Transparent Quality Optimization for Machine Learning-Based Regression in Neurology.

Area of Science:

• Neurology
• Computer Science
• Biomedical Engineering

Background:

• Clinical monitoring of walking generates vast datasets valuable for analysis.
• Machine learning (ML) is increasingly applied to analyze large, heterogeneous datasets in research.
• Software quality is crucial for data-driven ML applications in healthcare.

Purpose of the Study:

• To demonstrate the feasibility of using state-of-the-art ML to predict the 2-minute walk test (2MWT) distance.
• To develop a transparent and explainable ML approach optimizing software requirements for generic use.
• To benchmark a fractional-factorial design with standardized quality metrics for an optimized software prototype.

Main Methods:

• Applied a transparent, lean ML approach using a fractional-factorial design.
• Utilized 400 training and 100 validation data points for prediction.
• Evaluated software qualities alongside prediction accuracy using standardized metrics.

Main Results:

• Achieved a 6.1% relative error in distance prediction across optimized configurations.
• The Adadelta algorithm demonstrated superior performance, with 90% of predictions having <15 m absolute error.
• Factors like age, gender, or walking aids did not significantly affect relative error, but high walking impairment in MS patients showed a significant difference (24.0%).

Conclusions:

• It is feasible to develop a transparent ML prototype for medical use cases while ensuring software quality.
• The developed ML approach provides accurate predictions for the 2MWT, aiding in the assessment of walking endurance.
• The study highlights the potential of ML in clinical monitoring and neurological measurements.