Explainable AI for Time Series via Virtual Inspection Layers

Time series AI explanations lag behind other fields: While explainable AI has advanced significantly in image and text analysis, time series data—used in applications like medical monitoring and audio analysis—has received limited attention, creating a gap in understanding how AI makes decisions about sequential data.

A new method makes time series interpretable: The paper introduces "virtual inspection layers" that convert raw time series into understandable representations (specifically using Fourier transformation), allowing researchers to trace which parts of the data influenced an AI model's decision, similar to existing explanation methods for images.

The approach reveals hidden model behaviors: DFT-LRP helped researchers discover important insights, such as how models trained on different types of data (time-based vs. frequency-based) use different strategies to make predictions, and crucially, how models can exploit spurious correlations—patterns that appear meaningful but are actually coincidental—rather than learning genuine relationships.

Practical applications span healthcare and audio: The method was tested on real-world problems including electronic health records and audio classification, demonstrating that the approach works across different time series domains where explainability is critical for trust and safety.

Explainability enables responsible AI deployment: By making time series model decisions transparent and interpretable, this work supports more responsible AI use in sensitive domains like healthcare, where understanding why a model made a diagnosis or prediction is essential for clinical decision-making.