Time-Series & Sensors

DTW-Based Outlier Removal for Reliable FRDA Measurement

Enhancing Accuracy and Reliability of Objective Assessment of Friedreich Ataxia Using Dynamic Time Warping-Based Outlier Removal

Test-retest ICC = 0.97

Cleaning sensor recordings before scoring them: repeated attempts at a task are compared and the odd ones out are dropped.

The problem

Repeated task attempts vary because of protocol slips, sensor artefacts or momentary disturbances, which adds noise to severity estimates.

The approach

Multivariate dynamic time warping compares repeated cup-task executions, a median-based threshold flags outliers, and a multi-sensor consensus removes them before an Extra Trees model predicts severity.

The outcome

Higher accuracy and test-retest reliability, with an ICC of 0.97.

In detail

Developed a signal-level quality-control framework for AIM-C measurements using Dynamic Time Warping. Repeated task executions are compared within each session, with robust median/MAD thresholds and multi-sensor consensus used to remove non-representative segments. The approach improved prediction performance and substantially strengthened test-retest reliability, reaching an ICC of 0.97 with the two-sensor consensus strategy.

Built with

• Time-Series Analysis
• Dynamic Time Warping
• Clinical Reliability

More in Time-Series & Sensors

• Probabilistic Deep Ensembles for FRDA Severity Assessment
• Cross-Device Domain Adaptation for FRDA Assessment
• AIM-M: Multiview IMU Assessment of Friedreich Ataxia

Next project: FAIR: Interpretable Latent Severity Scoring for Neurological Disorders

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