates adaptive, goal-directed dialogue that aligns with established screening tools while unveiling a more nuanced understanding of the patient’s story.

Stacy is a phone-based bot built on the third ear platform. The bot is designed to administer the PHQ 9 via interview/conversation. for this initial proof of concept, we implemented two screening calls, conducted by the investigators.

After the user calls the phone, a neural network model transcribes their speech into text. This text is then analyzed for evidence of the depressive system in question and then the derived evidence is passed to the conversational engine, which then guides an open-source large language model to produce a response. This is finally passed to a state-of-the art text-to-speech system which then produces a voice response back on the phone call.

Results

Our primary finding was that Stacy can successfully carry out the PHQ-9, ask for elaborations when needed to get more confidence, and fill out the paper form. A typical interview took 3-5 minutes, with an average latency of 1200 milli seconds, which is about 100 ms/question. The users reported minimal delay in voice response and smooth conversational flow. Notably, the users reported that when they interrupted the Stacy bot mid conversation, it was able to pivot comparably to a human.

Conclusions

Using phone based AI tools has the potential to simplify and improve the efficacy of measurement based care, particularly for older adults. Our proof of concept focused only on a single scale – the PHQ 9. However, this approach can be expanded to include multiple measures and has the potential to reduce cost while improving data collection by deploying a conversational/interview approach. Our AI interviewers not only gather the standardized screening data clinicians rely on, but also uncover richer patient stories that can inform better diagnoses and follow-up care. This bridges the gap between simple surveys and clinician-led interviews, enabling more nuanced, accessible, and beneficial patient assessments at scale.