November 21 ~ 22, 2026, London, United Kingdom
Bala Subramanyan, Verifoxx, United Kingdom
Resource bounds established for individual tools do not compose automatically across autonomous control. A tool-using agent may transform each returned observation into retry, fan-out, fallback, or delegated work;cumulative resource demand therefore depends on the controller–tool feedback closure rather than on anyinvocation in isolation. This paper introduces Execution-Demand Hypergraphs (EDHs), a compositionalabstraction in which nodes carry independently certified local bounds and observation-conditioned hyper-edges encode multisets of successor obligations. The induced monotone demand operator has a least fixedpoint equal to the exact worst-case local-bound demand represented by the abstraction, yielding a finitecertificate if and only if that demand is finite. A sound abstraction theorem transfers EDH certificates toconcrete controller executions, and an exact extractor is established for a core controller language withsequence, choice, retry, fan-out, delegation, and controlled recurrence. EDH-Check matched independentreference semantics on all 16 finite hand-designed models and 20,000 generated controllers; observationand recurrence stress tests further exercised the soundness boundary. EDH thus provides a reusable bridgefrom component-level resource certification to whole-agent resource containment.
Tool-using AI, formal verification, resource analysis, autonomous agents, WebAssembly.
Dr. Alexandru I. Jittu, PE, AIJ R&D Engineering & Consulting Inc.Michigan, USA
This paper proposes AIJ-IERP , a distributed multimodal artificial intelligence and soft computing architecture for autonomous emergency response. The architecture connects satellite communications, regional and local AI stations, fire-station intelligence nodes, autonomous firefighting vehicles, aerial robotic platforms, humanoid emergency robots, multimodal sensors, and human operators. Thermal, visual, spatial, environmental, structural, and robotic information is fused into a continuously updated Emergency Digital Twin. A hybrid intelligence layer combines machine learning, fuzzy reasoning, probabilistic inference, anomaly detection, optimization, and confidence estimation to predict hazards, coordinate multiple agents, and select among available mitigation modalities. The proposed system operates through a closed-loop Sense Fuse Understand Predict Decide Act Verify Reassess cycle with supervised autonomy and human escalation for uncertain or safety-critical conditions. The work establishes a research framework and validation pathway rather than claiming completed operational deployment.
Artificial Intelligence, Autonomous Emergency Response, Multimodal Sensor Fusion, Soft Computing, Human AI Robot Collaboration.
Aniruddha Mukherjee, Vernon Rego1and Janche Sang2, 1Department of Computer Sciences, Purdue University, West Lafayette, IN, USA 2Department of Computer Science, Cleveland State University, Cleveland, OH,USA
We introduce the Hatter audit, a stochastic audit of causal exposure in parallel discrete-event simulation. A Hatter cell marks A → H → B, where a cross-process update must be incorporated before a later event commits. Using common-input ledgers, we compare unsafe, ordered-reference, conservative, and optimistic executions against the same sequential committed history. The experiments show that exposure probability is only the beginning: the same causal event may appear as wrong output, waiting, safety checks, rollback, replay, cancellation, stored history, or finite-account breach. Causal cost therefore cannot be inferred from exposure probability alone. It must be followed through the policy clocks and, ultimately, through the weighted global critical path.
Parallel discrete-event simulation, synchronization, causality, rollback, global critical path.
Siang-Ting Huang, Cheng-I Chen and Wei-Chung Lee, Singular Wings Medical Co. Ltd., Hsinchu, Taiwan (R.O.C.)
Continuous, non-invasive monitoring of blood glucose from physiological signals is an attractive alternative to invasive fingerstick testing or continuous glucose monitor sensors. This paper presents a deep learning architecture that classifies short, single-lead electrocardiogram (ECG) segments into blood glucose level categories, requiring no invasive measurement at inference time. Both a three-class (Low, Normal, High) and a two-class (Normal, High) formulation are supported. The proposed model combines a three-stage convolutional front end with a Transformer encoder: the convolutional stages compress the input waveform into a short sequence of feature vectors, and the encoder models temporal dependencies across that sequence before classification. Focal loss and balanced batch sampling address the severe class imbalance of the task. Across 49 subjects the model attains an overall sensitivity of 60.00% for three classes and 70.69% for two classes, against 58.77% and 66.38% for a CNN-LSTM baseline, using 44% fewer parameters.
Electrocardiogram, Transformer, Non-Invasive Blood Glucose Monitoring, Focal Loss .
Yu-Hsuan Lin, Cheng-I Chen and Wei-Chung Lee, Singular Wings Medical Co. Ltd., Hsinchu, Taiwan (R.O.C.)
Wearable electrocardiogram (ECG) devices enable continuous cardiac monitoring in daily life, but the recorded signals are frequently corrupted by poor electrode-skin contact, body motion, electrical interference and hardware constraints. Such artifacts can cause false R-peak detection and misleading clinical interpretation, so an efficient signal quality assessment step is required before further analysis. This paper proposes two lightweight filters for wearable ECG quality screening. The first, an out-of-range artifact filter, uses the Gini coefficient of the squared signal amplitudes to detect impulse noise within RR intervals, transient signal attenuation and signal loss; because the Gini coefficient is scale-invariant, a single threshold can be applied across subjects. The second, a T-wave distortion filter, maps the signal to its second derivative, combines the Teager-Kaiser energy operator with a slope zero-crossing rate to emphasize distorted segments relative to QRS complexes, and applies a three-stage decision rule within a heart-rate-adaptive sliding window. Both methods were evaluated on 10-second ECG segments recorded at 250 Hz in a clinical trial at Landseed International Hospital. With a Gini threshold of 0.23, the out-of-range artifact filter achieved 96.97% sensitivity and 100% specificity for one of two subjects and achieved 93.13% sensitivity and 98.4% specificity for the other subject. The T-wave distortion filter achieved 100% sensitivity and specificity for two of three subjects, and 99.32% sensitivity and 95.20% specificity for the third. As both filters rely only on basic algebraic and difference operations, they are well suited to real-time deployment on edge computing platforms and low-power wearable devices.
ECG Signal, Impulse Noise, T-Wave Distortion, Signal Filter
Chen Chang, Cheng-I Chen and Wei-Chung Lee, Singular Wings Medical Co. Ltd., Hsinchu, Taiwan (R.O.C.)
Wearable sensing and deep learning have made it attractive to infer physiological indices from non invasive signals, and the electrocardiogram (ECG), well suited to continuous long-term recording, has been used to detect glycemic events. Inter-subject heterogeneity, however, demands personalized modeling: each model is trained on the limited data of one subject, making sample size the principal constraint. Under such conditions, preserving and reusing the information already in the signal may matter more than increased capacity, and shortcut connections are the main mechanism for this. We therefore compared two personalized architectures for ECG-based glucose level classification: a nine-layer convolutional network followed by an LSTM, and a dense-shortcut residual network built on the dense weighted normalized shortcut of DSNet, with a dual-polarity stem that pools the positive and negative parts of the feature map separately so that both peaks and troughs are retained. The two networks were evaluated on ten participants labeled by continuous glucose monitoring (CGM), on a two-class and a three-class task. The dense-shortcut network attained the higher sensitivity for eight of the ten participants but the higher specificity for only five, and neither architecture was better for every participant. Architecture selection is therefore part of personalization, and combining architectures is a promising direction.
Electrocardiogram, non-invasive glucose monitoring, personalized deep learning, dense shortcut connections
Faaiza Zulfqar, Ana Maria Martinez-Enriquez, and Kazim Ali Korejo, Center for Research and Advanced Studies of the National Polytechnic Institute, Mexico
Financial question answering requires language models to generate accurate answers from domain-specific like finance and often complex information. This study investigates whether parameterefficient fine-tuning can improve the performance of compact Large Language Models in both ideal and retrieval-based evidence settings. Six LLMs ranging from 0.5B to 1.7B parameters were evaluated on the Financial-QA-10K dataset using gold context and fixed retrieved context. The evaluation was performed before and after QLoRA fine-tuning, resulting in 24 experimental configurations. QLoRA improved F1 for every model under both context conditions, with all 95% paired-bootstrap confidence intervals remaining above zero. Under gold context, Qwen3-1.7B achieved the highest fine-tuned F1 score of 71.40%, whereas SmolLM2-1.7B obtained the highest F1 score under retrieved context at 63.98%. Overall, the findings demonstrate the effectiveness of parameter-efficient fine-tuning for financial question answering while highlighting the influence of evidence quality on retrieval-augmented generation performance.
Financial Question Answering (FQA), Large Language Models (LLMs), Retrieval-Augmented Generation (RAG), QLoRA, Parameter-Efficient Fine-Tuning
Olayinka Jimoh1Kehinde Julius Adediran2 ,Anchal Garg3, 1NHS Foundation Trust, Sherwood Sutton-in-Ashfield, United Kingdom 2,Department of Digital and Pyhsical Science, University of Hull, United Kingdom 2,Department of Computing, University of Greater Manchester, United Kingdom
Machine-learning applications in observational healthcare research depend on consistent representations of patient data, yet differences in database structure and terminology can limit reproducibility and portability. The Observational Medical Outcomes Partnership Common (OMOP) Data Model (CDM) provides a standardised representation for observational healthcare data, but a standardised data model alone does not define an end-to-end machine-learning and research-software workflow. This study presents a proof-of-concept OMOP-to-machine-learning framework for patient-level gastrointestinal (GI) bleeding risk stratification. The framework demonstrates how OMOP CDM-formatted data can be transformed into patient-level features, used within a reproducible machine-learning development workflow, and operationalised as a research-software demonstrator. The current implementation uses the synthetic GiBleed dataset distributed through the OHDSI Eunomia ecosystem as a controlled development environment. Multiple classification approaches were evaluated to demonstrate the model-development and selection stages of the framework, with the resulting model subsequently operationalised in a research-software prototype. The significance of this work lies in demonstrating a structured pathway from standardised observational healthcare data to machine-learning development and operational demonstration, providing a reproducible foundation for future research using real-world data. The framework is intended primarily to benefit healthcare researchers, data scientists, and organisations working with observational health data by providing a structured approach that can support reproducible model development and facilitate the transition from data standardisation to machine-learning applications. In the longer term, a validated implementation could support healthcare organisations in integrating patient-level risk information into existing electronic health record (EHR) workflows, potentially making relevant risk information available for clinicians.Future work will focus on validation using real-world longitudinal OMOP data, followed by evaluation of clinical performance, transportability, and potential integration into existing hospital EHR systems.
OMOP Common Data Model, machine learning, patient-level prediction, gastrointestinal bleeding, EHR.