Improving Medical Trust with Time Series AI Explainability — AITRICS Paper Selected in Top 3% for ICML Spotlight

AITRICS (CEO Kwang joon Kim), a company specializing in artificial intelligence(AI) technology announced on  May 30 that its research paper, which improves AI explainability methods for time series data, has been officially accepted to the International Conference on Machine Learning (ICML) 2025—one of the world’s most prestigious machine learning conferences.

The accepted paper presents a novel explainability technique tailored for time series data, titled “TIMING: Temporality-Aware Integrated Gradients for Time Series Explanation.” This method contributes to clarifying the basis of predictions made by time series models and enhances transparency in the model’s decision-making process.

Traditional explainability methods for time series data primarily assess the magnitude of each time point’s contribution, without accounting for whether the contribution is positive or negative. This structural limitation makes it difficult to distinguish between opposing influences.

To overcome this issue, the AITRICS research team proposed an improved approach that incorporates directional attribution. They also introduced two new quantitative evaluation metrics: Cumulative Prediction Difference (CPD) and Cumulative Prediction Preservation (CPP).

Using these metrics, the team found that the classic Integrated Gradients (IG) method outperformed more recent time series explanation techniques. Building on IG’s strengths, they developed a new method optimized for time series data—TIMING. TIMING uses a segment-based random masking strategy to disrupt temporal dependencies and more accurately measure each time point’s contribution. Experiments across various time series datasets, including medical data, demonstrated that TIMING effectively identifies key moments critical to prediction and outperforms existing explainable AI (XAI) methods.

Changhoon Kim, a researcher at AITRICS, commented, “Time series data plays a vital role in domains like healthcare, where precision and safety are essential. This study is significant in that it enables evidence-based explanations for AI predictions and ensures transparency, which can meaningfully support clinical decision-making by medical professionals.”

ICML, first held in 1980, is the world’s top-tier machine learning conference, renowned for showcasing the latest advancements in AI research and technology. At its 42nd edition in 2025, AITRICS’ paper was selected among the top 3% of all submissions and recognized for a Spotlight presentation—an honor that highlights its technical excellence and research innovation.