Search Results (223)

Contemporary intensive care operates in an environment with high-complexity cases, large volumes of information, and vast physiological, biological, and therapeutic data, collected from laboratory results, investigations, and therapies for organ support, as well as from systems that operate in parallel. The lack of interoperability contributes to information overload, alarm fatigue, and delayed decision-making. The Smart ICU concept has been developed to address these limitations by integrating medical devices, information systems, and artificial intelligence into a unified system that allows interoperable data integration and predictive analytics. Aim: The purpose of this article is to provide a narrative review of the Smart ICU concept, with a specific focus on the cardiac intensive care unit. It summarizes Smart ICU architecture, data integration, clinical support, and applicability in monitoring perioperative inflammation in cardiac surgery. We describe the Smart ICU architecture, from data acquisition to storage and analytics, highlighting the differences between Smart ICU, artificial intelligence, and Tele-ICU, and we underline predictive analytics as a supportive tool, as well as its influence on clinical outcomes. Cardiac ICU application: Cardiac ICUs offer a data-dense, temporally well-defined model following cardiac surgery with cardiopulmonary bypass, where data concerning patients’ hemodynamics, perfusion data, and biological and inflammatory markers intertwine. Cardiac Smart ICU models could recognize early signs of hemodynamic compromise and low cardiac output states and identify early indicators of post-cardiac surgery complications. Neutrophil activation and complete blood count-derived indices may be used as dynamic biological data for Smart Cardiac ICU models. Conclusion: The Smart Cardiac ICU may support earlier risk stratification, and therefore earlier diagnostic and therapeutic interventions, but its clinical value requires prospective, multicenter validation. Cardiopulmonary bypass-induced inflammation may offer an ideal setting to integrate physiological, procedural, and immunological data into bedside predictive models. Full article

The Internet of Medical Things (IoMT) is a promising technology for enabling smart and efficient healthcare systems through continuous physiological monitoring, early anomaly detection, and proactive patient management. However, IoMT sensors are vulnerable to False Data Injection Attacks (FDIAs), in which adversaries can manipulate sensor measurements to compromise diagnostic accuracy, mislead clinical decision-making, and threaten patient safety. Existing detection approaches often rely on population-level statistical models that may not fully capture individual physiological variations or residual-based thresholds designed for relatively simple attack scenarios, limiting their ability to exploit the spatio-temporal dependencies of multi-sensor physiological streams and detect stealthy or adversarial FDIAs. This paper proposes a patient-specific FDIA detection framework based on a Graph Convolutional Network–Gated Recurrent Unit (GCN–GRU) digital twin that learns an individual patient’s normal physiological behaviour from clean baseline telemetry. The trained digital twin is integrated into a Kalman filter as the state prediction model, and the resulting standardised innovation residuals are used as detection features. To characterise stealthy attack behaviours, four complementary window-based feature groups are extracted from the innovation sequence: innovation statistics, sensor correlation drift, temporal smoothness, and uncertainty mismatch. A CNN-1D classifier is then trained to learn discriminative temporal attack patterns from these features for accurate detection. A structured attack taxonomy comprising five stealthy and adversarial FDIA scenarios is developed, where attacks are injected as smooth gradual or abrupt coordinated modifications to sensor measurements while remaining within plausible physiological ranges. Experiments conducted on the WUSTL-EHMS-2020 benchmark dataset demonstrate that the proposed framework achieves an F1-score of 94.3%, outperforming Isolation Forest and PCA Reconstruction by 34 percentage points. Furthermore, the proposed framework reduces the false alarm rate to 3.6%, compared with 35.1% and 9.2% achieved by Isolation Forest and PCA Reconstruction, respectively. These results demonstrate the effectiveness of the proposed framework for reliable detection of stealthy FDIAs in IoMT-based healthcare systems. Full article

Telemedicine in emergency and remote care relays medical images from ambulances and rural clinics to a hospital edge-computing server over a congested wireless uplink. Existing work prices such transmissions per bit or per quality-of-experience; neither metric captures the clinical value of a medical transmission. Diagnostic utility vanishes below a modality-specific acceptability floor rather than degrading gracefully, the deadline is determined by triage acuity rather than by the network, and a missed finding is far costlier than a false alarm. A per-bit clearing price therefore disadvantages the node that has expended local compute to produce a compact, diagnostically sufficient stream. We propose SemAuc, a triage-aware truthful mechanism for medical-image admission over a rate-splitting uplink, in which the shared semantic knowledge base rides the common stream, and case-specific residuals ride private streams. SemAuc filters tiers below the diagnostic floor and beyond the clinical deadline, reserves a regulated-price lane for life-threatening cases, and allocates remaining capacity through a single-parameter contestable auction whose bid-independent pre-selection step satisfies the conditions of Myerson’s lemma. The contestable lane is dominant-strategy truthful, individually rational, near-linear in the number of nodes, and achieves a constant-factor density-greedy welfare guarantee; the clinical lanes follow from triage policy without disturbing these properties. Diagnostic value is grounded by an offline kernel fitted on BraTS and CheXpert. On a Rayleigh-faded uplink at two hundred contending nodes, SemAuc preserves the high-acuity diagnostic service-level objective where bit-centric benchmarks fail, and tracks the offline optimum. Full article

Hypertension (HTN) is one of the greatest public health challenges of the 21st century. Its prevalence has reached alarming levels in recent decades. The search for effective prevention and treatment strategies includes lifestyle choices, such as dietary interventions. Although not applicable to everyone, particularly interesting in this context is the ketogenic diet (KD), known to clinicians and also applied in epilepsy treatment for over a century. Its possible beneficial effects are increasingly often reported in many other conditions, including HTN. The aim of the present study is to analyse the effect of a KD on blood pressure, and the mechanisms that may modulate that effect. Based on the available literature, key potential pathways of the influence of a KD on blood pressure regulation have been identified: (1) reduced body weight; (2) reduced visceral adipose tissue; (3) improved insulin sensitivity; (4) improved water and electrolyte balance; and (5) anti-inflammatory effect. Meta-analyses and randomised controlled trials consistently indicate that ketogenic dietary interventions are associated with reductions in blood pressure, although the magnitude of the effect varies considerably depending on the ketogenic diet model, study population, and comparator. In light of these observations, it has been found that in patients undergoing pharmacological treatment, it may be necessary to appropriately reduce antihypertensive medication dosage in advance. To maximise the hypotensive effect of a KD, the need to ensure an adequate supply of potassium, magnesium, and high-quality products, as well as proper hydration has been emphasised. Further studies are needed, with the effect of a KD on systolic and diastolic blood pressure as the primary endpoint, taking into account the role of the qualitative composition of the diet in the observed effects. Full article

Clinical artificial-intelligence systems are starting to act across a course of care, not to answer one question at a time. Their safety is checked by methods that sample the input space: a test suite tries some inputs, a language-model reviewer reads some cases, a physician panel audits some cases. A sampling check can pass a safety rule and still miss the rare input that breaks it, such as a documented obligation dropped several encounters later. This study measures that gap and releases CIV-Bench, a public benchmark of 832 clinical rule sets with safety properties across eight whole-person domains, in single-encounter and longitudinal forms, plus a computational stress tier, each with independently established ground truth. We compare formal verification, which uses a satisfiability-modulo-theories (SMT) solver to check every possible input at once, against the methods used in practice: random unit testing, language-model judges, and a blinded physician panel. Formal verification detected all 612 violations, raised no false alarm, and returned no unsound verdict; for each item it returned either a proof that the rule holds over every input or one concrete input that breaks it. A frontier language-model judge matched this detection, but it returned a pass rate over sampled cases rather than a guarantee, at three orders of magnitude more compute per item. The general open-weights judge returned unsound verdicts on the computational stress tier; the medically fine-tuned judge was unsound far more widely, collapsing on the longitudinal properties despite strong single-encounter medical detection, so medical fine-tuning did not close the gap. Unit testing and the physician panel missed the deep, cross-encounter violations that hold a course of care together. Formal verification is set apart not by a higher detection rate but by the kind of evidence it returns: a proof over the whole input space, a replayable counterexample, or an explicit statement that it cannot decide. The guarantee holds for the decisions placed in this layer, and it depends on the safety rule being specified correctly. Full article

Wireless sensor networks underpin the Internet of Medical Things (IoMT), where connected medical devices and wearable body-area sensors stream patient telemetry across hospital networks, and securing this traffic is safety-critical. Machine learning intrusion-detection systems for the IoMT are usually evaluated within a single public dataset, where they report near-perfect detection scores; whether that accuracy predicts performance in a different hospital has not been measured systematically. We frame the deployed detector as a lightweight human-in-the-loop edge AI agent that observes local traffic, asks an analyst to label a small set of representative flows, trains a tiny model on-premises, and monitors the traffic distribution for drift. Using cross-corpus shift across four public datasets, comprising three sensor-network corpora (WUSTL-EHMS-2020, CIC-IoMT-2024, TON-IoT) and an unrelated enterprise-network corpus, as a reproducible stand-in for cross-hospital deployment, we find the shift is near its theoretical maximum on all twelve transfer directions, that unlabeled domain-adaptation methods collapse on the hardest medical-telemetry targets, and that external pretraining adds no measurable benefit once training schedules are matched, consistent with a domain-adaptation error bound that the measured divergence renders vacuous. The same 1206-parameter network trained from scratch on ten to fifty locally labeled flows matches or exceeds every transfer alternative across all four corpora; on the six sensor-network transfer directions it also outperforms larger and classical local models. A closed-loop evaluation on simulated drifting streams shows the calibrated trigger detects drift within two 500-flow windows at under one false alarm per hundred stationary windows, retraining restores accuracy, and the agent tolerates ten percent analyst error for about five F1 points; we also map the detector’s exposure to white-box evasion and targeted label poisoning. Full article

Background: Large language models (LLMs) are increasingly used by patients for obtaining medical information; however, concerns remain regarding their clinical safety, reliability, and appropriateness of patient guidance. Evidence evaluating LLM performance in hemorrhoid-related patient questions remains limited. Objective: To compare the clinical accuracy, safety, and overall clinical adequacy of responses generated by ChatGPT, Gemini, and DeepSeek to hemorrhoid-related patient questions. Methods: In this cross-sectional comparative study, 25 hemorrhoid-related patient questions were developed and categorized into three predefined subgroups: basic informational questions, clinically significant scenarios, and misleading/risky patient statements. Responses generated by ChatGPT (GPT-5.3), Gemini (3.1), and DeepSeek (R1) were evaluated by two experienced surgeons using a consensus-based expert assessment approach and a structured 5-point scoring system assessing clinical accuracy, safety, appropriateness of patient guidance, and overall clinical adequacy. Critical errors and qualitative communication characteristics were also analyzed. Friedman and post hoc Conover tests with Bonferroni correction were used for statistical comparisons. Results: Overall response quality differed significantly among models (χ²(2) = 29.119, p < 0.001, Kendall’s W = 0.582). ChatGPT achieved the highest overall scores (5.00 ± 0.00), followed by Gemini (4.80 ± 0.41) and DeepSeek (4.12 ± 0.67). Significant differences were primarily observed between DeepSeek and the other models, whereas ChatGPT and Gemini showed comparable performance. Model divergence became more pronounced in clinically significant scenarios involving alarm symptoms, rectal bleeding, persistent symptoms, and acute anorectal pain. No model generated directly harmful medical recommendations or explicit guidance likely to result in substantial diagnostic delay. However, qualitative assessment demonstrated differences in communication style and risk communication. ChatGPT generally produced more balanced and context-appropriate responses, Gemini generated more explanatory responses, whereas DeepSeek showed a tendency toward disproportionately urgent or alarmist language in some higher-risk scenarios. Conclusions: Large language models demonstrated generally high clinical accuracy in answering hemorrhoid-related patient questions; however, notable model-specific differences were observed in clinical guidance, communication style, and risk communication, particularly in high-risk clinical scenarios. These findings suggest that LLMs may serve as useful supportive tools for patient education and health information delivery, although they should currently be regarded as systems that support rather than replace human clinical judgment. Full article

Background: Sarcomas are rare malignant tumors for which diagnostic delay is associated with poorer clinical outcomes. Primary care clinicians (PCCs) are often the first physicians to evaluate patients with suspicious soft tissue masses. This study aimed to assess training exposure, self-reported knowledge of soft tissue tumors/sarcoma (STT/S) alarm signs, and diagnostic practices among PCCs in Spain. Methods: We conducted a nationwide descriptive cross-sectional survey between January 2024 and May 2025. A structured questionnaire was distributed through scientific societies, in-person dissemination at healthcare centers, and direct professional contact. Participants included Primary Care physicians (PCPs), and medical intern residents (MIRs). The primary outcome was self-reported knowledge of STT/S alarm signs, assessed as a dichotomous variable. Results: A total of 642 clinicians participated, of whom 67% were female. Most respondents were Primary Care general practitioners (64%) or MIRs (31%), and 64% worked in urban settings. Overall, 38% of participants reported being aware of STT/S alarm signs. Undergraduate exposure to oncology- or sarcoma-related content was limited: 36% reported no training, 17% reported fewer than 10 h, and data were missing for 35%. Self-reported knowledge of sarcoma alarm signs was higher among younger participants, residents, and those with prior oncology training. Conclusions: Undergraduate exposure to sarcoma-related content and self-reported knowledge of STT/S alarm signs were suboptimal among PCCs in Spain. Targeted educational interventions and simplified referral pathways aligned with national recommendations may help improve earlier recognition and referral of suspected sarcoma. Full article

In this paper, we propose an outcome-based contrastive loss for imbalanced binary classification to alert to next-day health deterioration using care records and meteorological data. Long-term care facilities maintain daily care and observation records to monitor the health of older adults. Such objective records are particularly valuable when sudden deterioration occurs, enabling timely coordination with medical institutions. Predicting deterioration one day in advance could provide care staff with an actionable window to intensify observation and adjust care plans (e.g., scheduling additional vital checks or increasing fluid intake monitoring). This could potentially reduce emergency transports and ease the burden on already understaffed care facilities. However, for such predictions to be useful in practice, false positives must be suppressed. Because deterioration events are rare, class imbalance generates an excess of false positives, causing alert fatigue and increasing the risk that actual events go unnoticed. To address these challenges, we propose an outcome-based contrastive loss that contrasts actual deteriorating samples against false alarms conditioned on mini-batch prediction outcomes. The proposed loss contracts same-label pairs to shape local structure within each ground-truth label. The loss also separates actual deteriorating samples from false alarms among samples predicted as deteriorating, thereby directly reducing unnecessary alerts. As a result, compared with random oversampling with standard cross-entropy, the proposed model improved precision from 3.97% to 12.94% (+8.97 percentage points), while limiting the F1-score decrease to 0.71 percentage points (from 7.28% to 6.57%). Pair-design ablations and UMAP projections supported this mechanism by indicating clearer separation between actually deteriorating and false-alarm samples in the learned representation space. These results suggest a viable direction for alert systems that produce fewer unnecessary alerts, reducing alert fatigue and supporting more reliable deterioration detection in care settings. Full article

Prospective memory (i.e., remembering to complete future plans) and retrospective memory (i.e., memory for past events) are essential to daily functioning, but both are prone to everyday failures, such as forgetting to carry out an intended action (e.g., missing a medication dose) or inaccurately recalling past information (e.g., forgetting the details of a recent conversation). To mitigate these, individuals may rely on cognitive offloading, the use of physical actions or external tools to reduce retrieval effort (e.g., setting an alarm to avoid forgetting an important task). This study examined the impact of cognitive offloading on both prospective and retrospective memory, using two offloading strategies. In the first phase, 152 participants were instructed to send an email 48 h later at 7 p.m. and were randomly assigned to one of three conditions: a reminder (e.g., an alarm), a scheduled email for automatic delivery, or no reminder (internal memory). They also watched a news report. In the second phase, participants sent an email (prospective memory) and then completed a free recall question about the video (retrospective memory). Results show that both offloading conditions performed better in the prospective task. Notably, there were no significant differences in retrospective memory performance. Overall, cognitive offloading enhanced prospective memory and subjective confidence but did not influence retrospective recall, highlighting a dissociation between remembering when to act and remembering contextual information. Full article

Background: Aspirin, initially recognized for its anti-inflammatory, antipyretic and analgesic properties, holds a prominent role in the treatment of cardiovascular disease. The utility of aspirin in cancer therapeutics has been explored and stratified into COX-dependent and -independent mechanisms. COX2 gene expression has been demonstrated to be significantly upregulated in colorectal cancer and various other gastrointestinal malignancies including pancreatic, esophageal, and gastric cancer. This study investigates the relationship of aspirin use and outcomes in patients with colorectal cancer. Methods: The Nationwide Inpatient Sample (NIS) database from 2017 to 2022 was analyzed for patients age > 18 who were hospitalized for colorectal cancer and its decompensations using ICD-10 diagnostic codes. These patients were further stratified based on the long-term use of aspirin. The principal outcome of this investigation are the odds of in-hospital mortality, with secondary outcomes including odds of pulmonary embolism, portal vein thrombosis, acute kidney injury, septic shock, requiring an ICU level of care and odds of hepatic, pulmonary, gastrointestinal and peritoneal or retroperitoneal metastatic disease. Multivariate logistic regression accounting for hospital and patient characteristics was implemented for analysis, with the Charlson Comorbidity Index used to adjust for coexisting comorbidity burden; a p-value (p) of <0.05 was considered statistically significant. Results: In our analysis of the NIS, 596,160 patients were identified with colorectal cancer and 11.7% (69,750) of this population were identified with long-term use of aspirin. Aspirin use was identified to have a significantly reduced odds of in-patient mortality (adjusted odds ratio) [aOR] 0.530, p value < 0.001 95% CI (confidence interval): 0.460–0.617. Patients with aspirin use also demonstrated significantly reduced odds of adverse outcomes and gastrointestinal, hepatic, pulmonary and retroperitoneal/peritoneal metastasis; (aOR 0.606, 95% CI: 0.564–0.653, p < 0.001), (aOR 0.628, 95% CI: 0.582–0.678, p < 0.001), (aOR 0.676, 95% CI: 0.605–0.755, p < 0.001) and (aOR 0.751, 95% CI: 0.685–0.825, p < 0.001) respectively. Conclusions: In recent years, there has been an alarming increase in incidence of colorectal cancer, particularly amongst younger individuals with increased associated mortality. This mortality increase, albeit alarming, is a driving force for treatment innovation with continual examination of our repertoire of medications for possible repurposed applications. COX2-mediated signaling serves as a key promotor of tumorigenic molecular signaling that directly contributes to tumor cell proliferation, angiogenesis and metastasis in colorectal cancer. Aspirin use and its inhibitory action on COX2 demonstrated a significantly reduced odds of in-hospital mortality. Aspirin use is also associated with significantly reduced odds of developing metastatic disease to the liver, gastrointestinal system, lungs and peritoneum in patients with colorectal cancer. These findings convey that aspirin use reduces the likelihood of in-hospital mortality, major comorbid conditions and of developing metastatic disease as compared to those who do not use aspirin. Full article

Background and Objectives: Extracorporeal membrane oxygenation (ECMO) provides temporary respiratory or circulatory support when conventional therapies fail. Some patients do not recover and are not candidates for transplant or durable mechanical support. In these cases, continuing ECMO may no longer align with the patient’s goals. Compassionate ECMO discontinuation (CED) is the planned withdrawal of extracorporeal support with death anticipated. The term “compassionate” refers to the goal of minimizing suffering in the end-of-life process. This review proposes a reliability-oriented framework to standardize CED and reduce preventable distress for patients, families, and clinicians. Materials and Methods: We conducted a targeted narrative review of ethical analyses, consensus guidance, and empirical literature on planned ECMO withdrawal. The results of the narrative review were combined with our existing practical process for CED into this proposed reliability-oriented framework as a guide for clinicians. Recommendations were organized into a four-phase process model that emphasizes operational implementation, anticipatory guidance, and quality improvement. We included modality-specific considerations for veno-arterial (VA), veno-venous (VV) ECMO, and extracorporeal cardiopulmonary resuscitation (ECPR). Results: The framework includes four phases. Phase I, Anticipation and Alignment, emphasizes structured shared decision-making, early expectation setting, time-limited trials, palliative care integration, and predefined pathways for managing disagreement. Phase II, Preparation, includes interdisciplinary role assignment, a pre-withdrawal time out, family coaching on expected physiological changes, and preemptive comfort medications that account for ECMO-altered pharmacokinetics. Phase III, Implementation, prioritizes comfort first, pacing with explicit pause points, environmental controls to reduce alarms and visual distress, and modality-tailored sequencing. Phase VI, Aftercare and Learning Capture, includes bereavement support, standardized documentation, structured team debriefing, and recommended process measures to guide improvement. Conclusions: Viewing CED as a low-frequency, high-stakes clinical process supported by scripts, checklists, and iterative feedback can improve goal-concordant end-of-life (EOL) care, reduce suffering and family trauma, support clinicians, and strengthen ECMO program learning systems. Full article

The Internet of Medical Things (IoMT) is becoming a vital part of modern healthcare, enabling ongoing patient monitoring and remote diagnosis. However, as more devices connect to the internet, healthcare systems become more vulnerable to serious security issues such as unauthorized access, patient data manipulation, and Man-in-the-Middle attacks. Conventional Intrusion Detection Systems (IDSs) often struggle with the unclear and uncertain characteristics of IoMT traffic, which leads to reduced detection accuracy and increased false alarms. To address these challenges, this paper proposes ML-FSID-FIS, a multi-level feature selection-based Intrusion Detection System that employs a fuzzy inference system (FIS) for classification in IoMT networks. The model combines multiple feature selection techniques into a three-stage multi-level feature selection strategy to improve detection efficiency and strengthen the security of IoMT networks. In the first stage, four feature selection techniques—Random Forest, XGBoost, ReliefF, and Mutual Information—are applied to identify the most relevant features. In the second stage, a frequency-based consensus strategy is utilized to extract consistently selected features from the four top-ranked sets. In the third stage, an ensemble refinement using bagging-based ranking is employed to rank the remaining features, resulting in the selection of the top five features. From these, three candidate 3-feature groups are formed and evaluated, and the best-performing group is selected as the final input set for the fuzzy logic classifier. The FIS produces a continuous risk score that is mapped to a binary decision using a validation-selected threshold. When the proposed method was tested on the WUSTL-EHMS-2020 dataset and compared with other recent work using the same dataset, it showed strong detection performance while maintaining a very low false positive rate of 0.3%. This study is distinguished by its integrated design, which combines a three-stage multi-level feature selection strategy with fuzzy logic-based intrusion classification to improve feature efficiency and support interpretable intrusion detection in IoMT. Full article

Cancer is an alarming health concern and economic burden in both developed and developing countries. Recently, there has been a growing demand for new alternative medications with more effectiveness and fewer harmful effects. During the past decades, a set of chemotherapeutic agents has been developed to fight against a large spectrum of cancer types. Unfortunately, their use is associated with a high level of toxicity; they are expensive, also, and their deployment is restricted by the emergence of cellular resistance. Plant-based components are garnering attention due to their low toxicity, selectivity, efficiency, and ease of accessibility. Alkaloids are one of these targeted compounds. Indeed, they are a highly diverse group with basic heterocyclic nitrogen-containing alkaloids that exhibit potent anticancer effects against a large panel of solid and liquid tumors, such as lung, breast, leukemia, liver, and colon cancer. The main molecular mechanisms involved in alkaloids’ anticancer effect are the induction of apoptosis via the extrinsic and intrinsic pathways, DNA damage, and the inhibition of cell cycle progression. Amazingly, these auspicious compounds exhibited strenuous inhibitory effects against a whole range of key enzymes involved in cancer progression and metastasis, such as Cytochrome P450 (CYP450), Cyclooxygenase-2 (Cox-2), Lysine-Specific Demethylase 1 (LSD1), Poly [ADP-ribose] polymerase (PARP), and topoisomerase, mainly through two action modes, namely irreversible and reversible inhibition. Furthermore, several conventional extraction methods have been developed to extract bioactive compounds from natural matrices, such as Soxhlet and hot water extraction. However, these techniques have many drawbacks, as they require a large amount of organic solvents, which not only affect human health but also generate severe environmental issues. To overcome these limitations, multiple eco-extraction techniques have emerged as potential alternatives to traditional extraction methods such as ultrasonic extraction, microwave-assisted extraction, and supercritical fluid extraction. In fact, they are considered eco-friendly and efficient technologies with less time and solvent consumption. Overall, this review aims to provide an updated overview of the most prominent anticancer alkaloids that have not been well reviewed already, as well as the main green extraction techniques relevant to the extraction of antineoplastic alkaloids. Full article

Bone scintigraphy is a widely available and cost-effective modality for detecting skeletal metastases in prostate cancer, yet visual interpretation can be challenging due to heterogeneous uptake patterns, benign mimickers, and a high reporting workload, motivating robust computer-aided decision support. In this study, we present an experimental evaluation of fourteen convolutional neural network (CNN) architectures for binary metastasis classification in planar bone scintigraphy using a unified protocol. Fourteen models, CNN (baseline), AlexNet, VGG16, VGG19, ResNet18, ResNet34, ResNet50, ResNet50-attention, DenseNet121, DenseNet169, DenseNet121-attention, WideResNet50_2, EfficientNet-B0, and ConvNeXt-Tiny, were trained and tested on 600 scan images (300 normal, 300 metastatic) from the Jordanian Royal Medical Services under identical preprocessing and augmentation with stratified five-fold cross-validation. We report mean ± SD for AUC-ROC, accuracy, precision, sensitivity (recall), F1-score, specificity, and Cohen’s κ, alongside calibration via the Brier score and deployment indicators (parameters, FLOPs, model size, and inference time). DenseNet121 achieved the best overall balance of diagnostic performance and reliability, reaching AUC-ROC 96.0 ± 1.2, accuracy 89.2 ± 2.2, sensitivity 83.7 ± 3.4, specificity 94.7 ± 2.2, F1-score 88.5 ± 2.5, κ = 0.783 ± 0.045, and the strongest calibration (Brier 0.080 ± 0.013), with stable fold-to-fold behaviour. DenseNet121-attention produced the highest AUC-ROC (96.3 ± 1.1) but exhibited greater variability in specificity, indicating less consistent false-alarm control. Complexity analysis supported DenseNet121 as deployable (~7.0 M parameters, ~26.9 MB, ~92 ms/image), whereas heavier models yielded only limited additional clinical value. These results support DenseNet121 as a reliable backbone for automated metastasis detection in planar scintigraphy, with future work focusing on external validation, threshold optimisation, interpretability, and model compression for clinical adoption. Full article