ABH 2026 Abstracts


Area 1 - AI Ethics by Design

Short Papers
Paper Nr: 17
Title:

AI-Supported Development of a Picture-Based Questionnaire for Language-Independent Assessment of Quality of Life: A Feasibility Study

Authors:

Isabel Schubert, Jan P. Ehlers, Sebastian Unger and Thomas Ostermann

Abstract: Assessment and clinical diagnostics are essential pillars of psychology and psychotherapy to ensure the effectiveness of psychological interventions. Development of diagnostic tests is mainly based on standardized language-based assessments. Nonverbal diagnostics represent a complementary and promising approach in cases where communication barriers are present. This work describes the development of image-based, AI-supported questionnaire for quality of life. For this reason, the linguistic items of the WHO-5 questionnaire were converted into image-based versions using DALL-E and CHAT-GPT which was named WB-PiQ. In a pilot study with 50 participants the comprehensibility and clarity of the items of both questionnaires was checked. In the main study with another 100 participants item differences, correlations and agreement rates between the questionnaires were evaluated. The results show that the items of the WB-PiQ were highly comprehensible, with an average accuracy rate of 94.2% compared to the WHO-5. Analysis of internal consistency showed acceptable and nearly identical reliability of 0.77 and 0.78 for both test versions. Correlation analysis of the single items revealed moderate to high correlations between .385 and .672. The exact agreement between the items was notable lower, with Cohen's kappa values between 0.10 and 0.26 and between 0.43 and 0.68 for quadratic weighted kappa. In conclusion, our study delivered mixed results with respect to the feasibility of an AI-supported development of a picture-based questionnaire for language-independent assessment of quality of life. Despite the positive results, a joint effort between AI and human expertise might enhance the efficacy of such an approach.
Download

Paper Nr: 41
Title:

Privacy-Preserving Diabetes Risk Prediction with Fully Homomorphic Encryption

Authors:

Abdulelah Alasmari, Shafiq Bharwani, Louis Hwang and Giovanni Di Crescenzo

Abstract: Machine-learning models are increasingly used to support medical risk prediction, but deploying such models on remote infrastructure raises privacy concerns when patient data are stored with an untrusted server. We consider the problem of enabling a server-hosted classifier to operate directly on encrypted patient data, ensuring that the server never learns the client’s clinical features while still returning an accurate prediction. We present the design and evaluation of an end-to-end system for privacy-preserving diabetes risk prediction based on a decision-tree classifier compiled into a TFHE-based fully homomorphic encryption (FHE) circuit using Concrete-ML. In our architecture, the client locally quantizes and encrypts clinical features, the server evaluates the classifier entirely over ciphertexts, and only the client decrypts the resulting prediction. Using the Pima Indians Diabetes dataset, we conduct an extensive experimental study characterizing the tradeoffs among predictive accuracy, quantization precision, tree depth, compilation cost, and encrypted inference latency. We further investigate the impact of data quality, batching, and the performance gap between plaintext and encrypted execution. The results show that predictive accuracy remains largely preserved under quantization and encrypted inference, while the dominant costs are due to the interaction between tree depth and bit-width. In particular, data quality has a larger influence on predictive performance than quantization, whereas batching provides only limited reductions in per-sample latency. Finally, we discuss how blockchain-based governance and accountability mechanisms can complement encrypted inference and support realistic deployment scenarios for privacy-preserving medical machine learning. Overall, our results demonstrate that TFHE-based diabetes risk prediction provides strong privacy guarantees with no significant accuracy compromise over un-encrypted decision trees.
Download

Area 2 - Blockchain Engineering in Healthcare

Full Papers
Paper Nr: 14
Title:

Harnessing Visual Truthfulness: The Hybrid Authenticity Protocol for Ethical GenAI in Supplement E-Commerce

Authors:

Aleksandar-Jon Stefan

Abstract: The online market for dietary supplements focuses predominantly on visual cues, as purchasers cannot see the products themselves. Traditional packshots accurately convey data points but lack context, which may decrease engagement and fluency in processing those facts. Unlike in conventional sectors, generative AI (GenAI) enables the rapid creation of contextual product images. Still, it also poses significant downsides, including visual exaggeration, implied medical claims, and deceptive content in health-adjacent sectors. In this paper, we propose the Hybrid Authenticity Protocol - an engineering and governance framework - that restricts generative methods to the image background while preserving the pixel-level integrity of the original packaging. Using a multimodal paradigm, this protocol combines: (1) multimodal semantic extraction for prompt safety, (2) strict object preservation through bitwise alpha masking, (3) controlled background creation with diffusion-based inpainting, and (4) cryptographic audit trails through hashing, timestamping, versioning, and provenance records. A case study of BiVits ACTIVA brand is used to illustrate its implementation strategies, failure modes, and mitigation techniques. The contribution is a replicable, compliance-focused governance workflow rather than a novel underlying AI architecture for enterprises that roll out GenAI visuals in regulated e-commerce environments in the e-pharmacy field.
Download

Paper Nr: 31
Title:

Blockchain-Enabled Trust for Adaptive Artificial Intelligence in Healthcare

Authors:

Mugdha Srivastava, Abhishek Kaushik, Róisín Loughran and Fergal Mc Caffery

Abstract: The deployment of adaptive artificial intelligence (AI) in healthcare, where models continue to learn or retrain following initial deployment, introduces governance challenges that are foundationally distinct from those of static AI. In regulated clinical environments, these challenges manifest as four interrelated failure modes: concept drift and silent performance degradation, update poisoning and data integrity compromise, governance circumvention through ungoverned incremental modifications, and the collapse of model versioning and post-hoc auditability. This paper presents a blockchain-based trust architecture for adaptive healthcare AI comprising a four-layer system (user interaction, adaptive learning, blockchain governance, and IPFS storage), a version-controlled on-chain model registry, smart contract enforcement of update governance, and an explicit safety gate mechanism that activates a predetermined change control plan at the infrastructure level. The architecture extends a completed blockchain trust framework for static AI models and is motivated by a public cybersecurity survey in which blockchain was identified as the most frequently selected technology for improving healthcare system security following a data breach. The system is evaluated using the Architecture Trade-off Analysis Method (ATAM), with particular reference to the quality attributes of auditability, safety, scalability, and regulatory compliance.
Download

Short Papers
Paper Nr: 16
Title:

Antimicrobial Resistance Surveillance in England: An Explainable ML and Blockchain Framework for Outbreak Prediction

Authors:

Sanjitha Ganesh and Muthu Ramachandran

Abstract: Antimicrobial resistance (AMR) is among the most critical global public-health threats of the 21st century. This paper presents a structured descriptive analysis of 18 nationally reported AMR and antibiotic-stewardship indicators sourced from the NHS England FINGERTIPS platform, spanning hospital prescribing, community prescribing, infection rates, and clinical-governance compliance. Three supervised machine-learning (ML) classifiers — Logistic Regression, Random Forest, and XGBoost — were implemented and validated on 120 NHS Trust-level observations simulated from the real national distributions, with SMOTE applied to the training partition only to address class imbalance. Under stratified 5-fold cross-validation, Random Forest achieved the strongest performance (ROC-AUC = 0.979, F1 = 0.933, six false negatives), outperforming XGBoost (AUC = 0.965) and Logistic Regression (AUC = 0.882). SHAP analysis identified intravenous-to-oral switch non-compliance and Watch/Reserve prescribing as the strongest risk predictors, with UTI NICE compliance acting protectively. A novel Blockchain-Enabled AMR Surveillance Architecture (BEASA) is proposed, integrating Hyperledger Fabric for tamper-proof indicator storage, smart contracts for automated stewardship alerts, and HL7 FHIR-compliant NHS dashboard access. In Q1 2025/26, total hospital antibiotic prescribing reached 4,711 DDDs per 1,000 admissions, with Watch and Reserve categories constituting 46% — a 17-percentage-point gap from the WHO AWaRe 70% Access target. The combined framework provides transparent, accountable, and actionable AMR risk intelligence for NHS infection-prevention and control teams.
Download

Paper Nr: 25
Title:

Bittensor as Infrastructure for a Decentralised Healthcare AI Marketplace: Promise, Pitfalls, and Policy

Authors:

Robert Kellar and Richard Marcotte

Abstract: This position paper argues that Bittensor - an open public blockchain that coordinates a global network of AI specialists through token-based incentive mechanisms - represents a viable foundational layer for a decentralised healthcare AI marketplace. Drawing on an analysis of SafeScan, a former Bittensor subnet focused on skin-cancer detection, we argue that incentivised, open-source competition can produce clinically competitive diagnostic models without centralised infrastructure, venture capital, or proprietary data silos. SafeScan published preliminary data suggesting that it achieved a 2.7 percentage point (pp) improvement in classification accuracy and a 10.5 pp improvement in F-beta score over the ADAE state-of-the-art melanoma detection baseline, providing early empirical validation of the platform’s core thesis. We identify three structural tensions that must be resolved before Bittensor-based systems can reach clinical deployment at scale: (1) the risk that competition-optimised models overfit to internal metrics and fail to generalise in real-world clinical settings; (2) the data governance challenge of maintaining training provenance and demographic bias controls within a permissionless miner ecosystem; and (3) the misalignment between Bittensor’s short-term incentive structure and healthcare’s extended validation and regulatory requirements and timelines. We propose potential mitigations for these challenges and present policy-level recommendations for researchers, healthcare organisations, regulators, and subnet developers seeking to responsibly harness decentralised AI in clinical settings.
Download

Paper Nr: 29
Title:

A Multiagent AI Framework for Parallelised Prescription Decision Support in Primary Care

Authors:

Naveen Rajagopal, Abhijay Jagini and Muthu Ramachandran

Abstract: Primary care prescribing in the UK is conducted largely without access to a multidisciplinary team (MDT), exposing patients to unreviewed polypharmacy risk at scale. This paper presents a multiagent AI framework designed to simulate the parallel specialist review that an MDT provides but for every prescription request. The proposed architecture deploys five concurrent specialist agents (Pharmacist, Physiotherapist/ Occupational Therapist, Allergy Safety, Guidance, and Literature), each operating in a task-scoped context window to mitigate the accuracy degradation known to affect single-model approaches under high task volume. A compliance and timing optimisation layer identifies adherence-enhancing simplifications, and a structured safety double-check precedes GP sign-off. Agent slots expose standardised input/output contracts, allowing individual models to be retrained or upgraded independently. Evidence from multiagent systems research indicates that orchestrated multi-agent architectures sustain higher accuracy under clinical-scale workloads (90.6% versus 73.1% at low task volumes; 65.3% versus 16.6% at high volumes) than single-agent equivalents, which collapse under load. The paper argues that doubly parallelised multiagent AI - across agents within a case and across concurrent patient requests - is not merely an efficiency gain but a patient safety imperative given current UK primary care workforce constraint. Beyond drug-interaction checking, the agents apply STOPP/START criteria in older adults, screen for cross-reactivity and delayed hypersensitivity risk, and surface live drug-shortage and formulary information at the point of prescribing. The framework is designed for deployment within existing NHS infrastructure, interfacing directly with GP clinical systems, the Electronic Prescription Service, and NICE formulary data, minimising implementation and data-governance risk. We position the architecture as a credible and tractable candidate intervention whose value must be confirmed through prospective evaluation against clinician decisions.
Download

Paper Nr: 35
Title:

Ethical AI-Based Health Systems: Analysing Privacy Preserving and Post-Quantum Cryptography Standards for IoT-Based Blockchain Applications

Authors:

Kyriaki Tsantikidou and Nicolas Sklavos

Abstract: Smart healthcare systems integrate the Internet of Things (IoT) and blockchain technologies, cloud infrastructures and advanced data analytic models to enable continuous monitoring, predictive diagnosis and intelligent decision making. However, the distributed nature of these systems, together with blockchain-based infrastructures and third-party processing, introduces significant privacy and security challenges. In parallel, the emergence of quantum computing threatens the long-term security of conventional cryptographic mechanisms. This paper examines the security vulnerabilities and threats associated with AI-based healthcare systems. It analyses privacy-preserving techniques for protecting data during processing as well as post-quantum cryptographic algorithms for secure communication and authentication. Finally, key challenges and future directions are discussed, highlighting the potential of combining homomorphic encryption and post-quantum cryptography for developing secure, resilient and hardware-efficient healthcare systems.
Download

Paper Nr: 43
Title:

Domain-Engineered Foundation Models for AI–Blockchain Healthcare Integration: A Design-Science Framework with a Governance Prototype

Authors:

Muthu Ramachandran, Gopal Singh Jamnal and Kulvinder Panesar

Abstract: The integration of artificial-intelligence (AI) foundation models (FMs) and blockchain technology in healthcare holds transformative potential, yet current approaches lack the systematic domain grounding needed for safe, scalable, and regulatory-compliant deployment, contributing to a high rate of failed AI pilots. This paper argues that domain engineering — a structured methodology for identifying, organising, and managing reusable assets within an application domain — provides the strategic scaffold for developing proprietary healthcare FMs with native blockchain governance, and it develops that argument as a Design Science Research (DSR) study. Drawing on established domain-engineering techniques, ontological healthcare classification (HL7 FHIR, SNOMED CT), and the Generalist Medical AI (GMAI) paradigm, we present a four-layer AI–blockchain integration framework that maps healthcare subdomains to FM-type requirements and blockchain use cases, with subdomain fine-tuning realised through parameter-efficient adapters on a shared multimodal core. We then evaluate the framework along two complementary lines: an analytical capability comparison across five FM approaches, and a working proof-of-concept prototype that implements the governance layer — hash-chained provenance, consent and role-based access smart contracts, and on-chain/off-chain partitioning — and is measured under four experiments. The prototype shows that the governance layer adds under 0.02% to end-to-end latency relative to model inference, that hash-anchoring holds per-record on-chain storage constant at 128 bytes (an over 40,000× reduction for a 5 MB image), and that consent/access enforcement and tamper detection are correct in 100% of more than 5,500 tested cases. The contribution is a methodologically grounded, partially validated framework and an open prototype that move healthcare AI–blockchain integration from position to demonstrable design.
Download

Paper Nr: 45
Title:

ZTPIF-H: A Zero-Trust Patient Identity Engineering Framework for Healthcare Using AI and Blockchain

Authors:

Shailesh Kejadiwal and Muthu Ramachandran

Abstract: Healthcare ecosystems face persistent identity-fragmentation challenges that undermine secure, interoperable data access across Electronic Health Records (EHR), telemedicine platforms and clinical Application Programming Interfaces (APIs). Existing Identity and Access Management (IAM) solutions rely on static role assignments and reactive audit mechanisms ill-suited to dynamic clinical environments. This paper presents the Zero-Trust Patient Identity Engineering Framework for Healthcare (ZTPIF-H), which integrates decentralised identity, Artificial Intelligence (AI)-driven adaptive authorisation, smart-contract-based consent governance and blockchain-based immutable auditability. The framework is organised around four pillars Identity Assurance, Adaptive Trust Decisioning, Consent Governance and Immutable Accountability operationalised through an eight-phase lifecycle. Relative to prior work, this version adds a multi-dimensional comparison with state-of-the-art healthcare IAM approaches, a reproducible experimental protocol, an explicit threat model with a systematic security analysis, and an AI-trustworthiness analysis covering interpretability, calibration, robustness and fail-safe behaviour aligned with the NIST AI Risk Management Framework (AI RMF) and the EU AI Act. A proof-of-concept (PoC) evaluation in a simulated secondary-care environment shows a 42% reduction in authorisation latency, a 62% reduction in false-positive denials and a 77% reduction in standing-privilege exposure versus a conventional IAM baseline. ZTPIF-H aligns with HL7 FHIR R4, OpenID Connect (OIDC), W3C Verifiable Credentials (VC) and NIST Zero Trust Architecture (ZTA) principles, offering a vendor-neutral, incrementally adoptable pathway to modern healthcare identity engineering.
Download

Paper Nr: 28
Title:

The Immutable Hallucination: A Critical Analysis of AI-Blockchain Integration in Healthcare

Authors:

Sitalakshmi Venkatraman and John Minicz

Abstract: The convergence of Artificial Intelligence (AI) and blockchain technology is reshaping healthcare infrastructure, promising predictive intelligence and tamperproof data integrity. Yet this integration potentially introduces a fundamental paradox: the Immutable Hallucination. AI systems produce probabilistic outputs that may be incorrect, while blockchain systems enforce deterministic, irreversible data commitments. This paper critically examines the technical, clinical, legal, and regulatory risks that arise when AI generated errors, if left unchecked, become permanently embedded in distributed ledgers. We offer this paper as an opinion piece and draw on emerging case studies, architectural analyses, and regulatory developments, we identify key failure modes, evaluate mitigation strategies such as multi-signature medical oracles and reversible state layers, and identify gaps in AI-blockchain integration. We also offer thoughts for a governance framework that prioritises explainability, human oversight, and lifecycle centric regulation. While prior work has examined AI hallucinations and blockchain immutability separately, this paper offers three novel contributions: (1) a tripartite model of persistence (content, attestation, operational) that resolves the apparent contradiction between reversible corrections and immutable ledgers; (2) formalisation of the ‘certainty trap’ as a distinct clinical failure mode arising from probabilistic AI outputs triggering deterministic smart contracts; and (3) a hybrid regulatory framework that integrates FDA change control, EU post-market surveillance, and blockchain-specific mechanisms (e.g., multi-signature medical oracles, reversible state layers) for lifecycle-centric governance of AI-blockchain healthcare systems.
Download

Paper Nr: 38
Title:

An AI-Agent Framework for Accounting in Law Firms: Process Automation, Trust-Account Controls and Integrated Cybersecurity

Authors:

Rahul Nagarajan and Muthu Ramachandran

Abstract: Law firms occupy a sensitive position in the financial system: they hold high-value client monies in trust accounts, face strict professional and anti-money-laundering regulation, and concentrate confidential data that attracts cybercriminals. The average databreach cost for law firms reached USD 5.08 million in 2024, and enforcement actions for money-laundering failures are accelerating. This paper proposes a multi-agent artificial intelligence framework for the end-to-end accounting workflow of law firms, from client intake through trust-account management, billing, month-end close and regulatory reporting, with cybersecurity and anti-money-laundering controls embedded as first-class agents rather than afterthoughts. The paper contributes a combined process-and-threat model, an eight-agent architecture, a critical evaluation of seventeen academic and commercial systems, a mapping of agent controls to regulatory standards, and a working, openly specified validation: a Security-Agent application programming interface that detects fraudulent payee-bank-change requests. On a reproducible benchmark of three hundred cases the layered agent reaches an F1 score of 0.99, outperforming authentication-only and rule-only baselines, at sub-millisecond latency with faithful, auditable explanations. The architecture transfers directly to healthcare finance, where hospitals, NHS trusts and care homes share the same financial, safeguarding and data-protection duties.
Download

Area 3 - Blockchain Ethics by Design

Short Papers
Paper Nr: 23
Title:

AARD: A Risk-Aware Agentic AI Framework for Sequential Clinical Decision Support Using Large Language Models

Authors:

Gopal Singh Jamnal

Abstract: Modern healthcare systems require decision-support tools that can operate effectively in complex, uncertain, and data-rich environments. This paper proposes the Adaptive Agentic Risk-aware Decision (AARD) framework, a policy-based approach for modelling clinical decision-making as a sequential and adaptive process. The framework integrates structured clinical data, such as vital signs and laboratory measurements, with unstructured clinical text, where Large Language Models (LLMs) are used to extract contextual representations. AARD employs an agentic architecture in which decisions are generated through a learned policy and refined using a risk-aware mechanism. This allows the system to adapt actions over time based on evolving state representations while incorporating safety considerations into the decision process. A recursive learning mechanism enables continuous policy updates using observed state transitions, supporting consistent decision behaviour across sequential steps. The framework is evaluated using open-source healthcare datasets, with performance assessed through both predictive metrics and decision-oriented measures. The results indicate that integrating multimodal representations with risk-aware policy learning provides a structured approach to sequential decision support. Overall, the proposed framework offers a scalable and interpretable approach for combining agentic decision processes, LLM-based feature extraction, and risk-aware optimisation in health-care settings.
Download

Area 4 - Clinical and Practical Applications

Full Papers
Paper Nr: 21
Title:

Rough Set-Augmented Ensemble Machine Learning for OAB/BPS Differential Diagnosis: Experimental Validation with Figures

Authors:

Muthu Ramachandran and Kavitha.S.

Abstract: Accurate differentiation of Overactive Bladder (OAB) from Interstitial Cystitis/Bladder Pain Syndrome (IC/BPS) remains clinically challenging, with documented misdiagnosis rates of 38-42% in primary care settings. This paper presents and experimentally validates a Rough Set-augmented Ensemble Machine Learning (RST-EML) pipeline. RST reduct computation on a 22-feature synthetic cohort (n=300) identifies five core attributes: pain score, urgency score, nocturia score, symptom duration, and ICSI score. RST boundary-region analysis flags 38 patients (12.7%) as diagnostically ambiguous, enriched for the comorbid OAB+IC/BPS phenotype. The RST-EML system achieves 93.1% accuracy (AUC-ROC 0.965, F1 0.800) on the 262 non-boundary patients with 100% recall for pure OAB and IC/BPS. Six RST decision rules with certainty >= 0.96 provide audit-ready clinical outputs. Treatment concordance reaches 85.2% for OnabotulinumtoxinA in correctly classified OAB patients versus 62.1% for conventional anticholinergics (p=0.008). Six publication-quality figures illustrate ROC performance, SHAP-aligned feature importance, the boundary region in feature space, treatment outcomes, and the confusion matrix. These findings establish RST as a principled uncertainty-quantification and feature-selection complement to ensemble learning in urological diagnostic AI.
Download

Short Papers
Paper Nr: 13
Title:

A Comparative Deep Learning Study for Symptom-Based Overactive Bladder Prediction with SHAP Explainability

Authors:

Subha Subbiah and Muthu Ramachandran

Abstract: Background/Objectives. Overactive bladder (OAB) is a prevalent, yet frequently underdiagnosed condition characterised primarily by urinary urgency with associated increased frequency and nocturia. Early identification of individuals at risk may support timely clinical intervention and improve quality of life for patients. This study aimed to develop and evaluate explainable deep learning models for symptom-based OAB classification using nationally representative population data. Methods. Population-level health data were obtained from the National Health and Nutrition Examination Survey (NHANES). OAB classification was constructed using clinically aligned symptom definitions. Demographic variables, comorbidity indicators, and urinary symptom responses were incorporated as predictive features. Three neural network architectures, including a Deep Multi-Layer Perceptron, a Residual Neural Network, and a Wide and Deep model, were implemented and compared. Class imbalance was addressed using Synthetic Minority Over-sampling Technique (SMOTE). Model performance was evaluated using accuracy, precision, recall, F1-score, ROCAUC, and confusion matrix analysis. Model interpretability was assessed using SHapley Additive exPlanations (SHAP). Results. All architectures demonstrated high predictive performance with strong discrimination ability (ROC-AUC > 0.96). Minimal variation in performance across architectures was observed. SHAP analysis identified age and urinary urgency as the most influential predictors, while other factors demonstrated moderate correlations. Conclusions. Symptom-driven deep learning models can accurately classify OAB using structured questionnaire data while maintaining interpretability through explainable AI techniques. These findings support the feasibility of using deep learning approaches for population-level screening applications, although further validation using clinically confirmed diagnostic datasets is warranted.
Download

Paper Nr: 33
Title:

Fast-Bio: A High-Performance Multimodal Pipeline for Rapid Variant Prioritization in Clinical Settings

Authors:

Gabriella Trucco and Stelvio Cimato

Abstract: The diagnostic odyssey for patients with rare genetic disorders remains a critical challenge, as clinicians must often navigate thousands of variants to identify a single causative mutation. This paper presents Fast-Bio, an automated, multimodal Python-based pipeline designed for rapid and high-specificity variant prioritization. The system integrates genomic data with clinical phenotypic information standardized via the Human Phenotype Ontology (HPO). A key feature of Fast-Bio is its integrated scoring algorithm, which combines state-of-the-art AI-driven predictors, such as AlphaMissense, CADD, and REVEL, with a heuristic clinical weighting system. In the validation on a synthetic clinical benchmark replicating cases of Wilson’s Disease and familial Parkinson’s Disease, Fast-Bio processed an average of 32299 variants per sample in 2.61 seconds. The pipeline consistently ranked the known pathogenic variants (e.g., ATP7B, LRRK2) at the top of the clinical report, demonstrating a significant reduction in manual analysis time. By providing a lightweight and scalable decision support tool, Fast-Bio offers a robust solution for accelerating molecular diagnosis in complex Mendelian disorders.
Download

Paper Nr: 36
Title:

Artificial Intelligence in Mental Health: A Structured Review of Current Evidence across the Care Pathway

Authors:

Dr Ezhil Anand, Prof Muthu Ramachandran and Abhishek Subramaniam

Abstract: Artificial Intelligence (AI) is increasingly deployed across the mental health pathway, from screening and diagnosis through to intervention, monitoring and prognosis. This paper presents a structured review, aligned with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidance, of the current evidence base for AI in mental health. Working from four anchor reviews and a transparent, criteria-driven corpus of supporting primary and regulatory sources, we make four contributions. First, we propose a faceted taxonomy that classifies any mental-health AI system across paradigm, data modality, clinical task, autonomy/risk and evidence maturity. Second, we report an explicit search protocol with inclusion and exclusion criteria, so that the evidence base is reproducible rather than implicit. Third, we synthesise reported performance comparatively across application domains and grade the maturity of the evidence using a five-level scheme. Fourth, we propose a Responsible AI Pipeline that connects research activity to safe clinical deployment through explicit bias, validation, safety and regulatory gates. Reported strengths include accurate classification and risk prediction for common mental disorders, earlier case-finding, scalable chatbot-based self-help, and support for personalised treatment planning. However, the literature remains marked by methodological inconsistency, limited external validation, bias in training data, under-representation of people with intellectual disability and other marginalised groups, and unresolved issues around consent, explainability and regulation. We argue that AI should be framed as an augmentation of - not a replacement for - the clinical relationship, with equity, consent and explainability treated as first-order design constraints.
Download

Area 5 - Future Applications of Agent AI and Agentic AI in Healthcare

Short Papers
Paper Nr: 15
Title:

Robust Health Monitoring via AI Agents and Generalized Topological Data Analysis

Authors:

Ponraj A. P., Muthu Ramachandran and Hari Friedrich Schuth

Abstract: Conventional health monitoring systems depend on fixed numerical thresholds and frequently miss subtle non-linear physiological transitions in the presence of real-world noise. Although Topological Data Analysis (TDA) offers a structurally grounded approach to physiological signal analysis, it operates under the assumption of metric perturbations and therefore lacks robustness against structured clinical noise-such as motion artifacts and illumination variations arising in remote photoplethysmography (rPPG). This work introduces Generalized Topological Physiological Spaces (GTPS), a novel framework in which structured physiological noise is characterised as a hereditary class within a generalized topology. By coupling semi-H -open filtrations with persistent homology, the framework preserves topological invariants while suppressing non-metric perturbations. An AI-driven monitoring agent is developed to perform real-time analysis of physiological point clouds, detecting structural deviations in connectivity (H0) and cyclic patterns (H1). Experimental evaluation shows that the GTPS-refined approach improves topological stability and reduces false alarm rates by approximately 38% relative to classical TDA, while sustaining real-time performance at approximately 45 ms processing latency per frame. These results underscore the promise of topology-driven monitoring for robust, non-invasive, and continuous health assessment.
Download

Paper Nr: 30
Title:

Exploring Knowledge, Perceptions, and Preparedness for Artificial Intelligence among Health Informatics Students at the University of Hail: Qualitative Study

Authors:

Muneef Alshammari

Abstract: This qualitative study explores the knowledge, perceptions, and preparedness of health informatics students at the University of Hail regarding artificial intelligence (AI) in healthcare. Through semi-structured interviews with 15 undergraduate students, the research identifies widespread familiarity and generally positive attitudes toward AI tools such as ChatGPT and Google Gemini. Participants recognize AI as a supportive technology that enhances decision-making and precision medicine while emphasizing the continued importance of human oversight. Despite readiness to engage with AI, significant gaps exist in curriculum design, particularly the lack of practical, simulation-based training and ethical education. The study highlights challenges, including limited specialized courses, technical barriers, and ethical concerns, and presents student-driven recommendations for curriculum reform focused on experiential learning and tiered competency development. These findings contribute empirical evidence to the limited literature on AI preparedness in health informatics education and underscore the need for targeted educational reforms to equip future professionals to integrate AI responsibly and effectively in healthcare.
Download

Area 6 - NLP, Knowledge Graphs, and Multimodal AI

Full Papers
Paper Nr: 19
Title:

Methodological Feasibility of Multimodal Scene Sketch Segmentation for Detecting Demographic Differences in Children’s House-Tree-Person Drawings

Authors:

Sebastian Unger, Laura Anderle, Thomas Ostermann, David Martin and Bettina Berger

Abstract: In somatic medicine, pixel-level quantification has facilitated diagnostic accuracy, yet mental health is often assessed using verbal or textual instruments. However, as these instruments can present a barrier, especially for children, projective tests such as the House-Tree-Person Projective Drawing Test (HTP) offer an easy-to-understand and entertaining alternative. Despite this, there are hardly any digital solutions that facilitate HTP data collection and evaluation. In this cross-sectional study, the methodological feasibility of using a multimodal Scene Sketch Segmentation Model (SSSM) to bridge this gap was evaluated with 303 children’s HTP drawings. By validating the SSSM against manual ground truth, significant positive correlations for all three object categories (r ≥ .53) were found. Analyses also revealed that the model is sensitive to age and gender differences. While age differences were present in houses and trees (p ≤ .041), gender differences only occurred in houses (p = .008). In drawn persons, on the other hand, no significant demographic differences (p ≥ .923) were detected. To conclude, the SSSM approach can detect various demographic differences in HTP drawings, offering an objective and useful metric for clinical support. Nevertheless, further research is required to address potential biases and refine the approach.
Download