Research Project
Transcension
Adaptive guidance from real-time physiological signals
Can a system decide what to say next based on how a person's brain and body are actually responding, rather than following a fixed script? Transcension studies that question in guided relaxation and hypnosis, using multi-channel EEG, biofeedback, and synchronized audio.
The Question
A skilled practitioner watches the person in front of them and adjusts: slowing down, repeating, or moving on when they see the person is ready. A recorded script can't do that. Transcension asks whether a system can make the same kind of adjustment from measured signals: galvanic skin response, skin temperature, and multi-channel EEG.
The engineering problem is broader than hypnosis. It's about measuring a person's state reliably, choosing the next step from that measurement, and proving that adapting actually beats a fixed sequence.
The project brings together my graduate study in transpersonal psychology and the biofeedback hardware I prototyped at Somadome.
System Design
1Data Collection & Labeling
Volunteers take part in live sessions guided by a human practitioner. Each session records:
- Multi-channel EEG from scalp sensors.
- Galvanic skin response and skin temperature from fingertip sensors.
- Synchronized audio of both the practitioner and the participant.
Every sentence the practitioner speaks is labeled with its stage in the session (for example, introduction, induction, deepener, or suggestion), so each prompt can be lined up against the signals that followed it.
2Model Training
A model learns the relationship between what was said, how the participant responded verbally, and how their signals changed. The goal is to learn not only which prompt fits a moment, but when to deliver it.
3Adaptive Selection
In operation, a reinforcement learning loop picks the next prompt from a labeled library. Its reward is the participant's measured state moving toward a target profile. The initial target draws on Judith Pennington's "Awakened Mind" brainwave research, which I treat as a hypothesis to test rather than settled science.
Rigor & Ethics
Informed Consent
Participants know exactly what is recorded and why, and they can stop a session or withdraw their data at any time.
Human in the Loop
A trained practitioner leads every session. Until the system is validated, it only suggests the next prompt, and the practitioner decides.
Privacy by Design
Physiological recordings are kept apart from identity data and linked only by pseudonymous IDs. It's the same pattern I used for personal data at Voxivium.
Honest Validation
Adaptive sessions are compared against scripted sessions and human-led sessions, using measures defined before the data is collected. Each person's baseline is measured, because one signal profile won't fit everyone.
Engineering Challenges
Data Acquisition
The biggest challenge is a large, diverse, high-quality dataset. That means recruiting volunteers and recording hundreds of hours of carefully labeled sessions, with privacy and ethics built in from the start.
Signal Quality & Sync
EEG, skin, and audio streams have to share one precise clock. EEG is noisy: eye blinks, muscle movement, and poor electrode contact all have to be detected and removed before any signal can be trusted.
Cost & Complexity
Research-grade EEG and biometric sensors are expensive, and training on multi-modal data (audio, EEG, and body signals) takes substantial compute. The adaptive policy also has to stay varied and natural rather than repetitive.