Artificial Intelligence and Machine Learning

Multimodal explainability for ICU signals: metric and asymptotic results

Yu. V. Trofimov^ab, A. N. Averkin^cb, E. M. Kuznetsov^b, A. P. Eremeev^d, A. V. Nechaevskiy<sup>ab

a</sup> Meshcheryakov Laboratory of Information Technology, JINR, Dubna
^b Dubna State University, Dubna, Moscow Reg.
^c FRC Computer Science and Control RAS, Moscow
^d NRU MPEI, Moscow

Abstract: The paper presents the first mathematically rigorous multimodal explainability system for three-channel physiological signals (Electrocardiogram (ECG), Photoplethysmogram (PPG), Arterial Blood Pressure (ABP)) in distinguishing true from false ventricular tachycardia (VT) alarms in intensive care units (ICUs). A novel explanation consistency metric, Coherence, based on temporal attributions from Integrated Gradients between modalities, is introduced with theoretical justification of its connection to local surrogate stability. The developed ResNetFusionClassifier architecture with an adaptive attention mechanism provides specialized processing for each modality followed by intelligent feature fusion. Experimental validation on the extended VTaC dataset (1,247 episodes from 982 patients) [clifford2016false] demonstrated Accuracy 0.873, F1-score 0.873, AUC-ROC 0.926, with a statistically significant difference in the Coherence metric between true and false alarms ($p < 0.001$). Practical application of the detection system demonstrated high recall for critical cases (Recall = 0.878) alongside a significant reduction in false alarms, confirming the clinical applicability of the developed approach for addressing the problem of "alarm fatigue" in ICUs.

Keywords: multimodal explainability, ventricular tachycardia, mutual information, Shapley values, physiological signals, explanatory artificial intelligence.

UDC: 004.032.26, 519.6, 51-76

Received: 31.10.2025
Revised: 12.11.2025
Accepted: 08.12.2025

DOI: 10.26456/vtpmk757

Bibliographic databases:

•