How we model emotions in AI clinical simulations

How we model emotions in AI clinical simulations

Most AI systems treat emotion like a label. For example happy, sad or angry. Then they pick one, attach it to a response, and move on.

But that is not how emotions work in real clinical conversations.

A patient does not suddenly become frustrated, but builds that state over time. Trust does not appear instantly because of one empathetic sentence, but rises gradually through repeated signals of safety, validation, and agency. Anxiety does not stay fixed after one trigger, but can fade when the conversation shifts tone.

That is the problem we are trying to solve inside EmpatiQ.

We use a leaky integrate-and-fire model, adapted from computational neuroscience, to drive the emotional state of our simulated patients. In plain terms, every conversation creates emotional signals.

A dismissive phrase might raise frustration. A validating response might build trust. A rushed explanation might increase anxiety. A moment of silence might reduce pressure. Those signals accumulate over time, but they also fade if nothing reinforces them.

That is the “leaky” part. A patient who felt anxious earlier in the conversation may start to settle if the clinician changes approach.

The “fire” part is what happens when an emotional state crosses a threshold. At some point, frustration is no longer just slightly higher. The patient may shut down, become defensive, disengage, or push back.

That is how many difficult clinical conversations feel. They do not move in smooth, predictable lines, but shift when enough emotional pressure builds. We combine this with Plutchik’s emotion model, where basic emotions can form dyads.

For example: anticipation plus fear can become anxiety. Joy plus trust can become love. Surprise plus sadness can become disapproval.

That gives us a structured way to model mixed emotional states without writing a script for every possible reaction. So the simulated patient is not following a prompt that says "act upset now.", but their emotional state changes underneath the conversation.

If a clinician listens well, validates, and gives the patient agency, the patient may soften. If the clinician rushes, dismisses, or pressures them, the same patient may resist or shut down. We did not write every branch in advance. The emotional state evolved differently and that is why this matters for clinical training.

The goal is not to rehearse against a fixed script. The goal is to practice reading emotional changes as they happen, and responding before the conversation breaks, because in healthcare, the hardest part is often not the protocol.

It is what happens emotionally while you are trying to follow it.