Time-Series & Sensors

Cross-Device Domain Adaptation for FRDA Assessment

Domain Adaptation for IMU Data to Enhance Objective Assessment of Friedreich’s Ataxia

r = 0.78 (AIM-C) · 0.75 (AIM-S)

When every patient is precious data, borrow knowledge across devices: a model trained on one instrumented tool helps another one learn.

The problem

Friedreich ataxia is rare, so each sensor device has too little data to train a strong deep model on its own.

The approach

A CNN learns features directly from raw IMU signals, and a two-stage domain-adversarial training scheme transfers what it learns between the instrumented spoon and cup.

The outcome

Correlation with clinical severity rose from 0.74 to 0.78 for the cup and from 0.69 to 0.75 for the spoon.

In detail

Developed a two-stage CNN framework that learns directly from raw IMU signals and transfers knowledge between AIM-C and AIM-S. Domain-adversarial pre-training encourages device-invariant features, followed by supervised fine-tuning on the target device. The proposed approach improved correlation with mFARS to 0.78 for AIM-C and 0.75 for AIM-S, showing the value of cross-device learning in data-scarce rare-disease settings.

Built with

• Domain Adaptation
• CNNs
• Wearable Sensors

Source

• Domain Adaptation for IMU Data to Enhance Objective Assessment of Friedreich Ataxia, IEEE Journal of Biomedical and Health Informatics, 2026

More in Time-Series & Sensors

• Probabilistic Deep Ensembles for FRDA Severity Assessment
• AIM-M: Multiview IMU Assessment of Friedreich Ataxia
• DTW-Based Outlier Removal for Reliable FRDA Measurement

Next project: ClusterGuardFL: Secure Cluster-Weighted Client Aggregation

By Kanishka Ranaweera, PhD, Machine Learning Engineer in Melbourne, Victoria, Australia. hello@kanishkabuilt.me