Third Multimodal Sensing Grand Challenge for Next-Gen Pain Assessment (AI4Pain 2026)

Participating teams are to develop a machine learning model that classifies sample data in the EmoPain@Home dataset into three levels of pain (Low level pain (LP), Medium level Pain (MP), High level pain (HP)) based on motion capture data input.

Data

• EmoPain@Home (Participants with chronic pain) - This challenge track is centred on data from people with chronic pain in the EmoPain@Home dataset. 526 one-minute segments of joint position data for the right elbow, right wrist, mid spine, hip, right knee, and right ankle. These segments were extracted from 54 unique activity instances that cover 25 activity types from 9 participants with chronic pain. Segments from participants with chronic pain will include pain levels between 0 and 10 for 'no pain' to 'extreme pain' self-reported at the end of the segment. For this challenge track, the 0-to-10 pain level should be recoded as:

Low level Pain (LP) class - pain level less than 3 

Medium level Pain (MP) class - pain level between 3 and 6 (3 and 6 inclusive)

High level Pain (HP) class - pain level higher than 6

• EmoPain@Home (Participants without chronic pain) - 38 unsegmented joint position data collected from 9 healthy participants (i.e. participants without chronic pain) in the EmoPain@Home dataset are also provided. Please note that these data cannot be treated as 'Low level Pain' (LP) class.

•  EmoPain - The EmoPain dataset is additionally provided. The EmoPain dataset includes full-body joints position data and pain labels.

Protocol 

• During the testing phase of the challenge track, participating teams will submit their model's source code, allowing the organisers to (1) run leave-one-activity-instance-out cross-validation evaluation of the model; and (2) assess that the model is a valid machine learning model. Participating teams will need to provide sufficient details about their full machine learning pipeline and dependencies to ensure that the organisers are able to do these. The organisers would share the evaluation outcome with the respective team.

• Up to five model submissions will be allowed per participating team. 

• Participating teams are required to write and submit a paper detailing their methodologies, findings, and insights gained from the challenge, which will be presented in person at the PAAIn Workshop during the ACII 2026 conference in Puebla, Mexico. For more details, please refer to the ACII 2026 submission guidelines. Participating teams may report in their paper results using any standard validation strategies, but please note that performance based on leave-one-activity-instance-out cross-validation will be used for the final ranking for the Grand Challenge. 

• Best Performing award. Valid submissions to the challenge will be ranked based on nested leave-one-activity-instance-out cross-validation performance on the EmoPain@Home dataset. Note that this means a separate validation set representing an unseen activity instance will be additionally left out (separate from the training and test sets) in each cross-validation fold for early stopping, hyperparameter optimization, or similar required by the submitted model.

• Most Innovative award. Beyond the performance-based award for this challenge track, there will also be an innovation-based award. Teams are encouraged to explore rigorous and innovative approaches in terms of feature extraction, machine learning architectures, data augmentation / imputation, and transfer learning (the EmoPain dataset will be provided to facilitate transfer learning). Innovativeness will be assessed by a two-person panel and based on 6 criteria:

1. Architectural novelty: Innovative model design or novel combinations beyond standard architectures, with a clear and convincing rationale

2. Domain knowledge use: Integration of pain- and movement-specific insights into spatial–temporal modeling

3. Generalizability strategy: Use of data augmentation and domain generalization techniques to promote robustness to unseen data, especially to unseen people

4. Training innovation: Creative training pipelines beyond the standard, with a clear and convincing rationale

5. Addressing missing data: Novel approach to address missing data beyond standard techniques, with clear and convincing rationale

6. Overall intent: Demonstrated focus on real-world applicability rather than optimizing only for the EmoPain@Home or EmoPain datasets