International Journal of Neural Systems and Applications

AI-Driven Functional Safety Framework for Real-Time Biomedical Signal Processing Systems

2026-07-15T04:17:37+00:00

This work investigates a software functional safety framework for biomedical signal processing using advanced deep learning techniques to achieve reliable, accurate, and low-latency operation in real-time healthcare monitoring systems. Functional safety is a critical requirement in biomedical sensor applications, where incorrect signal interpretation or delayed responses may compromise patient safety. The proposed framework integrates safety-aware signal processing with deep learning-based classification to enhance both performance and dependability. Experiments were conducted using TensorFlow 2.15 on an Intel i9 processor with an NVIDIA RTX 4090 GPU under Ubuntu 22.04. Biomedical signals were sampled at 500 Hz and evaluated using Accuracy, Latency, and F1-score metrics. The proposed method was compared against conventional Digital Signal Processing (DSP), Convolutional Neural Network (CNN), and Long Short-Term Memory (LSTM) approaches. Experimental results demonstrate that the proposed framework achieves the highest classification accuracy, outperforming DSP (88.4%), CNN (93.2%), and LSTM (94.1%). Furthermore, the proposed approach exhibits the lowest processing latency of 6 ms, compared with 28 ms, 18 ms, and 12 ms for DSP, CNN, and LSTM methods, respectively. These improvements indicate that the framework effectively balances computational efficiency and predictive performance while satisfying real-time operational requirements. The incorporation of functional safety principles enhances system robustness by reducing the likelihood of hazardous failures and improving confidence in decision-making processes. The findings highlight the suitability of the proposed framework for deployment in safety-critical biomedical applications, including patient monitoring, wearable healthcare devices, and intelligent diagnostic systems. Overall, the study demonstrates that combining software functional safety mechanisms with deep learning can significantly improve the reliability, responsiveness, and accuracy of biomedical signal processing systems.

Neural Network Architectures: Taxonomy of Design Principles, Evolution, and Emerging Frontiers

2026-10-05T12:05:21+00:00

Neural network architectures have undergone a remarkable transformation over the past seven decades, evolving from the single-layer perceptron to billion-parameter foundation models defining the state of the art across vision, language, and multimodal reasoning. The proliferation of architectures, however, has outpaced theoretical understanding of underlying performance mechanisms, resulting in design decisions driven largely by empirical benchmarking rather than principled reasoning. This review presents a unified taxonomy of neural network architectures organized around four core design principles: locality, memory, attention, and hierarchy, supplemented by two secondary efficiency principles: parameter sharing and sparsity. The four core principles provide a stable explanatory framework for characterizing representational capabilities, whereas parameter sharing and sparsity function as cross-cutting efficiency mechanisms that regulate computational cost. Nine major architectural families are systematically examined: multilayer perceptrons, convolutional neural networks, recurrent networks and gated variants, autoencoders, generative adversarial networks, transformers, graph neural networks, diffusion models, and state space models. Analysis focuses on the mechanisms through which each architectural family instantiates the core principles, the resulting inductive biases, and associated computational and representational trade-offs. A systematic literature review synthesizes comparative evidence concerning parameter complexity, data efficiency, scalability, and representative benchmark performance. Cross-cutting analysis further examines trade-offs involving inductive biases, scaling laws, emergent behaviors, hybrid and modular architectures, computational efficiency, interpretability, and robustness. Emerging research frontiers, including state space models, neural architecture search, biologically inspired computing, neuro-symbolic integration, and generalist foundation models, are subsequently reviewed alongside open problems in theoretical understanding, architecture design, evaluation methodology, and sustainability. The analysis identifies four core design principles as persistent architectural invariants across successive eras of neural network development, while parameter sharing and sparsity operate primarily as efficiency-oriented modifiers. A principle-centered organization of neural network architectures provides a conceptual framework for understanding architectural evolution, a decision guide for architecture selection, and a research agenda supporting principle-driven rather than benchmark-driven design.

Implementation of Metastatic Cancer Detection using Supervised Machine Learning

2026-10-15T11:48:20+00:00

Cancer is one of the leading causes of death in the world, with an estimated 9.6 million deaths in 2018. Much of this mortality is due to metastasis, the spread of malignant cells from the primary tumour to other parts of the body, so the early detection of metastatic tissue is essential. Pathologists currently find metastases by examining stained tissue sections under a microscope, which is slow and can vary from one observer to another. This article presents a supervised machine learning approach for detecting metastatic tissue in histopathological images of lymph node sections. This study first reviews earlier techniques, including Artificial Neural Networks (ANN), Support Vector Machines (SVM), and Fuzzy Neural Networks (FNN), which reported accuracies between 71% and 96%. The study then implements a Convolutional Neural Network (CNN) based on the VGG16 architecture, pre-trained on ImageNet and adapted to 96 × 96 pixel patches from the Kaggle Histopathologic Cancer Detection dataset. The CNN learns its features directly from the image pixels and separates cancerous from non-cancerous patches with an accuracy of 97.35%, which is higher than that of the reviewed methods. The results suggest that CNN-based models can support pathologists in the early detection of metastatic cancer and reduce the time needed for diagnosis.

Optimal Spike Encoding Schemes for IMU-Data: A Systematic Study

2026-10-28T12:02:08+00:00

Inertial Measurement Units (IMUs) produce synchronous, real-valued acceleration and angular-rate streams, whereas Spiking Neural Networks (SNNs) process sparse, discrete events. The encoder that converts IMU samples into spikes therefore shapes information fidelity, input activity, robustness, latency, and deployment energy. This study presents a systematic scoping review and theory-driven design framework for selecting spike encoders for IMU data. The review covers rate coding, Time-To-First-Spike (TTFS) coding, binary and phase coding, temporal-contrast and delta-modulation methods, population coding, burst coding, and learned spiking front ends. Because the available studies differ in datasets, preprocessing, network design, training, validation, and energy-measurement boundaries, the evidence was synthesized narratively rather than pooled statistically. The evidence does not support a universally optimal encoder. In the most directly controlled IMU comparison, distribution-aware rate coding produced the highest classification accuracy, TTFS produced the lowest input firing rate, but the poorest robustness, 10-bit binary coding produced the highest reconstruction signal-to-noise ratio and lowest reported dynamic energy, and multi-threshold delta modulation produced the smallest accuracy loss under injected spike errors. A Pareto-based objective, a decision flow, and a pre-registrable validation protocol are proposed. For dynamic embedded IMU applications, adaptive multi-threshold delta modulation is proposed as a reasoned default starting point—not as an empirically proven universal optimum—with a sparse absolute-value anchor added when static orientation or gravity information is required.

CHAINWITNESS: AI-Powered Chain-of-Custody and Investigative Intelligence Platform

2026-10-28T12:22:00+00:00

Indian law enforcement agencies have been employing several disparate systems for case and evidence management: CCTNS for registering FIRs and case files, mobile apps like eSakshya for recording evidence, and manual mechanisms for managing the case flow and evidence integrity. Even though the ICJS project has taken an effort in attempting to integrate these systems at a national level of data exchange, there remains a lack of a unified intelligent working system for case registration and evidence handling by police stations and investigators. The proposed solution combines all three components, FIR/case management, evidence storage via cryptographic hashing, and investigation support through AI tools, into one system. The system cryptographically hashes evidence at upload, and any further access to that evidence becomes part of the immutable chain of custody. A similarity engine powered by natural language processing (NLP) searches for similarities between newly filed FIRs and previous cases. Geospatial analysis is used to locate and determine crime trends in different regions. The priority of cases can be assigned using a case prioritization module considering the gravity of the case, evidence, and age of the investigation to assist, rather than supersede, human judgment of higher authorities. The novelty of the platform lies in binding hash-verified evidence, a tamper-evident chain of custody, and assistive analytics (case similarity, geospatial analysis, and prioritization) to the same case record, which existing national platforms and stand-alone evidence tools treat separately. The prototype was evaluated on 500 synthetic cases: related-case retrieval reached Precision@3 = 0.51 and Recall@3 = 0.84 on the 29 test queries that have a planted relative, and the prioritization ranking was stable under ±10% weight changes (Kendall’s τ-b ≥ 0.97). The system is developed as an interoperable module that could work with current governmental architecture (like ICJS-based APIs), but not as a replacement for it, and tested using artificial data in order to prevent the use of a classified police database. This prototype shows the interrelation between full-stack development, cybersecurity (tamper-proof and access control), NLP, and data analysis in a single practical system.