Revolutionizing Friedreich Ataxia Assessment: How AI is Enhancing the Reliability of Motor Function Measurement

Reliable FRDA Assessment via Isolation Forest Anomaly Detection

0:00

Friedreich Ataxia (FRDA) is a debilitating neurodegenerative disorder that progressively impairs motor function, significantly affecting daily activities and quality of life. For clinicians and researchers, accurate and objective assessment of these motor impairments is crucial for monitoring disease progression, evaluating therapeutic efficacy, and guiding clinical management. While traditional clinical scales like the Friedreich Ataxia Rating Scale (FARS) and its modified version (mFARS) are widely used, they can suffer from inter-rater variability, making consistent tracking challenging.

The Promise and Challenge of Wearable Tech

Recently, wearable technologies, specifically those incorporating inertial measurement units (IMUs), have emerged as powerful tools for objectively quantifying motor deficits in FRDA. One such innovation is the Ataxia Instrumented Measurement Cup (AIM-C), an IMU-based device designed to capture precise kinematic and kinetic data during a simulated drinking task. This task is particularly relevant as it assesses upper limb motor coordination and control, a key aspect of daily living activities (ADLs). Features extracted from AIM-C data, such as movement smoothness and trajectory, have shown promise in correlating with disease severity.

However, the real-world application of these devices, especially for home-based monitoring, introduces a significant challenge: data anomalies. These anomalies can arise from irregular task execution by the participant or external disturbances in the environment. When present, these distortions in the raw IMU data can degrade the accuracy of extracted features and compromise the reliability of ataxia scores, undermining the device's utility.

Our Innovative Solution: Isolation Forest and Adaptive Denoising

To address this critical limitation, a novel framework has been developed that integrates Isolation Forest-based anomaly detection with an adaptive median filtering approach. This dual-step method is designed to enhance the reliability of AIM-C data by both identifying and mitigating the impact of anomalies.

Here’s how our proposed AIM-C Processing Workflow operates (as depicted in Fig. 1 in the sources):

1. Data Collection: Participants perform a functional drinking task using the AIM-C device, recording kinematic and kinetic data.

2. Segmentation & Anomaly Detection (Isolation Forest):

   • Raw signals are segmented into individual trials.

   • Isolation Forest, a robust unsupervised anomaly detection algorithm, is applied to identify anomalous segments. This algorithm works on the principle that anomalies are "different" from the majority of data points and thus require fewer "splits" to isolate within a decision tree.

   • Dynamic Thresholding: Unlike traditional methods that require a predefined contamination rate, our framework uses a dynamic thresholding mechanism. This approach computes anomaly scores for all segments and then selects the segment with the highest anomaly score (least isolated, most normal) as a baseline. A relative deviation from this baseline is calculated for all other segments, and a dynamic threshold (mean + d * standard deviation of these relative deviations) flags and removes anomalous segments. This adaptive design ensures robust detection even with varying data characteristics.

3. Adaptive Denoising:

   • For the remaining "clean" segments, an adaptive median filter is applied. This filter dynamically adjusts its kernel size based on each segment's anomaly score.

   • Segments with low anomaly scores (indicating high reliability) receive minimal smoothing to preserve fine details, while segments with higher scores (more noise) are smoothed more aggressively with larger kernels. This process is applied independently to each sensor channel (e.g., accelerometer X, Y, Z axes).

   • This intelligent filtering effectively removes outliers while preserving essential signal characteristics, leading to improved signal quality.

4. Feature Extraction & Score Prediction:

   • Feature values (from time, frequency, and complexity domains) are computed from the denoised segments.

   • These features are then fed into an Extra Trees Regressor model, chosen for its robustness and ability to handle high-dimensional datasets, to predict the AIM-C score, which reflects motor function severity.

     
     

Impressive Results: Enhanced Reliability and Precision

The effectiveness of this framework was rigorously evaluated using AIM-C data collected from 70 participants with FRDA. The study compared the reliability of the baseline AIM-C model (without preprocessing) against the enhanced model. Key metrics included:

• Inter-class correlation (ICC): Measures consistency across repeated measurements.

• Coefficient of Variation (CV): Assesses relative variability in predictions.

• Standard Error of Measurement (SEM): Evaluates prediction precision.

• Bland-Altman plots: Provides a visual assessment of agreement between test and retest scores.

The results were compelling:

• Significant Improvement in ICC: The enhanced model achieved an ICC of 0.954, surpassing the baseline model's ICC of 0.931. An ICC greater than 0.90 is considered excellent reliability.

• Reduced Variability and Improved Precision: The enhanced model also demonstrated lower CV (5.40 vs. 5.54) and SEM (1.46 vs. 1.91) compared to the baseline model, indicating reduced variability and improved prediction precision.

• Matching Gold Standard Reliability: Crucially, the enhanced AIM-C model's reliability metrics matched or even exceeded those of the mFARS scale, which is considered the most reliable traditional clinical tool for assessing ataxia severity (mFARS ICC: 0.95, CV: 5.57, SEM: 2.36). This comparison highlights the potential of AIM-C as a robust and objective alternative.

• Visual Confirmation: Bland-Altman plots visually corroborated these findings, showing narrower limits of agreement for the enhanced model, further confirming its consistency.

Clinical Significance

This study marks a significant step forward in the objective assessment of Friedreich Ataxia. By addressing the challenges of data anomalies and sensor noise, the proposed preprocessing framework significantly improves the reliability of AIM-C device data. This enhanced reliability provides clinicians with a more consistent and accurate tool for tracking disease progression and evaluating the effects of treatments. The AIM-C device, augmented with these advanced data preprocessing strategies, stands as a promising, reliable mechanism to enhance objective assessment capabilities in both FRDA research and clinical practice.