ElderMTL: teaching AI to read the emotions of older adults without mistaking wrinkles for sadness

The number of older adults is growing while there are not enough qualified staff to monitor their condition, and sensors or regular check-ups are often unavailable or expensive.

There is also a technical obstacle. Emotion recognition systems are trained on video of young adults, whereas the faces of older people are anatomically different: wrinkles, weaker muscles and slower facial movement. As a result, ordinary models routinely misread emotion.

ElderMTL was developed to solve this, together with the Russian-Armenian University, the Institute for Informatics and Automation Problems of the National Academy of Sciences of Armenia, and Sber AI Lab. The system analyses emotion from video in real time and requires no special equipment.

It assesses three things at once:

• the movement of facial muscles;

• the overall emotion — joy, sadness, anxiety;

• the level of arousal, from calm to agitated.

The key feature is that the model takes the approximate age of the person into account, and therefore distinguishes age-related changes in the face from genuine emotional reactions. Training used a multi-dataset approach that included a dedicated set of videos of older adults.

Accuracy rose from 43% to 74%, and the quality gap between younger and older subjects narrowed by almost half. A language model then turns the technical readings into text that staff can act on.

An important caveat: ElderMTL is an assistant for staff, not a replacement for a doctor or a nurse. Its use requires the consent of the older adults involved and proper protection of their data.

The paper has been accepted at IJCAI-ECAI 2026 and is available as a preprint.