Frontotemporal Dementia (FTD) presents significant challenges in disease progression forecasting due to its complex temporal dynamics and the logistical difficulties of acquiring longitudinal imaging data. Convolutional Neural Networks (CNNs) excel in image-based tasks but struggle with tabular data lacking spatial relationships. To address these challenges, we propose a Tensorized Image Generator (TIG) algorithm that converts longitudinal tabular data into structured image representations. TIG systematically maps features to specific pixel positions while preserving the spatial proximity of related features, enabling the use of CNNs for effective spatial feature extraction. We utilize longitudinal neuropsychiatric data from the FTLDNI archive and develop a hybrid deep learning framework integrating CNNs with Recurrent Neural Networks (RNNs), including Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRU), to model temporal dynamics. This approach forecasts FTD progression up to four years into the future. Benchmarking results reveal that CNN-LSTM and CNN-GRU models trained on TIG-generated images achieve significantly lower mean squared and mean absolute errors compared to models trained on original tabular data and images from existing conversion methods (e.g., IGTD). Additionally, TIG-generated images outperform spatio-temporal ConvLSTM2D models by preserving feature neighborhood structure more effectively, offering a novel and efficient solution for FTD progression forecasting.
Legitimacy is a key concern when introducing data about work in healthcare, especially as data-driven systems move closer to real-world application and begin shaping how care work is represented, interpreted, and shared across stakeholders. This paper reports healthcare workers’ (HCWs’) perspectives on data sharing in a design project that uses workflow data to support care, including accounting for the needs of patients and caregivers. Guided by the principle of “minimum viable datafication” [62]- limiting data production and sharing to what is meaningful and contextually appropriate - we engage Dourish’s [26] notion of the “legitimacy trap,” in which participatory processes appear consultative, but ultimately reinforce designer or managerial assumptions. Aiming to pro-actively achieve legitimacy, rather than falsely assume it, we contribute three lessons for the Human-Computer Interaction (HCI) and Health community: 1) data representations should attune to workers’ lived experiences; 2) HCWs should be enabled to care both for their own data-related needs and the impact of data sharing on patients and caregivers; and 3) space should be created for contestation and negotiation, recognizing that data sharing must remain responsive to how stakeholders experience the representation of work.
Accurate fracture detection in medical imaging is pivotal for intelligent orthopedic diagnostic systems, yet deploying high-capacity detection models on resource-constrained platforms remains a critical cyber-physical challenge. Conventional feature distillation based on Mean Squared Error (MSE) performs point-wise feature regression but does not explicitly model higher-order statistical relationships, which may lead to over-smoothed responses and weaken the preservation of subtle fracture structures. To overcome this systemic limitation, we propose Structured Feature Decoupling Knowledge Distillation (SFD-KD). This framework decomposes teacher features into linearly combined first-order semantics and multi-order statistics, bypassing computationally expensive covariance modeling. Concretely, SFD-KD integrates three specialized modules: the Multi-Order Statistic Extractor (MOSE) for multi-order structural alignment, First-Order Statistic Extractor (FOSE)-SVD for teacher-side semantic extraction via Singular Value Decomposition (SVD), and FOSE-FFT for student-side semantic stabilization using Fast Fourier Transform (FFT) with a Learnable Spectral Filter (LSF). Extensive experiments on the GRAZPEDWRI-DX and FracAtlas datasets demonstrate that SFD-KD consistently outperforms vanilla KD by +1.4~+4.4% mAP@0.5:0.95 across both Faster R-CNN and YOLOv8 frameworks. Ablation studies confirm the efficacy of the proposed statistical decoupling, while visualizations reveal superior preservation of multi-order fracture patterns. Notably, SFD-KD enables robust deployment in clinical cyber-physical systems with negligible inference overhead. Our code is available at: https://github.com/6720230811/SFD-KD.
The increasing complexity of medical data presents persistent challenges for traditional analytical methods. Conventional machine learning models often exhibit practical limitations in clinical settings, including a susceptibility to overfitting and a lack of transparency that can impede their adoption. This systematic review investigates the integration of attention mechanisms with ensemble learning, an approach this article formalises as Attention-Enhanced Ensemble Models (AEEM), to serve as a direct response to these issues. Conducted according to PRISMA guidelines, the review analyses 19 primary studies to evaluate the effectiveness of this combined approach in medical imaging. The analysis reveals that AEEMs demonstrate consistent performance improvements, with gains ranging from 1.05% to 29.5% over baseline models, alongside enhanced model interpretability and robustness. The systematic examination also identifies common patterns in component selection and highlights significant gaps in current reproducibility practices. The synthesis of these empirical findings informs a systematic decision framework designed to provide evidence-based guidance for the implementation of AEEMs.
The transport of healthcare materials is essential for a safe and well-functioning healthcare system. In remote areas, where transport can be particularly challenging, healthcare regions are investigating new approaches to maintain efficiency and safety whilst reducing costs. In this paper, we report on a field study with a design ethnographic sensibility that followed a pilot project of a healthcare region over the course of three months, testing a drone-based blood sample delivery system in the Gothenburg archipelago, Sweden. By adopting a holistic yet critical approach, we conducted observations of test flights, contextual interviews with nurses, project managers, transport coordinators, and a drone pilot, alongside a citizen survey. We combined this empirical data with other reports produced by the healthcare region. Through a reflexive thematic analysis, we show both opportunities and challenges—expressed as tensions—with the drone delivery system, questioning the techno-positive approach to the issue.
Hospitals and other healthcare settings benefit from the development and deployment of innovative interactive technology to better serve the needs of patients and employees alike. However, there are likely incompatibilities between the typical processes surrounding technology innovation and the processes that are central to a hospital’s functioning to facilitate patient care and meet employee needs. In this paper, we examine hospital administrators’ perspectives on the deployment of novel interactive technology in a real-world hospital setting. We conducted 13 semi-structured interviews with non-patient-facing employees at a rehabilitation hospital that had recently deployed novel smart patient rooms across all 75 rooms. Our findings indicate that prototyping was crucial for informing deployment decisions; however, scaling introduced unforeseen challenges, including financial and staffing complexities and an incomplete understanding of costs versus benefits. While the technology increased patient independence and reduced staff workload, reliability issues and overreliance on technology sometimes disrupted workflows and rehabilitation. These findings contribute by providing insights into the complex organizational effort involved in large-scale technology deployment, as well as the decision-making process for designing, deploying, and evaluating it. Future HCI researchers should explore opportunities to make HCI methods for ideation, prototyping, and evaluation more accessible in this setting.
This work explores the co-ideation of the Digital Pain Companion, a co-design proposal for a music-integrating wearable interface to improve self-management and person-centred care in chronic pain. We conducted two co-design workshops, informed by preliminary interviews, to explore the potential role and design of music-integrating technology in supporting daily pain self-management, particularly when interacting with the psychosocial dimensions of chronic pain. Our findings suggest that designing music-integrating technology to support daily psychosocial experience of pain needs to be more about distraction, self- selection, communication, and somaesthetic appreciation, rather than the design of the music itself and/or the acceleration of physical performance. The body emerged as central to the psychosocial experience of pain, linked not only to physical movement but also to embodied forms of psychological tension. We identified early evidence for two design opportunities—personal and social interaction—through which the Digital Pain Companion and similar interfaces could operate to increase ownership over the pain experience, facilitate effective pain communication, and reduce associated stigma. We contribute empirical and conceptual insights into how music can mediate the psychosocial experience of chronic pain, offering design implications for multimodal, person-centred technologies in digital healthcare and for the broader use of music-supported interventions in chronic health management.
This study presents PROVIMAPS, a framework for ensuring medical image data provenance and integrity in connected healthcare environments. The proposed approach introduces a software-based Device Fingerprint (DFP) generation method using intrinsic device parameters and SHA-256 hashing to uniquely identify the imaging source. The DFP is securely embedded into medical images using a DWT-based hybrid (DWT+DCT+SVD) watermarking technique, selected after a comparative analysis for its superior PSNR and SSIM performance across various medical image datasets. Additionally, an Image Average Intensity Profile (IAIP) is analyzed to detect image tampering and fused with the DFP for enhanced robustness. The combined signature ensures both device identification and content integrity. Comprehensive evaluation under various image attacks demonstrates the method’s high robustness, computational efficiency, and resilience against manipulation. The PROVIMAPS framework offers a secure, low-cost, and scalable solution for maintaining trust in telemedicine, Internet of Medical Things (IoMT), eHealth, and Medical Cyber-Physical Systems (MCPS) applications, enabling reliable verification of image authenticity and source in distributed healthcare settings.
Endometriosis is a complex and poorly understood chronic condition with variable symptoms. While many personal informatics tools support self-tracking, fewer leverage artificial intelligence (AI) to generate actionable insights for care. This research applied a digital phenotyping approach to characterize weekly health statuses using self-tracking data from the Phendo app — a research platform co-designed with endometriosis patients. Applying an unsupervised probabilistic mixed-membership model to week-level data, we generated temporal health status phenotypes that captured severity-based illness dynamics. We also created a rule-based phenotype to explore an alternate approach. We evaluated these phenotypes with a mixed-methods user study, which revealed complexities with the ability to align computational representations of health status with lived illness experiences as perceived by individuals. At the same time, we document the value of these representations, even when imperfect. We found significant mismatches between participants’ self-assessments and computational health status phenotypes, driven by personalized interpretations of health indicators, symptom under-reporting shaped by stigma, and the inherent complexity of characterizing subjective health experiences, which result in fluctuating perceptions of symptom severity. Future work will be needed to create mechanisms to better align computational representations of health status with human perceptions and values and individualized notions of health.
Human-in-the-loop principles promise to improve the performance and reliability of clinician-facing artificial intelligence (AI) by looping expert professionals into system development and evaluation. However, little is known about how such principles are realized in practice once AI systems are deployed in everyday clinical work. This paper presents findings from an ethnographic study of a long-standing automatic speech recognition system used for clinical documentation in a Danish hospital. We show how clinicians’ ongoing auditing of AI output becomes a necessary condition for system reliability while remaining ambiguous within existing clinical work organization. Our analysis identifies a set of three main tensions between clinical documentation and auditing work; auditing work and system alteration; as well as collaborative augmentation work and established professional accountabilities. To account for these dynamics, we characterize dual-purpose data work, capturing how clinical documentation simultaneously serves patient care and AI system development. We further describe the emergent form of overhead work to explain how human-in-the-loop is sustained through extensive, often invisible work required to make ongoing collaboration possible across organizational boundaries. Our findings advance understanding of post-deployment human-in-the-loop AI in healthcare and outline opportunities for computing, which makes auditing work meaningful, reduces overhead, and renders expert contributions transparent.
Accurate and efficient segmentation of gastro-intestinal (GI) tract images play a pivotal role in medical diagnosis. However, conventional deep learning models, such as U-Net, often incur high computational costs and memory overheads, hindering their deployment in resource-constrained clinical settings. This article proposes a novel hierarchical attention lightweight U-Net (HALU-Net) architecture designed specifically to address these limitations. HALU-Net incorporates a computationally efficient U-Net backbone, optimized through techniques like depth-wise separable convolutions and a hierarchical attention mechanism. While depth-wise separable convolution helps to reduce the model parameters, the hierarchical attention enables the model to selectively focus on both localized, fine-grained details and global contextual information within GI images at every decoding stage. This multi-level attention approach is essential for accurate segmentation of complex and variable GI tract structures. The HALU-Net model is rigorously evaluated on the publicly available UW-Madison GI tract image segmentation dataset. Results demonstrate that HALU-Net achieves competitive segmentation performance as measured by metrics such as dice coefficient and Jaccard index while boasting a substantially reduced parameter count compared to prevailing U-Net architectures. This significant decrease in computational complexity and memory requirements positions HALU-Net as a promising solution for facilitating accurate medical diagnoses of GI conditions, even in computation-constrained environments.
Artificial intelligence (AI) based mental health apps, especially chatbots, are increasingly being developed for youth, but rarely with their input, especially that of marginalised groups. This results in the development of apps that have low engagement and pose safety concerns. We use participatory methods to explore the preferences and design requirements of youth who face social exclusion, come from migrant backgrounds, or have low socioeconomic positions. We recruited 64 youths from youth work programs around the Netherlands and carried out 6 workshops. The first three explored the use of apps and large language models (LLMs) for well-being, while the last explored youth’s preferences for an LLM chatbot. Data was analysed thematically. Our results showed participants were open to using apps, preferring multifunctional apps, and identified human connection, self-development, and education as potential functions. However, they were reluctant to use chatbots, perceiving them as fake and lacking emotional intelligence. Instead, participants saw chatbots as providers of information, favouring shorter outputs with simple language, although they disagreed on how human-like chatbots should sound. Finally, the need for personalisation was emphasised, showing a desire for control with extensive customisation settings and clear privacy policies. Further work must be done to explore other relevant stakeholders’ views.
Radiology impression generation involves producing concise, clinically meaningful summaries from detailed imaging findings such as CT and MRI scans, serving as a critical aid in diagnosis and treatment planning. However, recent studies highlight a severe shortage of radiologists, particularly in low and middle-income countries, where there is fewer than one radiologist per 100,000 people, making timely expert interpretation a significant challenge. While advancements in AI, especially large language models (LLMs), offer promising potential to automate this task, current systems often suffer from hallucinations, omissions of key clinical details, and a lack of linguistic clarity, thereby raising serious concerns about their safety and reliability in real-world clinical settings. In this work, we attempted to address this issue by introducing RADO , a novel framework for radiology impression generation that integrates safety, faithfulness, and linguistic refinement rewards for preference optimization. To support robust evaluation, we introduce RIB , a real-world benchmark dataset curated and annotated by radiologists, spanning 1,429 annotated CT and MRI findings and impressions across 27 study types. RADO enforces critical safety and factuality constraints via carefully designed reward models and achieves state-of-the-art performance across multiple automatic and human evaluation metrics. Our framework significantly outperforms existing baselines, demonstrating improved factual consistency, reduced omissions, and higher clinical relevance, thus advancing the safety and reliability of generative AI in high-stakes medical applications. The code and dataset associated with the work are made available at RADO .
Social support is a key protective factor for young people's mental health, yet help-seekers may perceive supportive messages received through digital devices differently depending on content and style. We investigate how young people evaluate support messages from peers, adult mentors, therapists, and AI across four help-seeking scenarios. We conducted a mixed-methods secondary analysis of a survey study with 255 participants. We identified young people's contrasting opinions on various supportive communication elements that influenced whether they liked or disliked the message, shaping how well the support was received. These contrasting findings related to (1) message length, (2) normalization of the help-seeker's situation, (3) levels of encouragement, and (4) referral to formal support. These findings highlight how individual preferences moderate support message reception. We propose a reflection question bank to guide help-givers to consider how to incorporate help-seeker's unique supportive communication element preferences and trade-offs between contrasting perceptions, informing the design of more personalized and context-sensitive support in young people's mental health.
In the Netherlands, several public health platforms provide information and peer support to different patient groups such as people suffering from cancer or multiple sclerosis. With the rapid growth and development of Large Language Models (LLMs) in multiple sectors, these health platforms are considering the use of LLMs for personalized content generation and ease of information management. However, the sensitive nature of health information and personal experiences shared on such platforms introduces privacy and ethical risks, such as misinformation and bias. This study explores the concerns and risks associated with LLMs through interviews and focus groups with platform editors and users who use and access these platforms. Our findings show that risks related to content quality and quantity were most frequently identified. Moreover, the findings highlight the importance of disclosure if there is no human oversight, participants’ strong opposition to AI-generated blogs, and the potential of LLMs for personalization, provided users retain control over what they read. This work contributes to the ongoing discussion in human-centered computing about the ethical challenges and risks of adopting LLMs by presenting an empirical evaluation with editors and users. Moreover, the insights inform considerations and design guidelines for implementing LLMs on public health platforms.
Over the years, there has been an increase in the use of wearable sensors for high-precision Human Activity Recognition (HAR), ranging from personal fitness to remote patient monitoring. However, the continuous collection of biometric information poses a privacy risk, potentially leading to user profiling and data misuse. Today’s state-of-the-art approaches use standard encryption techniques to protect data in transit but leave it vulnerable during computation. For accurate HAR while maintaining the privacy of users’ data during collaborative analysis, we propose a hybrid, domain-agnostic framework that integrates Homomorphic Encryption (HE) with Secure Multi-Party Computation (SMPC). The proposed approach enables healthcare providers and device manufacturers to collaboratively analyze data without ever disclosing raw biometrics by utilizing HE for data encryption and secret sharing for collaborative computation. While this framework applies to general HAR scenarios, it will be crucially useful in the high-stakes domain of eldercare, where data privacy and regulatory compliance are critical. Reliable, Privacy-Aware Human Activity Sensing (RAHAS) achieved \(89.24\pm 0.95\%\) accuracy when tested on the widely used PAMAP2 dataset for HAR. Our analysis demonstrates that our privacy-preserving design provides side-channel resilience while maintaining utility comparable to clear-text baselines.
Video games have increasingly been used to support mental health through various game-based interventions (GBIs), including exergames, virtual reality therapy, and cognitive behavioral therapy-based games. While these approaches often emphasize novel game mechanics and gamification, such features may undermine intrinsic motivation, which is critical for meaningful internalization of mental health messaging. Narrative, a well-studied mechanism in communication theory for fostering engagement and persuasion, remains underexplored in GBIs for mental health. This systematic review investigates narrative GBIs (NGBIs) using the Preferred Reporting Items for Systematic Reviews and Meta-Analysis (PRISMA) guidelines. Seventeen studies met the inclusion criteria. The majority addressed depression and anxiety, and CBT was common as an evidence-based grounding. However, few of the studies investigated the narrative factorially, and the disparate measures used to evaluate success made it difficult to determine the overall efficacy of NGBIs for mental health. We conclude with suggestions for future directions for research and design of NGBIs.
LGBTQ+ communities are more likely to experience stigma-related stressors that lead to poorer mental health compared to cisgender and heterosexual individuals. For LGBTQ+ young adults, navigating the duality of their sexual orientation or gender identity and mental health can be challenging. Digital mental health (DMH) tools offer an alternative form of clinical care that may be useful for LGBTQ+ individuals wishing to remain anonymous, who need to access care from remote locations, or for those who do not have financial means to seek treatment. However, a user’s personal data can be contextualised in a way that perpetuates rigid gender norms. Gender perspectives that encompass the broad and fluid spectrum of gender identities and LGBTQ+ experiences are rarely embedded in design of DMH tools and intersectional differences remain largely underexplored in clinical research and practice. This study aimed to explore the perspectives and preferences of LGBTQ+ end-users on DMH tools, the data that it collects and how they may be developed to promote safety and wellbeing. We employed user-centred design methods in 3 workshops with gay men, lesbian women, and trans and gender diverse people (aged 22-32, n = 16). The study used a socio-cultural lens to consider how data justice can be maintained for LGBTQ+ individuals. We found that participants preferred to keep their identities private and control what and who their mental health data was shared with as a result of historical medicalization of their sexual or gender identities. Participants were concerned that heteronormative healthcare practitioners conflated their sexual or gender orientation with their mental health. Alternative support routes were found to be important for participants who felt their mental health challenges would be burdensome on their families. Socio-structural factors necessitate specialised DMH services and support for LGBTQ+ individuals. We then propose a set of design strategies for the future development of DMH tools for LGBTQ+ communities including the sensitive collection of demographic data, allowing the user to regularly update their gender or sexual identity, and integrated options for in-person or community-based support networks, along with the right for LGBTQ+ users to review and withdraw their data from DMH tools.
Large language models (LLMs) are increasingly attracting the attention of healthcare professionals for their potential to assist in diagnostic assessments, which could alleviate the strain on the healthcare system caused by a high patient load and a shortage of providers. For LLMs to be effective in supporting diagnostic assessments, it is essential that they closely replicate the standard diagnostic procedures used by clinicians. In this paper, we specifically examine the diagnostic assessment processes described in the Patient Health Questionnaire-9 (PHQ-9) for major depressive disorder (MDD) and the Generalized Anxiety Disorder-7 (GAD-7) questionnaire for generalized anxiety disorder (GAD). We investigate various prompting and fine-tuning techniques to guide both proprietary and open source LLMs in adhering to these processes, and we evaluate the agreement between LLM-generated diagnostic outcomes and expert-validated ground truth. For fine-tuning, we utilize the MentaLLaMa and Llama models, while for prompting, we experiment with proprietary models like GPT-3.5 and GPT-4o, as well as open source models such as llama-3.1-8b and mixtral-8x7b. Software Availability . We make all software artifacts available at this GitHub link ( https://github.com/kauroy1994/Large-Language-Models-for-Assisting-with-Mental-Health-Diagnostic-Assessments ). Institutional Review Board (IRB) . This study does not require approval from the IRB. It involves using clinician-annotated social media posts, authorized for research purposes. The primary objective is to evaluate the effectiveness of LLMs that incorporate diagnostic criteria for major depressive disorder and general anxiety disorder for assisting with mental health assessments.
Frustration is more than a byproduct of poor design–it is friction that exposes inaccessibility . People with mild cognitive impairment (PwMCI) constantly adapt, regularly confronting inaccessible technologies and environments, while undergoing cognitive changes that affect their lives. We consider frustration as a participant-driven design friction, which disabled designers argue illuminates and counters systemic ableism. We conducted participatory research with PwMCI and care partners, to understand their experiences with frustration. Participants reported frustration in most aspects of life, from how they communicate, navigate ability changes, and use technology. They designed social and technological interventions to counter frustration, supporting conflict navigation, home organization, and compensating for aphasia. We contribute a deep characterization of how PwMCI confront and reimagine frustration, and reflect on their agency in adapting to the challenges they face. In doing so, we highlight how frustration acts as a form of design friction, revealing systemic barriers that limit technology use. We critically examined these systemic barriers, often rooted in ableist societal norms, technological inequities, and accessibility barriers which all shape the use of technology by PwMCI. This paper guides designers to recognize the broader patterns of inaccessibility that affect PwMCI, situating their work within these dynamics to understand how frustration can reveal exclusionary practices and inspire more inclusive, equitable designs.
