Photoplethysmography (PPG)-based arrhythmia classification enables non-invasive and continuous cardiovascular monitoring, but its clinical adoption is limited by the lack of trustworthy and transparent AI models. In this work, we propose a novel framework that integrates uncertainty quantification (UQ) and explainable artificial intelligence (XAI) to enhance the reliability of arrhythmia detection systems from PPG signals. Beyond prediction confidence alone, the proposed approach explicitly evaluates the stability of model explanations to reduce overconfident yet potentially misleading decisions. We introduce two uncertainty-aware pipelines. The first leverages predictive uncertainty (PU), estimated via Monte Carlo Dropout with a self-calibration strategy, to identify unreliable predictions. The second pipeline extends this approach by incorporating a novel localization uncertainty (LU) metric, which quantifies the consistency of feature attribution maps across stochastic forward passes, thereby assessing the reliability of model explanations. Both pipelines employ data-driven thresholds calibrated on validation data to selectively identify predictions associated with elevated uncertain or unstable explanations. Experimental results show that the PU-based pipeline achieves up to 98 % accuracy and F1 score on high-confidence predictions, while the combined PU&LU strategy attains 95 % accuracy and F1 score while retaining a larger portion of analyzable data. Notably, integrating explanation stability improves discrimination between visually similar arrhythmias and reduces spurious model behavior. By jointly addressing prediction uncertainty and explanation reliability, the proposed framework supports selective and risk-aware AI-assisted decision-making for wearable and clinical cardiovascular monitoring applications.

Uncertainty-aware and explainability-driven framework for reliable cardiac arrhythmia classification from photoplethysmography signals / Seoni, S., Molinari, F., Lapia, F., Librale, F., Acharya, U.R., Salvi, M.. - In: BIOMEDICAL SIGNAL PROCESSING AND CONTROL. - ISSN 1746-8094. - 131:(2026). [10.1016/j.bspc.2026.111796]

Uncertainty-aware and explainability-driven framework for reliable cardiac arrhythmia classification from photoplethysmography signals

Seoni, Silvia;Molinari, Filippo;Salvi, Massimo
2026

Abstract

Photoplethysmography (PPG)-based arrhythmia classification enables non-invasive and continuous cardiovascular monitoring, but its clinical adoption is limited by the lack of trustworthy and transparent AI models. In this work, we propose a novel framework that integrates uncertainty quantification (UQ) and explainable artificial intelligence (XAI) to enhance the reliability of arrhythmia detection systems from PPG signals. Beyond prediction confidence alone, the proposed approach explicitly evaluates the stability of model explanations to reduce overconfident yet potentially misleading decisions. We introduce two uncertainty-aware pipelines. The first leverages predictive uncertainty (PU), estimated via Monte Carlo Dropout with a self-calibration strategy, to identify unreliable predictions. The second pipeline extends this approach by incorporating a novel localization uncertainty (LU) metric, which quantifies the consistency of feature attribution maps across stochastic forward passes, thereby assessing the reliability of model explanations. Both pipelines employ data-driven thresholds calibrated on validation data to selectively identify predictions associated with elevated uncertain or unstable explanations. Experimental results show that the PU-based pipeline achieves up to 98 % accuracy and F1 score on high-confidence predictions, while the combined PU&LU strategy attains 95 % accuracy and F1 score while retaining a larger portion of analyzable data. Notably, integrating explanation stability improves discrimination between visually similar arrhythmias and reduces spurious model behavior. By jointly addressing prediction uncertainty and explanation reliability, the proposed framework supports selective and risk-aware AI-assisted decision-making for wearable and clinical cardiovascular monitoring applications.
File in questo prodotto:
File Dimensione Formato  
1-s2.0-S1746809426023542-main.pdf

accesso riservato

Tipologia: 2a Post-print versione editoriale / Version of Record
Licenza: Non Pubblico - Accesso privato/ristretto
Dimensione 5.06 MB
Formato Adobe PDF
5.06 MB Adobe PDF   Visualizza/Apri   Richiedi una copia
Manuscript R2.docx

embargo fino al 19/11/2027

Tipologia: 2. Post-print / Author's Accepted Manuscript
Licenza: Creative commons
Dimensione 1.57 MB
Formato Microsoft Word XML
1.57 MB Microsoft Word XML   Visualizza/Apri   Richiedi una copia
Pubblicazioni consigliate

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11583/3018133