XAI explanations for heart rate estimation models lack direct performance correlation

Researchers have investigated the reliability and cross-dataset transferability of explainable AI (XAI) methods when applied to RhythmFormer, a model used for remote photoplethysmography (rPPG) which estimates cardiovascular pulse from facial videos. The study assessed various XAI techniques, including attention maps and saliency-guided faithfulness coefficients, across different datasets like NCKU-rPPG and UBFC-rPPG. Findings indicate that while some XAI methods, particularly 'Beyond Intuition', showed good skin coverage and faithfulness, these metrics did not consistently correlate with the model's actual performance in estimating heart rate. The research suggests that attribution to skin regions does not guarantee accurate rPPG estimates, and XAI explanations primarily reveal where a model focuses rather than how faithfully it represents the underlying physiological signal. AI

IMPACT Explains that current XAI methods may not reliably indicate the performance of physiological signal estimation models, suggesting a need for more robust evaluation techniques.

RANK_REASON Academic paper detailing methodology and results of an AI research study. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• Beyond intuition and instinct blindness: toward an evolutionarily rigorous cognitive science
• NCKU-rPPG
• RhythmFormer
• SaCo
• Salience-guided Faithfulness Coefficient
• Torbjörn Nordling
• UBFC-rPPG

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →