Time-Series & Sensors

Probabilistic Deep Ensembles for FRDA Severity Assessment

A Probabilistic Deep Ensemble Framework for the Objective Assessment of Friedreich Ataxia

A model that rates disease severity from a simple drinking task and also says how confident it is, so a clinician knows when to trust the number.

The problem

Clinical rating scales for Friedreich ataxia are subjective and vary between raters, and a single predicted score hides how certain the model is.

The approach

Two complementary time-series networks (InceptionTimePlus and RNN-FCNPlus) with custom probabilistic output heads, combined into an ensemble that produces bounded predictions and uncertainty intervals.

The outcome

Pearson r = 0.807 with clinical scores and RMSE of 11.48, with well-calibrated confidence intervals (AUUCC 9.62) from a limited dataset.

In detail

Designed a probabilistic deep ensemble for objective FRDA severity prediction from AIM-C time-series data. InceptionTimePlus and RNN-FCNPlus models are equipped with bounded probabilistic heads that output both severity estimates and predictive variance. Ensemble aggregation improves robustness, while Uncertainty Characteristic Curves are used to evaluate calibration and derive clinically interpretable confidence intervals.

Built with

Source

More in Time-Series & Sensors

Next project: Cross-Device Domain Adaptation for FRDA Assessment

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