When patients cannot be reached yet risk exists, GP trainees must make proportionate safety decisions without the stabilising anchor of a live consultation. This 'unreachable patient' problem is common in results handling, radiology follow-up, medication monitoring, and safeguarding, but is rarely taught explicitly as a clinical skill. The Unreachable Patient Algorithm (UPA) is a structured educational intervention that trains trainees to treat reachability as a variable in clinical reasoning. It combines rapid risk stratification, a graded action ladder, and a documentation standard that supports confidentiality, accountability, and loop closure. Supervisors introduce a single trigger (for example, a critical laboratory alert, abnormal imaging report, or safeguarding concern) and ask trainees to state the risk tier, the minimum safe action, the escalation threshold, and the fallback plan if contact remains unsuccessful. Teaching occurs through inbox simulation, case-based tutorials, and supervised 'pause and plan' moments in real clinics. A one-page note template and brief portfolio prompts reinforce consistency across placements. Early reflections suggest trainees escalate more appropriately, write clearer contingency plans, and rely less on vague statements such as 'tried to call'. UPA is low-cost, fits routine GP training, and offers a replicable method to reduce avoidable harm when contact fails in practice.
Brachial plexus birth injury (BPBI) results in individualized impairments in upper extremity (UE) mobility. A patient-specific understanding of these movement limitations is critical for optimizing decision-making and outcomes; however, current assessments may fail to capture the entirety of a patient's UE mobility. Reachable workspace quantifies an individual's global UE mobility by measuring the regions that can be reached by their hand. Despite most UE activities of daily living requiring adequate close-to-body function, current workspace approaches only assess far-from-body mobility. This study assessed the ability of a motion capture-based workspace approach to evaluate inner, close-to-body UE mobility in children with BPBI. It was hypothesized that the BPBI affected limb would have less inner, close-to-body workspace than the unaffected limb, especially in regions requiring UE movements commonly impaired in BPBI. Fifteen children with unilateral BPBI were assessed with motion capture using real-time visual feedback to measure UE workspace in all regions surrounding the body. All inner, close-to-body points reached by the hand were recorded. A two-way repeated measures ANOVA evaluated percentage workspace reached in each region surrounding the head, thorax, and abdomen. The affected limb had significantly less workspace reached than the unaffected limb for 8 of 9 regions (mean interlimb differences by region, 17.0-49.2%). Affected limb workspace deficits corresponded to common movement impairments in BPBI demonstrating the clinical relevance of this tool. Assessment of inner, close-to-body reachable workspace may provide a valuable new perspective on UE mobility to help guide clinical decision-making and outcomes assessment.
This paper formulates the "Minimum Vertex Cut with Reachable Set" (MVCRS) problem as an optimization framework to suppress botnet propagation in networked systems, and clarifies its computational complexity and algorithmic solutions. Building a firewall to minimize damage is essential for addressing botnet propagation in Internet of Things (IoT) networks. We define the basic MVCRS problem as minimizing the sum of the weight of the deployed resources and the resulting propagation scope. While we demonstrate that the constrained version of the problem is NP-complete, we show that the fundamental trade-off optimization model can be solved in polynomial time by reducing it to the maximum flow-minimum cut problem. This provides a theoretical baseline for optimal resource allocation in cybersecurity. Experimental evaluations reveal the limitations of conventional heuristics. In community-structured networks, the degree-based greedy algorithm overlooks critical bridge nodes, yielding an optimality gap of up to 72.6% above the theoretical minimum cost. Conversely, our exact algorithm consistently guarantees the optimal minimum cost (a 0% gap) with high statistical stability across diverse topologies. Furthermore, it scales efficiently to solve 100,000-node IoT networks within practical time limits, proving to be a reliable and efficient foundation for botnet suppression in complex real-world systems.
This study validates a clinically accessible approach for quantifying the Upper Extremity Reachable Workspace (UERW) using monocular AI-driven Markerless Motion Capture (MMC). Objective validation of such techniques for clinically oriented tasks is essential to support their adoption in clinical motion analysis. Nine adults without impairments performed the standardized UERW task, reaching targets distributed across a virtual sphere centered on the torso and displayed via VR headset. Movements were simultaneously captured with a marker-based system and eight FLIR cameras; monocular analysis was applied to two videos representing frontal and offset camera configurations. Agreement was assessed by comparing the percentage workspacereached across six of eight workspace octants between the systems. The frontal camera demonstrated strong agreement with the marker-based reference (mean bias: 0.61±0.12% reachspace per octant), whereas the offset view underestimated workspace reached -5.66±0.45%. Depth-related errors in the frontal configuration were confined to posterior octants, whereas the offset view introduced inaccuracies in both contralateral and posterior octants. These findings support the feasibility of a frontal monocular camera for UERW assessment, particularly for anterior workspace evaluation. While posterior accuracy remains limited by depth estimation and anatomical occlusion errors, the overall results demonstrate clinical potential for practical, monocular-camera assessments.
Imprecise probability models generated from data represent epistemic uncertainty by replacing the precise empirical distribution with a set of compatible probability distributions. When this set is described by reachable probability intervals, the induced bounds are tight, so the represented imprecision is not inflated by unattainable interval limits. This paper studies the informational effect of this replacement through the epistemic entropy gap, defined as the difference between the maximum entropy over the induced credal set and the Shannon entropy of the empirical distribution. The gap is a differential quantity: it measures the additional uncertainty introduced by the imprecise model beyond the observed frequencies. We analyze it for three reachable interval models generated from multinomial data: the Imprecise Dirichlet Model, the ϵ-contamination model and the approximated Non-Parametric Predictive Inference model. The analysis covers its main properties, its asymptotic behavior and its role in entropy equivalent calibration of model parameters. The results show that the entropy gap offers a common informational scale for comparing how different imprecise models represent the same empirical evidence, and helps interpret the degree of caution associated with limited data reliability and with empirical distributions that may otherwise lead to overconfident uncertainty assessments.
Community health volunteers (CHVs) are an important resource for supporting health service delivery, surveillance, and social programmes. However, retention and attrition of CHVs remain a big challenge. This study explored factors affecting the retention and attrition of CHVs working as village reporters (VRs) responsible for community-based death notification in the Malaria Vaccine Implementation Program (MVIP) in Malawi. This mixed-methods exploratory study, which intersected with the case studies, was conducted from November 2022 to March 2023 in nine rural districts in southern and central Malawi. Purposive sampling was used to select 64 study participants for qualitative interviews. Using case studies, we conducted six in-depth interviews (IDIs) with CHVs who had dropped out, were reachable, and agreed to be interviewed-many were dispersed, hesitant to attend meetings, or unreachable for focus group discussions (FGDs). We held five FGDs (n = 50) with CHVs who remained in the MVIP for shared norms and experiences and eight key informant interviews (KIIs) with health workers, opinion leaders, and program staff to provide insights into health workers motivation, supervisory and program perspectives. Thematic analysis and the social capital framework (roles, relationships, and empowerment) were used to analyse and interpret the qualitative data. The qualitative study was complemented by a cross-sectional survey involving 696 randomly selected participants from a pool of 2,861 CVHs to demonstrate the trends of retention over time. Descriptive statistics were computed from quantitative survey data, with retention as a primary outcome (defined as whether a CHVs was willing to stay in the program). At the start of the program in 2019, a total of 2,861 CHVs were recruited by March 2023; only 295 (10.3%) had dropped out. Among 696 CHVs surveyed, the most commonly reported factors associated with retention were incentives (643; 92%), participation in exchange visits (377; 54%), and managing a small geographical area (275; 40%). Qualitative data from FGDs and KIIs corroborated these findings and identified compassion for serving others, financial and non‑financial incentives, and flexibility to work across multiple programmes as key motivators for continued participation. Factors associated with attrition included experiences of ridicule or disrespect, lack of opportunities for personal development, and limited career progression. IDIs with CHVs who left the programme provided in‑depth accounts of these individual‑level drivers, which helped explain patterns observed in the survey. Engagement of CHVs in community-based programs can be promoted by offering opportunities to serve others, incentives, and flexibility to work on multiple programmes. However, it is also important to address ridicule-making fun or rude comments and limited personal and career development, which act as barriers to the continued engagement of lay health workers.
The epidemic potential of infectious diseases depends on how contacts connect individuals over time-a form of temporal connectivity that has rarely been quantified in resource-poor settings. We used the forward-reachable path (FRP)-the proportion of population reachable from an index person via direct or indirect connections-to quantify temporal connectivity in contact networks relevant for acute respiratory transmission. From empirical social-contact data collected in rural and urban Tamil Nadu, India, we derived contact-location-specific network statistics. These statistics were used to parameterize dynamic network models, simulate daily networks over one year, and compute FRPs. In both rural and urban networks, mean FRPs rose sharply on day 1, then either increased steadily at school and work or plateaued at home and at locations included in the other layer (that is, locations other than home, school, and work). By day 365, mean FRPs followed the order: home (0.06% [rural] and 0.03% [urban]) < school (11.96% [rural] and 9.14% [urban]) < work (12.55% [rural] and 26.72% [urban]) < other (40.54% [rural] and 67.99% [urban]). The mean FRP peaked at home among those aged ≥60 years, at school among those aged 10-19 years, and at work among those aged 40-59 years. Although FRP at home was bounded by household size, reachability expanded substantially through school, work, and other contacts. These findings indicate high temporal connectivity and substantial epidemic potential for acute respiratory transmission in these settings.
Soft robotic hands are well suited for handling fragile and geometrically diverse objects, yet many existing designs still rely on fixed finger layouts, which limits grasping adaptability when object size varies substantially. To address this issue, this study proposes a four-finger pneumatic soft robotic hand with a synchronous variable-stroke base mechanism. The design combines a rigid reconfigurable base with compliant soft fingers, allowing the radial positions of the fingers to be adjusted before grasping. A system-level kinematic model is established to describe the relationship between base stroke, finger bending, and the reachable workspace of the hand. A prototype is fabricated, and comparative grasping experiments are conducted under fixed-stroke and variable-stroke configurations using objects with different grasping cross-sections. The results show that the proposed mechanism achieves stable geometric reconfiguration and improves grasping performance when the initial finger spacing is matched to the object size. In particular, the variable-stroke configuration provides better grasp stability and a wider usable grasping range than the fixed-stroke configuration. These findings indicate that geometric reconfiguration at the hand level is an effective way to enhance the adaptability of multi-finger soft robotic hands.
This study develops a formally grounded verification framework for blockchain consensus mechanisms and smart contract behavior using Event-B and the Rodin platform. Unlike prior approaches that rely primarily on simulation or case-based validation of isolated contracts, this work integrates Finite State Machine (FSM) abstraction, invariant-driven proof, refinement modeling, and temporal logic verification to analyze Proof of Work (PoW), Proof of Stake (PoS), and mechanisms for double-spending prevention. Solidity smart contracts are abstracted into FSMs and encoded as Event-B machines, enabling the formal specification of state transitions and safety constraints. Safety properties-including transaction uniqueness, state consistency, access control enforcement, and ledger invariant preservation-are verified through automatically generated proof obligations in Rodin. A total of 312 proof obligations were generated, of which 287 (92%) were automatically discharged, and 25 were proven interactively, resulting in complete invariant coverage. Liveness properties were specified in Computation Tree Logic (CTL) and validated via model checking, confirming deadlock freedom and eventual validator selection under PoS conditions. Double-spending prevention was formally enforced using state-consistent ledger modeling, where uniqueness constraints were proven across all reachable states. Protocol-level consensus logic for PoW and PoS was refined across three abstraction levels, ensuring block integrity and validator correctness through stepwise refinement. The results demonstrate that machine-checked proofs provide verifiable correctness guarantees beyond simulation-based evaluation, establishing a rigorous and reproducible verification pipeline that enhances correctness assurance and protocol-level robustness in blockchain systems.
Generative artificial intelligence (GenAI) is rapidly entering consumer health information environments, yet patient readiness for safe adoption in cancer survivorship remains unclear. This study assessed preparedness for GenAI adoption among Chinese cancer survivors and identified correlates relevant to equitable implementation. We conducted a multi-center, cross-sectional online survey among digitally reachable adult cancer survivors recruited via clinical encounters, WeChat patient groups, and peer referral across three oncology centers in Sichuan, China. The questionnaire was administered on Wenjuanxing. The primary outcome was a theory-informed, study-specific 0-100 GenAI adoption preparedness composite derived from five Likert items: perceived usefulness, perceived ease of use, access to guidance/support, privacy concern after reverse coding, and near-term intention. Secondary outcomes included GenAI awareness and prior use, willingness for report explanation and symptom advice scenarios, and health information ability. Multivariable linear regression with robust standard errors estimated adjusted associations with preparedness, with sensitivity analyses addressing data quality flags and recruitment pathway. From 1,062 survey visits, 876 participants comprised the analytic sample. Mean preparedness was 57.8 (SD 24.2) with acceptable internal consistency (Cronbach's alpha 0.75). Awareness of GenAI was 61.6 and 40.6% reported prior use. Near-term intention to try GenAI for survivorship information tasks was endorsed by 51.3%. Willingness was higher for test report explanation (56.1% agree/strongly agree) than for symptom advice with referral prompts (40.3%). Preparedness was lower among participants older than 60 years versus 18-45 years (beta -8.1, 95% CI -12.0 to -4.3) and higher with prior generative AI use (beta 9.3, 95% CI 5.7-12.9), higher self-rated generative AI knowledge (beta 5.4 per 1-point, 95% CI 3.7-7.0), and greater health information ability (beta 2.0 per 10 points, 95% CI 1.2-2.7). The model explained 36% of variance (R2 0.36). Among digitally reachable cancer survivors in Sichuan, preparedness for GenAI adoption was moderate and strongly use-case dependent, with lower readiness among older survivors. The online-only sampling strategy means that the observed preparedness level may overestimate readiness in the broader survivorship population. Implementation should begin with lower-risk applications, such as report explanation and question preparation, paired with guidance on verification, privacy protection, and clear escalation to clinicians.
Near-limit emergency maneuvers are dominated by nonlinear tire saturation and stability constraints, calling for a short-horizon handling envelope that can be queried online. We formulate a stability-constrained finite-time reachability tube for a planar single-track vehicle with Magic-Formula tires. For a discretized primitive library and an environment channel [Formula: see text], where μ denotes tire-road friction, s road grade, and w a crosswind surrogate, the tube [Formula: see text] aggregates all states that remain dynamically admissible and reachable for [Formula: see text]. We prove compactness and, under a local robustness-of-admissibility condition, a Hausdorff-continuity result, and derive sampling-complexity scaling with time resolution, geometric tolerance, and environment dimension. Offline, we construct tube slices using analytic stability screening and nonconvex α-shape reconstruction in a normalized feature space, enabling fast membership queries for online feasibility screening. The main high-fidelity benchmark is carried out on the nominal flat-road/no-wind slice [Formula: see text] with friction-only variation; the Supplementary Information extends the study to fixed grade-only, wind-only, and combined off-nominal slices, and includes a stratified CarSim spot-check on the separated grade-only and wind-only cases. On an extreme-scenario benchmark, the tube screen provides a conservative triage gate and deployment-oriented diagnostics, including conflict-time statistics, risk-coverage frontiers, baseline comparisons, and guard-band sensitivity. The approach bridges reachability analysis and reliability-facing safety screening for near-limit vehicle operation.
Urban venues serve as arenas for social mixing, yet less is known about how public transit infrastructure shapes the geography of mixing at specific locations. This study examines how transit catchment diversity-the socioeconomic heterogeneity of populations reachable by public transit-associates with visitor diversity at points of interest (POIs) in nine Swedish and three US cities. Using mobile phone GPS traces and aggregated foot traffic data from 2024, we compute visitor diversity indices based on visitors' home-neighborhood birth-background composition and employ spatial regression models and geographically weighted regression (GWR). Transit catchment diversity positively predicts visitor diversity across nearly all cities, but this association is robust only in the largest metropolitan areas; in smaller cities, the coefficient attenuates to insignificance once geographic catchment composition, centrality, and venue density are controlled. Spatial spillovers in visitor diversity follow general geographic proximity rather than shared transit-stop connectivity, suggesting that the association operates through catchment population composition rather than station-level linkages. Transit-diversity hotspots occur not in already-diverse venues, but in lower-diversity POIs with lower commercial density, greater distance from transit in US cities, and greater centrality in Sweden. These patterns are consistent with transit-accessible population composition being associated with visitor diversity, particularly where alternative pathways to diverse co-presence are limited.
Safety-critical Industrial Internet of Things (IIoT) sensor networks deployed in disaster scenarios require intelligent routing mechanisms that prioritize mission-critical packets without relying on centralized coordination. Federated learning on resource-constrained edge nodes presents three primary challenges: the absence of an interpretable supervisory signal, the inability to act conservatively based on per-inference confidence, and vulnerability to partial node availability. The proposed FedCARE framework addresses these issues by employing a Mamdani Fuzzy Inference System to generate traceable criticality labels from multi-modal sensor telemetry, a dropout-aware aggregation protocol that normalizes over only reachable nodes, and a confidence-gated resolver that defers to symbolic fuzzy classification when model confidence is insufficient, otherwise applying an auditable maximization rule to prevent under-prioritization of safety-critical data. Evaluation on 50-, 100-, and 200-node Watts-Strogatz topologies under fault rates up to 50%, using the Edge-IIoTset and WUSTL-IIoT-2021 benchmarks, demonstrates 99.00% critical recall and up to 1.8× higher overall-packet delivery compared to RPL-RP under severe fault conditions. Routing improvements are primarily attributed to fuzzy criticality labeling and multi-path replication. These findings indicate that fuzzy-supervised federated inference offers a practical and interpretable solution for safety-critical IIoT routing, with an observed energy overhead of 7.8% per delivered packet.
This article presents time-dependent outputfeedback and state-feedback sampled-data control strategies for achieving both state and output reachability in permanent magnet synchronous generator-based wind turbine systems using a fuzzy approach. First, a nonlinear wind turbine model is represented as a set of fuzzy linear subsystems subject to bounded disturbances and parametric uncertainty. Unlike conventional sampled-data control schemes, a unified samplingtime- dependent fuzzy control framework is developed for both state-feedback and output-feedback cases. The framework varies across sampling periods and incorporates Bernoulli random packet dropouts, thereby forming a closed-loop system. Next, the fundamental Lyapunov component is modified by incorporating aperiodic sampling with various weighting levels. A samplingvariable- dependent discontinuous Lyapunov-Krasovskii functional, combined with a fuzzy membership function-dependent $\mathcal {H}\_\infty$ technique, is employed to derive sufficient reachability conditions. Finally, the simulation results, including comparative studies with existing approaches, demonstrate the applicability of the proposed control strategies and confirm improvements in terms of allowable maximum sampling period, reduced H$\mathcal {H}\_\infty$ performance bounds, tighter reachable-set ellipsoids, and fewer decision variables.
Neural circuits are remarkably robust to perturbations that threaten their function. Activity-dependent homeostatic plasticity (ADHP) is a stabilizing mechanism that supports robustness by tuning neuronal ion conductances to combat chronic over- or under-activity. Its restorative capacity has been demonstrated in the pyloric circuit of the crustacean stomatogastric ganglion, whose neurons must burst in a specific order to coordinate digestive muscles. After disruption by physical and pharmacological manipulations, this circuit reliably recovers not only the activity levels of constituent neurons, but also the proper burst order. But how could ADHP, operating only on local information about each neuron's average activity, maintain higher-order circuit properties? We explored this question in a computational model of the pyloric pattern generator. We first optimized a set of pyloric-like networks, then optimized ADHP mechanisms for each network to restore its pyloric character after parametric perturbations. This was possible for some networks and impossible for others, so we aimed to explain this disparity. We found that successful homeostatic regulators target average neural activity levels which happen to occur only among pyloric circuits and not among non-pyloric ones, within the set of reachable circuit configurations. Therefore, in subsets of parameter space where such dissociation is possible, activity carries indirect information about burst order, which ADHP can exploit to maintain pyloricness. Other subsets, whose pyloric averages are inseparable from non-pyloric ones, cannot be perfectly regulated. This separability property may explain differences in recovery capacity across perturbations and across individuals.
The design of active pneumatic upper limb exoskeletons is complicated by the challenge of reliably determining a kinematically safe workspace. Existing analytical kinematic methods are not sufficient to predict geometric collisions between elements of closed kinematic chains, which poses risks of mechanical damage and threats to user safety during exoskeleton operation. This paper proposes a hybrid algorithm for verifying the workspace of a pneumatic exoskeleton, combining analytical modelling in MATLAB R2020b based on the Product of Exponentials (PoE) method with high-performance static simulation in the Unity environment. At the initial stage, a discrete set comprising 758 million positions of the upper exoskeleton manipulator was generated. Subsequently, a multithreaded two-stage filtering process was implemented: analytical verification of rod stroke limits and angular constraints, followed by the detection of physical intersections of solid-state meshes using the PhysX engine. The results indicate that while the analytical model filters out 99.6% of invalid configurations. Yet, among the remaining positions-formally correct from a mathematical standpoint-up to 50% lead to critical geometric collisions or breaks in the kinematic chain. The computational efficiency of the proposed architecture enabled full static workspace verification in under 20 min. A reachable zone topology was established, revealing pronounced asymmetry and the presence of a "manoeuvrability core" in the user's anterior hemisphere. The developed algorithm generates a verified set of kinematically safe exoskeleton states, providing a foundation for the kinematic safety layer of a hierarchical control system. These findings demonstrate the necessity of complementing analytical kinematics with physical collision detection when designing hybrid kinematic mechanisms, and the approach can be applied to verify collision-free movement trajectories in various robotic systems. The approach can be applied to verify collision-free movement trajectories in simulation, with physical validation deferred to future work.
The optimal prehospital transport strategy for patients eligible for mechanical thrombectomy remains debated. We evaluated how direct versus secondary transfer to a comprehensive stroke center (CSC) influences outcomes in a large metropolitan stroke network. This prospective registry-based cohort study included all patients undergoing thrombectomy between 2015 and 2022 in a city of 1.7 million inhabitants. The network comprised 11 primary stroke centers (PSCs) referring to a single CSC, all reachable within 60 minutes. Among 2,017 patients, 242 (12.0%) were transported directly to the CSC, whereas 1,775 (88.0%) were secondarily transferred after initial assessment at a PSC. Functional outcome was assessed using the modified Rankin Scale (mRS) at 90 days. Multivariable logistic and Cox regression models were used to identify independent predictors of good functional outcome and survival. The median onset-to-CSC admission time was shorter after direct versus secondary transfer (101 [7-1,071] vs. 240 [40-1,411] min, P<0.001). A good functional outcome (mRS 0-2) was more frequent after direct transport (50.3% vs. 40.5%, P=0.015). After adjustment for age, National Institutes of Health Stroke Scale score, thrombolysis, posterior circulation stroke, and comorbidities, direct transport independently predicted good functional outcome (odds ratio: 1.74; 95% confidence interval [CI] 1.25-2.42; P=0.001). Among patients receiving combined thrombolysis and thrombectomy (n=1,110), direct transport was associated with improved long-term survival (log-rank P=0.026; adjusted hazard ratio: 0.67; 95% CI 0.47-0.95; P=0.026). Even within a metropolitan network with 1-hour access, direct transport to a CSC shortens treatment delays, improves functional recovery, and enhances survival among patients undergoing combined reperfusion therapy.
Suicide is a concern in rural veterans, particularly following acute mental health admission. Rural veterans can experience barriers accessing treatment, including mental health treatment. The MISSION Act of 2018 aims to alleviate some challenges through access to Community Care, which has since raised concerns about suicide prevention effort discrepancies and barriers between Veterans Affairs (VA) and non-VA providers. Despite concerns, limited studies exist of suicide prevention strategies in rural Veterans receiving acute mental health Community Care. We conducted a pilot study of the suicide prevention program, VA Brief Intervention and Contact Program (VA BIC), in rural veterans who accessed acute mental health Community Care across Northern New England. VA BIC supports treatment engagement and health-promotion behaviors in veterans after Community Care discharge. We developed a process to recruit eligible veterans into a three-month study of VA BIC. We assessed the feasibility of VA BIC in Community Care veterans and collected mental health outcome pilot data. Among 44 eligible and reachable patients, 45.5% (N=20) consented. Retention was high with 95.0% of patients completing all assessments. Among the 10 VA BIC participants, adherence was high with 90% completing all eight visits and 100% completing six visits. Suicidal ideation, hopelessness, social connectedness, and suicide-related coping trended towards improvement in the VA BIC group at follow-up. It is feasible to study VA BIC following discharge from acute mental health Community Care, and the intervention may benefit veterans. Future studies should confirm the efficacy of VA BIC in reducing suicide risk in non-VA settings.
Multi-modal models that fuse neuroimaging with clinical assessment data represent the current state of the art for automated Alzheimer's disease detection, yet their adversarial robustness remains poorly understood. We systematically investigated adversarial vulnerability in CogniNetMM, a model that fuses 3D structural MRI volumes with neuropsychological clinical variables, using the Fast Gradient Sign Method and DeepFool in three modality configurations: MRI only, clinical variables only, and both jointly. We further introduced a mean attack framework, in which the average perturbation vector across samples serves as a fixed-direction probe of the decision boundary, decoupling the contribution of attack direction from that of sample position.During training, we identified modality collapse, a training instability in which the fusion layer progressively suppresses the MRI pathway. Collapse probability decreased with larger batch sizes and was further reduced by stratified sampling on class-imbalanced data. Across all clinical variable configurations and both attack methods, the joint multi-modal attack achieved higher success than the average of the two unimodal attacks. For DeepFool, the advantage was strong enough that the joint attack outperformed each unimodal attack individually near the decision boundary. The mean attack replicated this result under a fixed perturbation direction, confirming that the advantage is a structural property of the fused decision boundary rather than an artifact of per-sample gradient alignment. These findings are consistent with a concave original-class region in the joint input space: non-linear modality coupling creates adversarially reachable regions that neither modality perturbation can access independently. Together, these results show that heterogeneous data fusion introduces emergent adversarial vulnerabilities beyond what unimodal analysis predicts, and that standard training practices on imbalanced medical datasets carry a risk of silent modality suppression.
Given the underrepresentation of Latinx or Hispanic-Identifying (LHI) individuals in STEM fields, this research examines the effectiveness of a parent intervention focused on the community cultural wealth for Spanish-speaking parents of LHI STEM undergraduate students in their first year of college who navigate systemic barriers such as racism, inequity, and marginalization (Yosso, 2005). Using a longitudinal quasi-experimental design with propensity score weighting, this study compared six- and 12-month post-test parent-child outcomes (Parent STEM Conversations, Parent STEM Values, Parent Support, & Family Obligations) for LHI undergraduate students whose parents participated in an intervention with two control groups: parents who were interested but unable to attend (C1) and parents not interested in or reachable for the intervention (C2). Longitudinal latent trait-state structural equation models indicated that the treatment group reported higher state parental STEM conversations than C1 and C2 at the six-month follow-up, and higher state parental support than C2 at the 12-month follow-up. Although the treatment group reported higher state family obligations than the C2 group at the six-month follow-up, group differences dissipated by the 12-month follow-up. Student perceptions of parents' STEM values did not differ between the treatment and control groups. These findings underscore the complexity of addressing cultural and contextual factors in parental interventions, implicating the need for more tailored approaches to fully support LHI students in STEM fields that emphasize cultural wealth. This research contributes to the growing body of literature on the role of family support in the academic success of underrepresented students in STEM disciplines.