2026 Finland Tech Arena
2026 Finland Tech Arena
AI in Health
TOPIC TITLE DESCRIPTION
Access to Example Code, Sample Data, and Proposal Template
Stress or Sweat? Decoding the Body's Signals from the Wrist
A racing heart can mean a stressful meeting, a hard sprint, or a steady jog, and today's wearables often can't tell the difference. Using only heart (PPG/IBI), skin conductance (EDA), and skin-temperature signals, with no motion data, build a model that classifies a person's state as rest/recovery, acute stress, or physical activity. Solve a real health-tech problem, and top teams pitch to Huawei experts at the final in Helsinki.
The goal of this challenge is to build a model that classifies a person's physiological state as rest/recovery, acute stress, or physical activity using only wrist-worn PPG/BVP/IBI, EDA, and skin-temperature signals, without any motion or activity data.
Requirements
Your model should be trained and be compatible with the following versions of dependencies:
Your model should be trained and be compatible with the following versions of dependencies:
- Python 3.13
All deliverables must be in Python. - Only 1 CPU core is used and multi-threading is disabled when the submission runs.
Dataset
Training dataset is provided as one zip file, which contains the data and documentation of the data.
Training dataset is provided as one zip file, which contains the data and documentation of the data.
Model inference
- Model is expected to take a Pandas timeseries DataFrame as an input. The DataFrame will contain following columns: BVP, EDA, HR, IBI_dt, IBI_duration, Temperature
- The model produces a Pandas timeseries DataFrame with column “Activity” that contains a detected activity from these options: Stress, physical activity, Rest
Model submission
Create a Python file exposing get_model() that returns a class implementing ADM_Base base class as defined in adm_base_class.py. The trained model in ONNX format must also be uploaded next to the source file, using the same basename and .onnx extension.
Example (your_model.py): [CODE SAMPLE]
These two files should be uploaded in one .zip file. For the next submission, edit the previous one by changing the .zip file.
Create a Python file exposing get_model() that returns a class implementing ADM_Base base class as defined in adm_base_class.py. The trained model in ONNX format must also be uploaded next to the source file, using the same basename and .onnx extension.
Example (your_model.py): [CODE SAMPLE]
These two files should be uploaded in one .zip file. For the next submission, edit the previous one by changing the .zip file.
Model evaluation
The model is evaluated against an evaluation data set containing different activities.
The model quality score is based on macro-F1 score, which is an average of F1 scores of Stress, physical activity, and Rest states. The final quality score is macro-F1 * 80 producing a score in range 0-80.
The model speed score is based on mean execution time (MET) of 50 runs of the detect() function of ADM_Base(). The final speed score is (1.0s – MET) * 20.0 producing a score in range 0-2 (MET > 1.0s produces 0 points).
The final score is: quality score (0, 80) + time score (0, 20) = total (0, 100)
The model is evaluated against an evaluation data set containing different activities.
The model quality score is based on macro-F1 score, which is an average of F1 scores of Stress, physical activity, and Rest states. The final quality score is macro-F1 * 80 producing a score in range 0-80.
The model speed score is based on mean execution time (MET) of 50 runs of the detect() function of ADM_Base(). The final speed score is (1.0s – MET) * 20.0 producing a score in range 0-2 (MET > 1.0s produces 0 points).
The final score is: quality score (0, 80) + time score (0, 20) = total (0, 100)
Add ons
- Provided input
- Expected output
- Reference(s)
- NDA Requirement