Research / published + ongoing

What should an adaptive or generated system do when its inputs are uncertain and a person must live with the output?

My recent work approaches that question from two directions: audio-to-visual mappings that people experience differently, and clinical referral drafts that must remain subject to human review.

Publication

2026

Co-author Pervasive Health 2025

An Empirical Study on the Design of Dynamic Music Visualization for Therapy Based on Synesthesia Feature Mapping

Hua Ma, Qin Zhou, Yixin Chen, Qixuan Chu, Dujiao Yang, and Sun Xu. Springer, LNICST 693, pp. 512-527 (2026).

Open DOI ↗

What we tested

A 30-participant controlled comparison of soft and intense music-driven visual conditions, using PSS-10, PANAS, aesthetic-pleasure measures, and semi-structured interviews.

My role

I built the TouchDesigner visualization system and audio-to-visual mappings, contributed to the study work, and co-authored the paper.

What the evidence says

The paper reports stronger short-term relief of negative affect for the softer condition and a relationship between aesthetic pleasure, positive affect, and stress relief.

What it does not say

A small, homogeneous sample in one controlled setting cannot establish long-term therapeutic effectiveness or a universally preferred visual mapping.

Questions in progress

I separate finished evidence from work that is still being formed.

Each thread below states the question, the evidence I actually have, and the boundary I do not want the portfolio to blur.

01 Ongoing system work

Generated drafts that remain reviewable

Question: How should an AI-generated draft preserve context, show what happened, and give a domain expert a safe path to correct or reject it?
Evidence so far: In Remeda, I built a backend referral-draft workflow with generation metadata, persistence, doctor-only review actions, immutable audit events, duplicate-review protection, and a deterministic fallback.
Scope and limits: This work validates workflow and review boundaries; it is not evidence of clinical effectiveness or a deployed provider-backed generation system.

02 Published study

Adaptive music visualization

Question: Which audio-to-visual mappings support relaxation, and where do individual perception and context make one fixed mapping insufficient?
Evidence so far: I built the TouchDesigner visualization system and feature mappings, contributed to the study work, and co-authored a 2026 Springer chapter. The published study reports a 30-participant comparison of soft and intense visual conditions.
Scope and limits: The study measured short-term responses in a small, homogeneous sample and does not establish long-term therapeutic effectiveness.

03 Current manuscript work

From fixed mappings to cautious adaptation

Question: Can multimodal signals inform adaptive media without turning noisy human data into overconfident personalization?
Evidence so far: During the final phase of my UNNC research assistantship, I drafted and revised manuscript sections and synthesized literature across affective computing, multimodal sensing, and conservative offline RL methods including BCQ, IQL, and CQL.
Scope and limits: This is ongoing manuscript research. I present the framing and literature work as my contribution, not a completed or clinically validated adaptive system.

How I read for the MedQX review

Question → evidence → system implication → uncertainty

I am trying to make the literature review useful to implementation, not just complete as a bibliography. For each paper I track the population or setting, what the system observes, what it produces, who reviews the output, and what kind of evaluation supports the claim.

Research contact

If one of these questions overlaps with your work, I’d be glad to hear from you.

[email protected]