Citation

Sungheetha, Akey and R., Rajesh Sharma and Balusamy, Balamurugan and Mapari, Shrikant and Karthik, P. and Yogarayan, Sumendra (2026) Biologically inspired neuromorphic-XAI synergy for transparent and low-carbon healthcare intelligence. Scientific Reports, 16 (1). ISSN 2045-2322

Text s41598-026-39515-2.pdf - Published Version Restricted to Repository staff only Download (6MB)

Abstract

The exponential growth of artificial intelligence in healthcare has created unprecedented computational demands, contributing significantly to carbon emissions while often lacking transparency in critical medical decisions. Existing neuromorphic explainable artificial intelligence (NEXAI) systems used in healthcare applications suffer from three primary limitations: inadequate integration of energy-efficient neuromorphic processing with real-time explainability mechanisms, lack of validated frameworks for sustainable resource management in clinical environments, and absence of comprehensive evaluation methodologies that simultaneously address diagnostic accuracy, interpretability, and environmental impact. We develop the NEXAI-Health framework by processing continuous spike streams, iteratively sampling spike rates in the range rs = 480–520 spikes/s, with cycle-to-cycle variations of ±7 spikes confirming stable neuromorphic firing behavior. Event-driven thresholds are dynamically tuned to θt = 0.42 ± 0.03, and simulation sweeps further validate threshold drift within the narrow interval [0.39, 0.45]. The integrated explainability module processes gradient-based attributions using sample magnitudes ∇Φ=0.87–0.94, internally expanding to per-layer saliency scores {0.91, 0.88, 0.93} across representative trials. Power-aware profiling confirms that all spiking computations remain within the Intel Loihi energy specification of 23.6 pJ per event, supporting sustainable deployment. Experimental iterations on 109,446 MIT-BIH heartbeat samples yield mean diagnostic accuracy of 94.7 ± 2.1% with explainability scores of 0.92 ± 0.04, and projected energy-efficiency gains converging to 67.3 ± 5.2% over conventional AI baselines. Statistical validation employs 10-fold stratified cross-validation with Bonferroni-corrected paired t-tests (α = 0.0125), demonstrating significant improvements over conventional approaches (Cohen’s d = 2.84, p < 0.001). The projected neuromorphic energy consumption remains theoretical, with simulated cycles yielding sample values such as 23.6pJ–28.2pJ per spike under a modeled firing rate of rs = 145–162 Hz. Claims regarding biodegradable substrate integration are likewise conceptual, assuming provisional material constants κm = 1.12–1.34 for tensile–thermal coupling. Clinical translation further mandates regulatory approval and structured physician training, while algorithmic correctness is supported through iterative validation on the MIT-BIH dataset (109, 446 labeled beats). Ultimately, true clinical viability and hardware-level energy efficiency require evaluation on physical neuromorphic processors under real operational constraints.This study presents a theoretical framework validated through software simulation using publicly available MIT-BIH Arrhythmia Database; no physical neuromorphic hardware implementation, clinical trials, or human participants were involved.

Item Type:	Article
Uncontrolled Keywords:	Neuromorphic computing
Subjects:	Q Science > QA Mathematics > QA71-90 Instruments and machines > QA75.5-76.95 Electronic computers. Computer science R Medicine > R Medicine (General) > R858-859.7 Computer applications to medicine. Medical informatics
Divisions:	Faculty of Information Science and Technology (FIST)
Depositing User:	Ms Rosnani Abd Wahab
Date Deposited:	04 May 2026 02:12
Last Modified:	07 May 2026 08:05
URII:	http://shdl.mmu.edu.my/id/eprint/15825

Downloads

Edit (login required)